Duplicate from Maikou/Michelangelo

Co-authored-by: Zhao <[email protected]>

.gitattributes

@@ -0,0 +1,35 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,3 @@
+.idea
+.vscode
+__pycache__

README.md

@@ -0,0 +1,13 @@
+---
+license: lgpl-3.0
+pipeline_tag: text-to-3d
+tags:
+- image-to-3d
+---
+
+# Michelangelo
+
+* [Project Page](https://neuralcarver.github.io/michelangelo/)
+* [Paper](https://arxiv.org/abs/2306.17115)
+* [Code](https://github.com/NeuralCarver/Michelangelo)
+* [Demo](https://huggingface.co/spaces/Maikou/Michelangelo)

checkpoints/aligned_shape_latents/shapevae-256.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0391b81c36240e8f766fedf4265df599884193a5ef65354525074b9a00887454
+size 3934164973

checkpoints/clip/clip-vit-large-patch14

@@ -0,0 +1 @@
+Subproject commit 8d052a0f05efbaefbc9e8786ba291cfdf93e5bff

checkpoints/image_cond_diffuser_asl/image-ASLDM-256.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83eda8e4f81034dee7674b3ce1ff03a4900181f0f0d7bc461e1a8692fb379b0f
+size 1999253985

checkpoints/text_cond_diffuser_asl/text-ASLDM-256.ckpt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:af546b1f877a41d71f63c3a11394779e77c954002c50dc8e75359338224f615b
+size 4076140813

configs/aligned_shape_latents/shapevae-256.yaml

@@ -0,0 +1,46 @@
+model:
+  target: michelangelo.models.tsal.asl_pl_module.AlignedShapeAsLatentPLModule
+  params:
+    shape_module_cfg:
+      target: michelangelo.models.tsal.sal_perceiver.AlignedShapeLatentPerceiver
+      params:
+        num_latents: 256
+        embed_dim: 64
+        point_feats: 3   # normal
+        num_freqs: 8
+        include_pi: false
+        heads: 12
+        width: 768
+        num_encoder_layers: 8
+        num_decoder_layers: 16
+        use_ln_post: true
+        init_scale: 0.25
+        qkv_bias: false
+        use_checkpoint: true
+    aligned_module_cfg:
+      target: michelangelo.models.tsal.clip_asl_module.CLIPAlignedShapeAsLatentModule
+      params:
+        clip_model_version: "./checkpoints/clip/clip-vit-large-patch14"
+
+    loss_cfg:
+      target: michelangelo.models.tsal.loss.ContrastKLNearFar
+      params:
+        contrast_weight: 0.1
+        near_weight: 0.1
+        kl_weight: 0.001
+
+    optimizer_cfg:
+      optimizer:
+        target: torch.optim.AdamW
+        params:
+          betas: [0.9, 0.99]
+          eps: 1.e-6
+          weight_decay: 1.e-2
+
+      scheduler:
+        target: michelangelo.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler
+        params:
+          warm_up_steps: 5000
+          f_start: 1.e-6
+          f_min: 1.e-3
+          f_max: 1.0

configs/image_cond_diffuser_asl/image-ASLDM-256.yaml

@@ -0,0 +1,97 @@
+model:
+  target: michelangelo.models.asl_diffusion.clip_asl_diffuser_pl_module.ClipASLDiffuser
+  params:
+    first_stage_config:
+      target: michelangelo.models.tsal.asl_pl_module.AlignedShapeAsLatentPLModule
+      params:
+        shape_module_cfg:
+          target: michelangelo.models.tsal.sal_perceiver.AlignedShapeLatentPerceiver
+          params:
+            num_latents: &num_latents 256
+            embed_dim: &embed_dim 64
+            point_feats: 3   # normal
+            num_freqs: 8
+            include_pi: false
+            heads: 12
+            width: 768
+            num_encoder_layers: 8
+            num_decoder_layers: 16
+            use_ln_post: true
+            init_scale: 0.25
+            qkv_bias: false
+            use_checkpoint: false
+        aligned_module_cfg:
+          target: michelangelo.models.tsal.clip_asl_module.CLIPAlignedShapeAsLatentModule
+          params:
+            clip_model_version: "./checkpoints/clip/clip-vit-large-patch14"
+
+        loss_cfg:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: michelangelo.models.conditional_encoders.encoder_factory.FrozenCLIPImageGridEmbedder
+      params:
+        version: "./checkpoints/clip/clip-vit-large-patch14"
+        zero_embedding_radio: 0.1
+
+    first_stage_key: "surface"
+    cond_stage_key: "image"
+    scale_by_std: false
+
+    denoiser_cfg:
+      target: michelangelo.models.asl_diffusion.asl_udt.ConditionalASLUDTDenoiser
+      params:
+        input_channels: *embed_dim
+        output_channels: *embed_dim
+        n_ctx: *num_latents
+        width: 768
+        layers: 6   # 2 * 6 + 1 = 13
+        heads: 12
+        context_dim: 1024
+        init_scale: 1.0
+        skip_ln: true
+        use_checkpoint: true
+
+    scheduler_cfg:
+      guidance_scale: 7.5
+      num_inference_steps: 50
+      eta: 0.0
+
+      noise:
+        target: diffusers.schedulers.DDPMScheduler
+        params:
+          num_train_timesteps: 1000
+          beta_start: 0.00085
+          beta_end: 0.012
+          beta_schedule: "scaled_linear"
+          variance_type: "fixed_small"
+          clip_sample: false
+      denoise:
+        target: diffusers.schedulers.DDIMScheduler
+        params:
+          num_train_timesteps: 1000
+          beta_start: 0.00085
+          beta_end: 0.012
+          beta_schedule: "scaled_linear"
+          clip_sample: false   # clip sample to -1~1
+          set_alpha_to_one: false
+          steps_offset: 1
+
+    optimizer_cfg:
+      optimizer:
+        target: torch.optim.AdamW
+        params:
+          betas: [0.9, 0.99]
+          eps: 1.e-6
+          weight_decay: 1.e-2
+
+      scheduler:
+        target: michelangelo.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler
+        params:
+          warm_up_steps: 5000
+          f_start: 1.e-6
+          f_min: 1.e-3
+          f_max: 1.0
+
+    loss_cfg:
+      loss_type: "mse"

configs/text_cond_diffuser_asl/text-ASLDM-256.yaml

@@ -0,0 +1,98 @@
+model:
+  target: michelangelo.models.asl_diffusion.clip_asl_diffuser_pl_module.ClipASLDiffuser
+  params:
+    first_stage_config:
+      target: michelangelo.models.tsal.asl_pl_module.AlignedShapeAsLatentPLModule
+      params:
+        shape_module_cfg:
+          target: michelangelo.models.tsal.sal_perceiver.AlignedShapeLatentPerceiver
+          params:
+            num_latents: &num_latents 256
+            embed_dim: &embed_dim 64
+            point_feats: 3   # normal
+            num_freqs: 8
+            include_pi: false
+            heads: 12
+            width: 768
+            num_encoder_layers: 8
+            num_decoder_layers: 16
+            use_ln_post: true
+            init_scale: 0.25
+            qkv_bias: false
+            use_checkpoint: true
+        aligned_module_cfg:
+          target: michelangelo.models.tsal.clip_asl_module.CLIPAlignedShapeAsLatentModule
+          params:
+            clip_model_version: "./checkpoints/clip/clip-vit-large-patch14"
+
+        loss_cfg:
+          target: torch.nn.Identity
+
+    cond_stage_config:
+      target: michelangelo.models.conditional_encoders.encoder_factory.FrozenAlignedCLIPTextEmbedder
+      params:
+        version: "./checkpoints/clip/clip-vit-large-patch14"
+        zero_embedding_radio: 0.1
+        max_length: 77
+
+    first_stage_key: "surface"
+    cond_stage_key: "text"
+    scale_by_std: false
+
+    denoiser_cfg:
+      target: michelangelo.models.asl_diffusion.asl_udt.ConditionalASLUDTDenoiser
+      params:
+        input_channels: *embed_dim
+        output_channels: *embed_dim
+        n_ctx: *num_latents
+        width: 768
+        layers: 8   # 2 * 6 + 1 = 13
+        heads: 12
+        context_dim: 768
+        init_scale: 1.0
+        skip_ln: true
+        use_checkpoint: true
+
+    scheduler_cfg:
+      guidance_scale: 7.5
+      num_inference_steps: 50
+      eta: 0.0
+
+      noise:
+        target: diffusers.schedulers.DDPMScheduler
+        params:
+          num_train_timesteps: 1000
+          beta_start: 0.00085
+          beta_end: 0.012
+          beta_schedule: "scaled_linear"
+          variance_type: "fixed_small"
+          clip_sample: false
+      denoise:
+        target: diffusers.schedulers.DDIMScheduler
+        params:
+          num_train_timesteps: 1000
+          beta_start: 0.00085
+          beta_end: 0.012
+          beta_schedule: "scaled_linear"
+          clip_sample: false   # clip sample to -1~1
+          set_alpha_to_one: false
+          steps_offset: 1
+
+    optimizer_cfg:
+      optimizer:
+        target: torch.optim.AdamW
+        params:
+          betas: [0.9, 0.99]
+          eps: 1.e-6
+          weight_decay: 1.e-2
+
+      scheduler:
+        target: michelangelo.utils.trainings.lr_scheduler.LambdaWarmUpCosineFactorScheduler
+        params:
+          warm_up_steps: 5000
+          f_start: 1.e-6
+          f_min: 1.e-3
+          f_max: 1.0
+
+    loss_cfg:
+      loss_type: "mse"

example_data/surface/surface.npz

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0893e44d82ada683baa656a718beaf6ec19fc28b6816b451f56645530d5bb962
+size 1201024

gradio_app.py

@@ -0,0 +1,372 @@
+# -*- coding: utf-8 -*-
+import os
+import time
+from collections import OrderedDict
+from PIL import Image
+import torch
+import trimesh
+from typing import Optional, List
+from einops import repeat, rearrange
+import numpy as np
+from michelangelo.models.tsal.tsal_base import Latent2MeshOutput
+from michelangelo.utils.misc import get_config_from_file, instantiate_from_config
+from michelangelo.utils.visualizers.pythreejs_viewer import PyThreeJSViewer
+from michelangelo.utils.visualizers import html_util
+
+import gradio as gr
+
+
+gradio_cached_dir = "./gradio_cached_dir"
+os.makedirs(gradio_cached_dir, exist_ok=True)
+
+save_mesh = False
+
+state = ""
+device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
+
+box_v = 1.1
+viewer = PyThreeJSViewer(settings={}, render_mode="WEBSITE")
+
+image_model_config_dict = OrderedDict({
+    "ASLDM-256-obj": {
+        "config": "./configs/image_cond_diffuser_asl/image-ASLDM-256.yaml",
+        "ckpt_path": "./checkpoints/image_cond_diffuser_asl/image-ASLDM-256.ckpt",
+    },
+})
+
+text_model_config_dict = OrderedDict({
+    "ASLDM-256": {
+        "config": "./configs/text_cond_diffuser_asl/text-ASLDM-256.yaml",
+        "ckpt_path": "./checkpoints/text_cond_diffuser_asl/text-ASLDM-256.ckpt",
+    },
+})
+
+
+class InferenceModel(object):
+    model = None
+    name = ""
+
+
+text2mesh_model = InferenceModel()
+image2mesh_model = InferenceModel()
+
+
+def set_state(s):
+    global state
+    state = s
+    print(s)
+
+
+def output_to_html_frame(mesh_outputs: List[Latent2MeshOutput], bbox_size: float,
+                         image: Optional[np.ndarray] = None,
+                         html_frame: bool = False):
+    global viewer
+
+    for i in range(len(mesh_outputs)):
+        mesh = mesh_outputs[i]
+        if mesh is None:
+            continue
+
+        mesh_v = mesh.mesh_v.copy()
+        mesh_v[:, 0] += i * np.max(bbox_size)
+        mesh_v[:, 2] += np.max(bbox_size)
+        viewer.add_mesh(mesh_v, mesh.mesh_f)
+
+    mesh_tag = viewer.to_html(html_frame=False)
+
+    if image is not None:
+        image_tag = html_util.to_image_embed_tag(image)
+        frame = f"""
+        <table border = "1">
+            <tr>
+                <td>{image_tag}</td>
+                <td>{mesh_tag}</td>
+            </tr>
+        </table>
+        """
+    else:
+        frame = mesh_tag
+
+    if html_frame:
+        frame = html_util.to_html_frame(frame)
+
+    viewer.reset()
+
+    return frame
+
+
+def load_model(model_name: str, model_config_dict: dict, inference_model: InferenceModel):
+    global device
+
+    if inference_model.name == model_name:
+        model = inference_model.model
+    else:
+        assert model_name in model_config_dict
+
+        if inference_model.model is not None:
+            del inference_model.model
+
+        config_ckpt_path = model_config_dict[model_name]
+
+        model_config = get_config_from_file(config_ckpt_path["config"])
+        if hasattr(model_config, "model"):
+            model_config = model_config.model
+
+        model = instantiate_from_config(model_config, ckpt_path=config_ckpt_path["ckpt_path"])
+        model = model.to(device)
+        model = model.eval()
+
+        inference_model.model = model
+        inference_model.name = model_name
+
+    return model
+
+
+def prepare_img(image: np.ndarray):
+    image_pt = torch.tensor(image).float()
+    image_pt = image_pt / 255 * 2 - 1
+    image_pt = rearrange(image_pt, "h w c -> c h w")
+
+    return image_pt
+
+def prepare_model_viewer(fp):
+    content = f"""
+      <head>
+        <script 
+          type="module" src="https://ajax.googleapis.com/ajax/libs/model-viewer/3.1.1/model-viewer.min.js">
+        </script>
+      </head>
+      <body>
+        <model-viewer 
+          style="height: 150px; width: 150px;"
+          rotation-per-second="10deg"
+          id="t1"
+          src="file/gradio_cached_dir/{fp}" 
+          environment-image="neutral" 
+          camera-target="0m 0m 0m"
+          orientation="0deg 90deg 170deg"
+          shadow-intensity="1"
+          ar:true 
+          auto-rotate 
+          camera-controls>
+        </model-viewer>
+      </body>
+    """
+    return content
+
+def prepare_html_frame(content):
+    frame = f"""
+    <html>
+      <body>
+        {content}
+      </body>
+    </html>
+    """
+    return frame
+
+def prepare_html_body(content):
+    frame = f"""
+      <body>
+        {content}
+      </body>
+    """
+    return frame
+
+def post_process_mesh_outputs(mesh_outputs):
+    # html_frame = output_to_html_frame(mesh_outputs, 2 * box_v, image=None, html_frame=True)
+    html_content = output_to_html_frame(mesh_outputs, 2 * box_v, image=None, html_frame=False)
+    html_frame = prepare_html_frame(html_content)
+
+    # filename = f"{time.time()}.html"
+    filename = f"text-256-{time.time()}.html"
+    html_filepath = os.path.join(gradio_cached_dir, filename)
+    with open(html_filepath, "w") as writer:
+        writer.write(html_frame)
+
+    '''
+    Bug: The iframe tag does not work in Gradio.
+         The chrome returns "No resource with given URL found" 
+    Solutions:
+         https://github.com/gradio-app/gradio/issues/884
+         Due to the security bitches, the server can only find files parallel to the gradio_app.py.
+         The path has format "file/TARGET_FILE_PATH"
+    '''
+
+    iframe_tag = f'<iframe src="file/gradio_cached_dir/{filename}" width="600%" height="400" frameborder="0"></iframe>'
+
+    filelist = []
+    filenames = []
+    for i, mesh in enumerate(mesh_outputs):
+        mesh.mesh_f = mesh.mesh_f[:, ::-1]
+        mesh_output = trimesh.Trimesh(mesh.mesh_v, mesh.mesh_f)
+
+        name = str(i) + "_out_mesh.obj"
+        filepath = gradio_cached_dir + "/" + name
+        mesh_output.export(filepath, include_normals=True)
+        filelist.append(filepath)
+        filenames.append(name)
+
+    filelist.append(html_filepath)
+    return iframe_tag, filelist
+
+def image2mesh(image: np.ndarray,
+               model_name: str = "subsp+pk_asl_perceiver=01_01_udt=03",
+               num_samples: int = 4,
+               guidance_scale: int = 7.5,
+               octree_depth: int = 7):
+    global device, gradio_cached_dir, image_model_config_dict, box_v
+
+    # load model
+    model = load_model(model_name, image_model_config_dict, image2mesh_model)
+
+    # prepare image inputs
+    image_pt = prepare_img(image)
+    image_pt = repeat(image_pt, "c h w -> b c h w", b=num_samples)
+
+    sample_inputs = {
+        "image": image_pt
+    }
+    mesh_outputs = model.sample(
+        sample_inputs,
+        sample_times=1,
+        guidance_scale=guidance_scale,
+        return_intermediates=False,
+        bounds=[-box_v, -box_v, -box_v, box_v, box_v, box_v],
+        octree_depth=octree_depth,
+    )[0]
+
+    iframe_tag, filelist = post_process_mesh_outputs(mesh_outputs)
+
+    return iframe_tag, gr.update(value=filelist, visible=True)
+
+
+def text2mesh(text: str,
+              model_name: str = "subsp+pk_asl_perceiver=01_01_udt=03",
+              num_samples: int = 4,
+              guidance_scale: int = 7.5,
+              octree_depth: int = 7):
+    global device, gradio_cached_dir, text_model_config_dict, text2mesh_model, box_v
+
+    # load model
+    model = load_model(model_name, text_model_config_dict, text2mesh_model)
+
+    # prepare text inputs
+    sample_inputs = {
+        "text": [text] * num_samples
+    }
+    mesh_outputs = model.sample(
+        sample_inputs,
+        sample_times=1,
+        guidance_scale=guidance_scale,
+        return_intermediates=False,
+        bounds=[-box_v, -box_v, -box_v, box_v, box_v, box_v],
+        octree_depth=octree_depth,
+    )[0]
+
+    iframe_tag, filelist = post_process_mesh_outputs(mesh_outputs)
+
+    return iframe_tag, gr.update(value=filelist, visible=True)
+
+example_dir = './gradio_cached_dir/example/img_example'
+
+first_page_items = [
+    'alita.jpg',
+    'burger.jpg'
+    'loopy.jpg'
+    'building.jpg',
+    'mario.jpg',
+    'car.jpg',
+    'airplane.jpg',
+    'bag.jpg',
+    'bench.jpg',
+    'ship.jpg'
+]
+raw_example_items = [
+    # (os.path.join(example_dir, x), x)
+    os.path.join(example_dir, x)
+    for x in os.listdir(example_dir)
+    if x.endswith(('.jpg', '.png'))
+]
+example_items = [x for x in raw_example_items if os.path.basename(x) in first_page_items] + [x for x in raw_example_items if os.path.basename(x) not in first_page_items]
+
+example_text = [
+    ["A 3D model of a car; Audi A6."],
+    ["A 3D model of police car; Highway Patrol Charger"]
+    ],
+
+def set_cache(data: gr.SelectData):
+    img_name = os.path.basename(example_items[data.index])
+    return os.path.join(example_dir, img_name), os.path.join(img_name)
+
+def disable_cache():
+    return ""
+
+with gr.Blocks() as app:
+    gr.Markdown("# Michelangelo")
+    gr.Markdown("## [Github](https://github.com/NeuralCarver/Michelangelo) | [Arxiv](https://arxiv.org/abs/2306.17115) | [Project Page](https://neuralcarver.github.io/michelangelo/)")
+    gr.Markdown("Michelangelo is a conditional 3D shape generation system that trains based on the shape-image-text aligned latent representation.")
+    gr.Markdown("### Hint:")
+    gr.Markdown("1. We provide two APIs: Image-conditioned generation and Text-conditioned generation")
+    gr.Markdown("2. Note that the Image-conditioned model is trained on multiple 3D datasets like ShapeNet and Objaverse")
+    gr.Markdown("3. We provide some examples for you to try. You can also upload images or text as input.")
+    gr.Markdown("4. Welcome to share your amazing results with us, and thanks for your interest in our work!")
+
+    with gr.Row():
+        with gr.Column():
+            
+            with gr.Tab("Image to 3D"):
+                img = gr.Image(label="Image")
+                gr.Markdown("For the best results, we suggest that the images uploaded meet the following three criteria: 1. The object is positioned at the center of the image, 2. The image size is square, and 3. The background is relatively clean.")
+                btn_generate_img2obj = gr.Button(value="Generate")
+                
+                with gr.Accordion("Advanced settings", open=False):
+                    image_dropdown_models = gr.Dropdown(label="Model", value="ASLDM-256-obj",choices=list(image_model_config_dict.keys()))
+                    num_samples = gr.Slider(label="samples", value=4, minimum=1, maximum=8, step=1)
+                    guidance_scale = gr.Slider(label="Guidance scale", value=7.5, minimum=3.0, maximum=10.0, step=0.1)
+                    octree_depth = gr.Slider(label="Octree Depth (for 3D model)", value=7, minimum=4, maximum=8, step=1)
+
+                
+                cache_dir = gr.Textbox(value="", visible=False)
+                examples = gr.Gallery(label='Examples', value=example_items, elem_id="gallery", allow_preview=False, columns=[4], object_fit="contain")
+            
+            with gr.Tab("Text to 3D"):
+                prompt = gr.Textbox(label="Prompt", placeholder="A 3D model of motorcar; Porche Cayenne Turbo.")
+                gr.Markdown("For the best results, we suggest that the prompt follows 'A 3D model of CATEGORY; DESCRIPTION'. For example, A 3D model of motorcar; Porche Cayenne Turbo.")
+                btn_generate_txt2obj = gr.Button(value="Generate")
+                
+                with gr.Accordion("Advanced settings", open=False):
+                    text_dropdown_models = gr.Dropdown(label="Model", value="ASLDM-256",choices=list(text_model_config_dict.keys()))
+                    num_samples = gr.Slider(label="samples", value=4, minimum=1, maximum=8, step=1)
+                    guidance_scale = gr.Slider(label="Guidance scale", value=7.5, minimum=3.0, maximum=10.0, step=0.1)
+                    octree_depth = gr.Slider(label="Octree Depth (for 3D model)", value=7, minimum=4, maximum=8, step=1)
+
+                gr.Markdown("#### Examples:")
+                gr.Markdown("1. A 3D model of a coupe; Audi A6.")
+                gr.Markdown("2. A 3D model of a motorcar; Hummer H2 SUT.")
+                gr.Markdown("3. A 3D model of an airplane; Airbus.")
+                gr.Markdown("4. A 3D model of a fighter aircraft; Attack Fighter.")
+                gr.Markdown("5. A 3D model of a chair; Simple Wooden Chair.")
+                gr.Markdown("6. A 3D model of a laptop computer; Dell Laptop.")
+                gr.Markdown("7. A 3D model of a lamp; ceiling light.")
+                gr.Markdown("8. A 3D model of a rifle; AK47.")
+                gr.Markdown("9. A 3D model of a knife; Sword.")
+                gr.Markdown("10. A 3D model of a vase; Plant in pot.")
+
+        with gr.Column():
+            model_3d = gr.HTML()
+            file_out = gr.File(label="Files", visible=False)
+
+        outputs = [model_3d, file_out]
+        
+        img.upload(disable_cache, outputs=cache_dir)
+        examples.select(set_cache, outputs=[img, cache_dir])
+        print(f'line:404: {cache_dir}')
+        btn_generate_img2obj.click(image2mesh, inputs=[img, image_dropdown_models, num_samples,
+                                                       guidance_scale, octree_depth],
+                                   outputs=outputs, api_name="generate_img2obj")
+
+        btn_generate_txt2obj.click(text2mesh, inputs=[prompt, text_dropdown_models, num_samples,
+                                                      guidance_scale, octree_depth],
+                                   outputs=outputs, api_name="generate_txt2obj")
+
+app.launch(server_name="0.0.0.0", server_port=8008, share=False)

inference.py

@@ -0,0 +1,181 @@
+# -*- coding: utf-8 -*-
+import os
+import time
+from collections import OrderedDict
+from typing import Optional, List
+import argparse
+from functools import partial
+
+from einops import repeat, rearrange
+import numpy as np
+from PIL import Image
+import trimesh
+import cv2
+
+import torch
+import pytorch_lightning as pl
+
+from michelangelo.models.tsal.tsal_base import Latent2MeshOutput
+from michelangelo.models.tsal.inference_utils import extract_geometry
+from michelangelo.utils.misc import get_config_from_file, instantiate_from_config
+from michelangelo.utils.visualizers.pythreejs_viewer import PyThreeJSViewer
+from michelangelo.utils.visualizers import html_util
+
+def load_model(args):
+
+    model_config = get_config_from_file(args.config_path)
+    if hasattr(model_config, "model"):
+        model_config = model_config.model
+
+    model = instantiate_from_config(model_config, ckpt_path=args.ckpt_path)
+    model = model.cuda()
+    model = model.eval()
+
+    return model
+
+def load_surface(fp):
+    
+    with np.load(args.pointcloud_path) as input_pc:
+        surface = input_pc['points']
+        normal = input_pc['normals']
+    
+    rng = np.random.default_rng()
+    ind = rng.choice(surface.shape[0], 4096, replace=False)
+    surface = torch.FloatTensor(surface[ind])
+    normal = torch.FloatTensor(normal[ind])
+    
+    surface = torch.cat([surface, normal], dim=-1).unsqueeze(0).cuda()
+    
+    return surface
+
+def prepare_image(args, number_samples=2):
+    
+    image = cv2.imread(f"{args.image_path}")
+    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+    
+    image_pt = torch.tensor(image).float()
+    image_pt = image_pt / 255 * 2 - 1
+    image_pt = rearrange(image_pt, "h w c -> c h w")
+    
+    image_pt = repeat(image_pt, "c h w -> b c h w", b=number_samples)
+
+    return image_pt
+
+def save_output(args, mesh_outputs):
+    
+    os.makedirs(args.output_dir, exist_ok=True)
+    for i, mesh in enumerate(mesh_outputs):
+        mesh.mesh_f = mesh.mesh_f[:, ::-1]
+        mesh_output = trimesh.Trimesh(mesh.mesh_v, mesh.mesh_f)
+
+        name = str(i) + "_out_mesh.obj"
+        mesh_output.export(os.path.join(args.output_dir, name), include_normals=True)
+
+    print(f'-----------------------------------------------------------------------------')
+    print(f'>>> Finished and mesh saved in {args.output_dir}')
+    print(f'-----------------------------------------------------------------------------')        
+
+    return 0
+
+def reconstruction(args, model, bounds=(-1.25, -1.25, -1.25, 1.25, 1.25, 1.25), octree_depth=7, num_chunks=10000):
+
+    surface = load_surface(args.pointcloud_path)
+    
+    # encoding
+    shape_embed, shape_latents = model.model.encode_shape_embed(surface, return_latents=True)    
+    shape_zq, posterior = model.model.shape_model.encode_kl_embed(shape_latents)
+
+    # decoding
+    latents = model.model.shape_model.decode(shape_zq)
+    geometric_func = partial(model.model.shape_model.query_geometry, latents=latents)
+    
+    # reconstruction
+    mesh_v_f, has_surface = extract_geometry(
+        geometric_func=geometric_func,
+        device=surface.device,
+        batch_size=surface.shape[0],
+        bounds=bounds,
+        octree_depth=octree_depth,
+        num_chunks=num_chunks,
+    )
+    recon_mesh = trimesh.Trimesh(mesh_v_f[0][0], mesh_v_f[0][1])
+    
+    # save
+    os.makedirs(args.output_dir, exist_ok=True)
+    recon_mesh.export(os.path.join(args.output_dir, 'reconstruction.obj'))    
+    
+    print(f'-----------------------------------------------------------------------------')
+    print(f'>>> Finished and mesh saved in {os.path.join(args.output_dir, "reconstruction.obj")}')
+    print(f'-----------------------------------------------------------------------------')
+    
+    return 0
+
+def image2mesh(args, model, guidance_scale=7.5, box_v=1.1, octree_depth=7):
+
+    sample_inputs = {
+        "image": prepare_image(args)
+    }
+    
+    mesh_outputs = model.sample(
+        sample_inputs,
+        sample_times=1,
+        guidance_scale=guidance_scale,
+        return_intermediates=False,
+        bounds=[-box_v, -box_v, -box_v, box_v, box_v, box_v],
+        octree_depth=octree_depth,
+    )[0]
+    
+    save_output(args, mesh_outputs)
+    
+    return 0
+
+def text2mesh(args, model, num_samples=2, guidance_scale=7.5, box_v=1.1, octree_depth=7):
+
+    sample_inputs = {
+        "text": [args.text] * num_samples
+    }
+    mesh_outputs = model.sample(
+        sample_inputs,
+        sample_times=1,
+        guidance_scale=guidance_scale,
+        return_intermediates=False,
+        bounds=[-box_v, -box_v, -box_v, box_v, box_v, box_v],
+        octree_depth=octree_depth,
+    )[0]
+    
+    save_output(args, mesh_outputs)
+    
+    return 0
+
+task_dick = {
+    'reconstruction': reconstruction,
+    'image2mesh': image2mesh,
+    'text2mesh': text2mesh,
+}
+
+if __name__ == "__main__":
+    '''
+    1. Reconstruct point cloud
+    2. Image-conditioned generation
+    3. Text-conditioned generation
+    '''
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--task", type=str, choices=['reconstruction', 'image2mesh', 'text2mesh'], required=True)
+    parser.add_argument("--config_path", type=str, required=True)
+    parser.add_argument("--ckpt_path", type=str, required=True)
+    parser.add_argument("--pointcloud_path", type=str, default='./example_data/surface.npz', help='Path to the input point cloud')
+    parser.add_argument("--image_path", type=str, help='Path to the input image')
+    parser.add_argument("--text", type=str, help='Input text within a format: A 3D model of motorcar; Porsche 911.')
+    parser.add_argument("--output_dir", type=str, default='./output')
+    parser.add_argument("-s", "--seed", type=int, default=0)
+    args = parser.parse_args()
+    
+    pl.seed_everything(args.seed)
+
+    print(f'-----------------------------------------------------------------------------')
+    print(f'>>> Running {args.task}')
+    args.output_dir = os.path.join(args.output_dir, args.task)
+    print(f'>>> Output directory: {args.output_dir}')
+    print(f'-----------------------------------------------------------------------------')
+    
+    task_dick[args.task](args, load_model(args))

michelangelo/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

michelangelo/data/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

michelangelo/data/templates.json

@@ -0,0 +1,69 @@
+{
+    "shape": [
+        "a point cloud model of {}.",
+        "There is a {} in the scene.",
+        "There is the {} in the scene.",
+        "a photo of a {} in the scene.",
+        "a photo of the {} in the scene.",
+        "a photo of one {} in the scene.",
+        "itap of a {}.",
+        "itap of my {}.",
+        "itap of the {}.",
+        "a photo of a {}.",
+        "a photo of my {}.",
+        "a photo of the {}.",
+        "a photo of one {}.",
+        "a photo of many {}.",
+        "a good photo of a {}.",
+        "a good photo of the {}.",
+        "a bad photo of a {}.",
+        "a bad photo of the {}.",
+        "a photo of a nice {}.",
+        "a photo of the nice {}.",
+        "a photo of a cool {}.",
+        "a photo of the cool {}.",
+        "a photo of a weird {}.",
+        "a photo of the weird {}.",
+        "a photo of a small {}.",
+        "a photo of the small {}.",
+        "a photo of a large {}.",
+        "a photo of the large {}.",
+        "a photo of a clean {}.",
+        "a photo of the clean {}.",
+        "a photo of a dirty {}.",
+        "a photo of the dirty {}.",
+        "a bright photo of a {}.",
+        "a bright photo of the {}.",
+        "a dark photo of a {}.",
+        "a dark photo of the {}.",
+        "a photo of a hard to see {}.",
+        "a photo of the hard to see {}.",
+        "a low resolution photo of a {}.",
+        "a low resolution photo of the {}.",
+        "a cropped photo of a {}.",
+        "a cropped photo of the {}.",
+        "a close-up photo of a {}.",
+        "a close-up photo of the {}.",
+        "a jpeg corrupted photo of a {}.",
+        "a jpeg corrupted photo of the {}.",
+        "a blurry photo of a {}.",
+        "a blurry photo of the {}.",
+        "a pixelated photo of a {}.",
+        "a pixelated photo of the {}.",
+        "a black and white photo of the {}.",
+        "a black and white photo of a {}",
+        "a plastic {}.",
+        "the plastic {}.",
+        "a toy {}.",
+        "the toy {}.",
+        "a plushie {}.",
+        "the plushie {}.",
+        "a cartoon {}.",
+        "the cartoon {}.",
+        "an embroidered {}.",
+        "the embroidered {}.",
+        "a painting of the {}.",
+        "a painting of a {}."
+    ]
+
+}

michelangelo/data/transforms.py

@@ -0,0 +1,407 @@
+# -*- coding: utf-8 -*-
+import os
+import time
+import numpy as np
+import warnings
+import random
+from omegaconf.listconfig import ListConfig
+from webdataset import pipelinefilter
+import torch
+import torchvision.transforms.functional as TVF
+from torchvision.transforms import InterpolationMode
+from torchvision.transforms.transforms import _interpolation_modes_from_int
+from typing import Sequence
+
+from michelangelo.utils import instantiate_from_config
+
+
+def _uid_buffer_pick(buf_dict, rng):
+    uid_keys = list(buf_dict.keys())
+    selected_uid = rng.choice(uid_keys)
+    buf = buf_dict[selected_uid]
+
+    k = rng.randint(0, len(buf) - 1)
+    sample = buf[k]
+    buf[k] = buf[-1]
+    buf.pop()
+
+    if len(buf) == 0:
+        del buf_dict[selected_uid]
+
+    return sample
+
+
+def _add_to_buf_dict(buf_dict, sample):
+    key = sample["__key__"]
+    uid, uid_sample_id = key.split("_")
+    if uid not in buf_dict:
+        buf_dict[uid] = []
+    buf_dict[uid].append(sample)
+
+    return buf_dict
+
+
+def _uid_shuffle(data, bufsize=1000, initial=100, rng=None, handler=None):
+    """Shuffle the data in the stream.
+
+    This uses a buffer of size `bufsize`. Shuffling at
+    startup is less random; this is traded off against
+    yielding samples quickly.
+
+    data: iterator
+    bufsize: buffer size for shuffling
+    returns: iterator
+    rng: either random module or random.Random instance
+
+    """
+    if rng is None:
+        rng = random.Random(int((os.getpid() + time.time()) * 1e9))
+    initial = min(initial, bufsize)
+    buf_dict = dict()
+    current_samples = 0
+    for sample in data:
+        _add_to_buf_dict(buf_dict, sample)
+        current_samples += 1
+
+        if current_samples < bufsize:
+            try:
+                _add_to_buf_dict(buf_dict, next(data))  # skipcq: PYL-R1708
+                current_samples += 1
+            except StopIteration:
+                pass
+
+        if current_samples >= initial:
+            current_samples -= 1
+            yield _uid_buffer_pick(buf_dict, rng)
+
+    while current_samples > 0:
+        current_samples -= 1
+        yield _uid_buffer_pick(buf_dict, rng)
+
+
+uid_shuffle = pipelinefilter(_uid_shuffle)
+
+
+class RandomSample(object):
+    def __init__(self,
+                 num_volume_samples: int = 1024,
+                 num_near_samples: int = 1024):
+
+        super().__init__()
+
+        self.num_volume_samples = num_volume_samples
+        self.num_near_samples = num_near_samples
+
+    def __call__(self, sample):
+        rng = np.random.default_rng()
+
+        # 1. sample surface input
+        total_surface = sample["surface"]
+        ind = rng.choice(total_surface.shape[0], replace=False)
+        surface = total_surface[ind]
+
+        # 2. sample volume/near geometric points
+        vol_points = sample["vol_points"]
+        vol_label = sample["vol_label"]
+        near_points = sample["near_points"]
+        near_label = sample["near_label"]
+
+        ind = rng.choice(vol_points.shape[0], self.num_volume_samples, replace=False)
+        vol_points = vol_points[ind]
+        vol_label = vol_label[ind]
+        vol_points_labels = np.concatenate([vol_points, vol_label[:, np.newaxis]], axis=1)
+
+        ind = rng.choice(near_points.shape[0], self.num_near_samples, replace=False)
+        near_points = near_points[ind]
+        near_label = near_label[ind]
+        near_points_labels = np.concatenate([near_points, near_label[:, np.newaxis]], axis=1)
+
+        # concat sampled volume and near points
+        geo_points = np.concatenate([vol_points_labels, near_points_labels], axis=0)
+
+        sample = {
+            "surface": surface,
+            "geo_points": geo_points
+        }
+
+        return sample
+
+
+class SplitRandomSample(object):
+    def __init__(self,
+                 use_surface_sample: bool = False,
+                 num_surface_samples: int = 4096,
+                 num_volume_samples: int = 1024,
+                 num_near_samples: int = 1024):
+
+        super().__init__()
+
+        self.use_surface_sample = use_surface_sample
+        self.num_surface_samples = num_surface_samples
+        self.num_volume_samples = num_volume_samples
+        self.num_near_samples = num_near_samples
+
+    def __call__(self, sample):
+
+        rng = np.random.default_rng()
+
+        # 1. sample surface input
+        surface = sample["surface"]
+
+        if self.use_surface_sample:
+            replace = surface.shape[0] < self.num_surface_samples
+            ind = rng.choice(surface.shape[0], self.num_surface_samples, replace=replace)
+            surface = surface[ind]
+
+        # 2. sample volume/near geometric points
+        vol_points = sample["vol_points"]
+        vol_label = sample["vol_label"]
+        near_points = sample["near_points"]
+        near_label = sample["near_label"]
+
+        ind = rng.choice(vol_points.shape[0], self.num_volume_samples, replace=False)
+        vol_points = vol_points[ind]
+        vol_label = vol_label[ind]
+        vol_points_labels = np.concatenate([vol_points, vol_label[:, np.newaxis]], axis=1)
+
+        ind = rng.choice(near_points.shape[0], self.num_near_samples, replace=False)
+        near_points = near_points[ind]
+        near_label = near_label[ind]
+        near_points_labels = np.concatenate([near_points, near_label[:, np.newaxis]], axis=1)
+
+        # concat sampled volume and near points
+        geo_points = np.concatenate([vol_points_labels, near_points_labels], axis=0)
+
+        sample = {
+            "surface": surface,
+            "geo_points": geo_points
+        }
+
+        return sample
+
+
+class FeatureSelection(object):
+
+    VALID_SURFACE_FEATURE_DIMS = {
+        "none": [0, 1, 2],                              # xyz
+        "watertight_normal": [0, 1, 2, 3, 4, 5],        # xyz, normal
+        "normal": [0, 1, 2, 6, 7, 8]
+    }
+
+    def __init__(self, surface_feature_type: str):
+
+        self.surface_feature_type = surface_feature_type
+        self.surface_dims = self.VALID_SURFACE_FEATURE_DIMS[surface_feature_type]
+
+    def __call__(self, sample):
+        sample["surface"] = sample["surface"][:, self.surface_dims]
+        return sample
+
+
+class AxisScaleTransform(object):
+    def __init__(self, interval=(0.75, 1.25), jitter=True, jitter_scale=0.005):
+        assert isinstance(interval, (tuple, list, ListConfig))
+        self.interval = interval
+        self.min_val = interval[0]
+        self.max_val = interval[1]
+        self.inter_size = interval[1] - interval[0]
+        self.jitter = jitter
+        self.jitter_scale = jitter_scale
+
+    def __call__(self, sample):
+
+        surface = sample["surface"][..., 0:3]
+        geo_points = sample["geo_points"][..., 0:3]
+
+        scaling = torch.rand(1, 3) * self.inter_size + self.min_val
+        # print(scaling)
+        surface = surface * scaling
+        geo_points = geo_points * scaling
+
+        scale = (1 / torch.abs(surface).max().item()) * 0.999999
+        surface *= scale
+        geo_points *= scale
+
+        if self.jitter:
+            surface += self.jitter_scale * torch.randn_like(surface)
+            surface.clamp_(min=-1.015, max=1.015)
+
+        sample["surface"][..., 0:3] = surface
+        sample["geo_points"][..., 0:3] = geo_points
+
+        return sample
+
+
+class ToTensor(object):
+
+    def __init__(self, tensor_keys=("surface", "geo_points", "tex_points")):
+        self.tensor_keys = tensor_keys
+
+    def __call__(self, sample):
+        for key in self.tensor_keys:
+            if key not in sample:
+                continue
+
+            sample[key] = torch.tensor(sample[key], dtype=torch.float32)
+
+        return sample
+
+
+class AxisScale(object):
+    def __init__(self, interval=(0.75, 1.25), jitter=True, jitter_scale=0.005):
+        assert isinstance(interval, (tuple, list, ListConfig))
+        self.interval = interval
+        self.jitter = jitter
+        self.jitter_scale = jitter_scale
+
+    def __call__(self, surface, *args):
+        scaling = torch.rand(1, 3) * 0.5 + 0.75
+        # print(scaling)
+        surface = surface * scaling
+        scale = (1 / torch.abs(surface).max().item()) * 0.999999
+        surface *= scale
+
+        args_outputs = []
+        for _arg in args:
+            _arg = _arg * scaling * scale
+            args_outputs.append(_arg)
+
+        if self.jitter:
+            surface += self.jitter_scale * torch.randn_like(surface)
+            surface.clamp_(min=-1, max=1)
+
+        if len(args) == 0:
+            return surface
+        else:
+            return surface, *args_outputs
+
+
+class RandomResize(torch.nn.Module):
+    """Apply randomly Resize with a given probability."""
+
+    def __init__(
+        self,
+        size,
+        resize_radio=(0.5, 1),
+        allow_resize_interpolations=(InterpolationMode.BICUBIC, InterpolationMode.BILINEAR, InterpolationMode.BILINEAR),
+        interpolation=InterpolationMode.BICUBIC,
+        max_size=None,
+        antialias=None,
+    ):
+        super().__init__()
+        if not isinstance(size, (int, Sequence)):
+            raise TypeError(f"Size should be int or sequence. Got {type(size)}")
+        if isinstance(size, Sequence) and len(size) not in (1, 2):
+            raise ValueError("If size is a sequence, it should have 1 or 2 values")
+
+        self.size = size
+        self.max_size = max_size
+        # Backward compatibility with integer value
+        if isinstance(interpolation, int):
+            warnings.warn(
+                "Argument 'interpolation' of type int is deprecated since 0.13 and will be removed in 0.15. "
+                "Please use InterpolationMode enum."
+            )
+            interpolation = _interpolation_modes_from_int(interpolation)
+
+        self.interpolation = interpolation
+        self.antialias = antialias
+
+        self.resize_radio = resize_radio
+        self.allow_resize_interpolations = allow_resize_interpolations
+
+    def random_resize_params(self):
+        radio = torch.rand(1) * (self.resize_radio[1] - self.resize_radio[0]) + self.resize_radio[0]
+
+        if isinstance(self.size, int):
+            size = int(self.size * radio)
+        elif isinstance(self.size, Sequence):
+            size = list(self.size)
+            size = (int(size[0] * radio), int(size[1] * radio))
+        else:
+            raise RuntimeError()
+
+        interpolation = self.allow_resize_interpolations[
+            torch.randint(low=0, high=len(self.allow_resize_interpolations), size=(1,))
+        ]
+        return size, interpolation
+
+    def forward(self, img):
+        size, interpolation = self.random_resize_params()
+        img = TVF.resize(img, size, interpolation, self.max_size, self.antialias)
+        img = TVF.resize(img, self.size, self.interpolation, self.max_size, self.antialias)
+        return img
+
+    def __repr__(self) -> str:
+        detail = f"(size={self.size}, interpolation={self.interpolation.value},"
+        detail += f"max_size={self.max_size}, antialias={self.antialias}), resize_radio={self.resize_radio}"
+        return f"{self.__class__.__name__}{detail}"
+
+
+class Compose(object):
+    """Composes several transforms together. This transform does not support torchscript.
+    Please, see the note below.
+
+    Args:
+        transforms (list of ``Transform`` objects): list of transforms to compose.
+
+    Example:
+        >>> transforms.Compose([
+        >>>     transforms.CenterCrop(10),
+        >>>     transforms.ToTensor(),
+        >>> ])
+
+    .. note::
+        In order to script the transformations, please use ``torch.nn.Sequential`` as below.
+
+        >>> transforms = torch.nn.Sequential(
+        >>>     transforms.CenterCrop(10),
+        >>>     transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
+        >>> )
+        >>> scripted_transforms = torch.jit.script(transforms)
+
+        Make sure to use only scriptable transformations, i.e. that work with ``torch.Tensor``, does not require
+        `lambda` functions or ``PIL.Image``.
+
+    """
+
+    def __init__(self, transforms):
+        self.transforms = transforms
+
+    def __call__(self, *args):
+        for t in self.transforms:
+            args = t(*args)
+        return args
+
+    def __repr__(self):
+        format_string = self.__class__.__name__ + '('
+        for t in self.transforms:
+            format_string += '\n'
+            format_string += '    {0}'.format(t)
+        format_string += '\n)'
+        return format_string
+
+
+def identity(*args, **kwargs):
+    if len(args) == 1:
+        return args[0]
+    else:
+        return args
+
+
+def build_transforms(cfg):
+
+    if cfg is None:
+        return identity
+
+    transforms = []
+
+    for transform_name, cfg_instance in cfg.items():
+        transform_instance = instantiate_from_config(cfg_instance)
+        transforms.append(transform_instance)
+        print(f"Build transform: {transform_instance}")
+
+    transforms = Compose(transforms)
+
+    return transforms
+

michelangelo/data/utils.py

@@ -0,0 +1,59 @@
+# -*- coding: utf-8 -*-
+
+import torch
+import numpy as np
+
+
+def worker_init_fn(_):
+    worker_info = torch.utils.data.get_worker_info()
+    worker_id = worker_info.id
+
+    # dataset = worker_info.dataset
+    # split_size = dataset.num_records // worker_info.num_workers
+    # # reset num_records to the true number to retain reliable length information
+    # dataset.sample_ids = dataset.valid_ids[worker_id * split_size:(worker_id + 1) * split_size]
+    # current_id = np.random.choice(len(np.random.get_state()[1]), 1)
+    # return np.random.seed(np.random.get_state()[1][current_id] + worker_id)
+
+    return np.random.seed(np.random.get_state()[1][0] + worker_id)
+
+
+def collation_fn(samples, combine_tensors=True, combine_scalars=True):
+    """
+
+    Args:
+        samples (list[dict]):
+        combine_tensors:
+        combine_scalars:
+
+    Returns:
+
+    """
+
+    result = {}
+
+    keys = samples[0].keys()
+
+    for key in keys:
+        result[key] = []
+
+    for sample in samples:
+        for key in keys:
+            val = sample[key]
+            result[key].append(val)
+
+    for key in keys:
+        val_list = result[key]
+        if isinstance(val_list[0], (int, float)):
+            if combine_scalars:
+                result[key] = np.array(result[key])
+
+        elif isinstance(val_list[0], torch.Tensor):
+            if combine_tensors:
+                result[key] = torch.stack(val_list)
+
+        elif isinstance(val_list[0], np.ndarray):
+            if combine_tensors:
+                result[key] = np.stack(val_list)
+
+    return result

michelangelo/graphics/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

michelangelo/graphics/primitives/__init__.py

@@ -0,0 +1,9 @@
+# -*- coding: utf-8 -*-
+
+from .volume import generate_dense_grid_points
+
+from .mesh import (
+    MeshOutput,
+    save_obj,
+    savemeshtes2
+)

michelangelo/graphics/primitives/mesh.py

@@ -0,0 +1,114 @@
+# -*- coding: utf-8 -*-
+
+import os
+import cv2
+import numpy as np
+import PIL.Image
+from typing import Optional
+
+import trimesh
+
+
+def save_obj(pointnp_px3, facenp_fx3, fname):
+    fid = open(fname, "w")
+    write_str = ""
+    for pidx, p in enumerate(pointnp_px3):
+        pp = p
+        write_str += "v %f %f %f\n" % (pp[0], pp[1], pp[2])
+
+    for i, f in enumerate(facenp_fx3):
+        f1 = f + 1
+        write_str += "f %d %d %d\n" % (f1[0], f1[1], f1[2])
+    fid.write(write_str)
+    fid.close()
+    return
+
+
+def savemeshtes2(pointnp_px3, tcoords_px2, facenp_fx3, facetex_fx3, tex_map, fname):
+    fol, na = os.path.split(fname)
+    na, _ = os.path.splitext(na)
+
+    matname = "%s/%s.mtl" % (fol, na)
+    fid = open(matname, "w")
+    fid.write("newmtl material_0\n")
+    fid.write("Kd 1 1 1\n")
+    fid.write("Ka 0 0 0\n")
+    fid.write("Ks 0.4 0.4 0.4\n")
+    fid.write("Ns 10\n")
+    fid.write("illum 2\n")
+    fid.write("map_Kd %s.png\n" % na)
+    fid.close()
+    ####
+
+    fid = open(fname, "w")
+    fid.write("mtllib %s.mtl\n" % na)
+
+    for pidx, p in enumerate(pointnp_px3):
+        pp = p
+        fid.write("v %f %f %f\n" % (pp[0], pp[1], pp[2]))
+
+    for pidx, p in enumerate(tcoords_px2):
+        pp = p
+        fid.write("vt %f %f\n" % (pp[0], pp[1]))
+
+    fid.write("usemtl material_0\n")
+    for i, f in enumerate(facenp_fx3):
+        f1 = f + 1
+        f2 = facetex_fx3[i] + 1
+        fid.write("f %d/%d %d/%d %d/%d\n" % (f1[0], f2[0], f1[1], f2[1], f1[2], f2[2]))
+    fid.close()
+
+    PIL.Image.fromarray(np.ascontiguousarray(tex_map), "RGB").save(
+        os.path.join(fol, "%s.png" % na))
+
+    return
+
+
+class MeshOutput(object):
+
+    def __init__(self,
+                 mesh_v: np.ndarray,
+                 mesh_f: np.ndarray,
+                 vertex_colors: Optional[np.ndarray] = None,
+                 uvs: Optional[np.ndarray] = None,
+                 mesh_tex_idx: Optional[np.ndarray] = None,
+                 tex_map: Optional[np.ndarray] = None):
+
+        self.mesh_v = mesh_v
+        self.mesh_f = mesh_f
+        self.vertex_colors = vertex_colors
+        self.uvs = uvs
+        self.mesh_tex_idx = mesh_tex_idx
+        self.tex_map = tex_map
+
+    def contain_uv_texture(self):
+        return (self.uvs is not None) and (self.mesh_tex_idx is not None) and (self.tex_map is not None)
+
+    def contain_vertex_colors(self):
+        return self.vertex_colors is not None
+
+    def export(self, fname):
+
+        if self.contain_uv_texture():
+            savemeshtes2(
+                self.mesh_v,
+                self.uvs,
+                self.mesh_f,
+                self.mesh_tex_idx,
+                self.tex_map,
+                fname
+            )
+
+        elif self.contain_vertex_colors():
+            mesh_obj = trimesh.Trimesh(vertices=self.mesh_v, faces=self.mesh_f, vertex_colors=self.vertex_colors)
+            mesh_obj.export(fname)
+
+        else:
+            save_obj(
+                self.mesh_v,
+                self.mesh_f,
+                fname
+            )
+
+
+

michelangelo/graphics/primitives/volume.py

@@ -0,0 +1,21 @@
+# -*- coding: utf-8 -*-
+
+import numpy as np
+
+
+def generate_dense_grid_points(bbox_min: np.ndarray,
+                               bbox_max: np.ndarray,
+                               octree_depth: int,
+                               indexing: str = "ij"):
+    length = bbox_max - bbox_min
+    num_cells = np.exp2(octree_depth)
+    x = np.linspace(bbox_min[0], bbox_max[0], int(num_cells) + 1, dtype=np.float32)
+    y = np.linspace(bbox_min[1], bbox_max[1], int(num_cells) + 1, dtype=np.float32)
+    z = np.linspace(bbox_min[2], bbox_max[2], int(num_cells) + 1, dtype=np.float32)
+    [xs, ys, zs] = np.meshgrid(x, y, z, indexing=indexing)
+    xyz = np.stack((xs, ys, zs), axis=-1)
+    xyz = xyz.reshape(-1, 3)
+    grid_size = [int(num_cells) + 1, int(num_cells) + 1, int(num_cells) + 1]
+
+    return xyz, grid_size, length
+

michelangelo/models/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

michelangelo/models/asl_diffusion/__init__.py

@@ -0,0 +1 @@
+# -*- coding: utf-8 -*-

michelangelo/models/asl_diffusion/asl_diffuser_pl_module.py

@@ -0,0 +1,483 @@
+# -*- coding: utf-8 -*-
+
+from omegaconf import DictConfig
+from typing import List, Tuple, Dict, Optional, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.optim import lr_scheduler
+import pytorch_lightning as pl
+from pytorch_lightning.utilities import rank_zero_only
+
+from einops import rearrange
+
+from diffusers.schedulers import (
+    DDPMScheduler,
+    DDIMScheduler,
+    KarrasVeScheduler,
+    DPMSolverMultistepScheduler
+)
+
+from michelangelo.utils import instantiate_from_config
+# from michelangelo.models.tsal.tsal_base import ShapeAsLatentPLModule
+from michelangelo.models.tsal.tsal_base import AlignedShapeAsLatentPLModule
+from michelangelo.models.asl_diffusion.inference_utils import ddim_sample
+
+SchedulerType = Union[DDIMScheduler, KarrasVeScheduler, DPMSolverMultistepScheduler]
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class ASLDiffuser(pl.LightningModule):
+    first_stage_model: Optional[AlignedShapeAsLatentPLModule]
+    # cond_stage_model: Optional[Union[nn.Module, pl.LightningModule]]
+    model: nn.Module
+
+    def __init__(self, *,
+                 first_stage_config,
+                 denoiser_cfg,
+                 scheduler_cfg,
+                 optimizer_cfg,
+                 loss_cfg,
+                 first_stage_key: str = "surface",
+                 cond_stage_key: str = "image",
+                 cond_stage_trainable: bool = True,
+                 scale_by_std: bool = False,
+                 z_scale_factor: float = 1.0,
+                 ckpt_path: Optional[str] = None,
+                 ignore_keys: Union[Tuple[str], List[str]] = ()):
+
+        super().__init__()
+
+        self.first_stage_key = first_stage_key
+        self.cond_stage_key = cond_stage_key
+        self.cond_stage_trainable = cond_stage_trainable
+
+        # 1. initialize first stage. 
+        # Note: the condition model contained in the first stage model.
+        self.first_stage_config = first_stage_config
+        self.first_stage_model = None
+        # self.instantiate_first_stage(first_stage_config)
+
+        # 2. initialize conditional stage
+        # self.instantiate_cond_stage(cond_stage_config)
+        self.cond_stage_model = {
+            "image": self.encode_image,
+            "image_unconditional_embedding": self.empty_img_cond,
+            "text": self.encode_text,
+            "text_unconditional_embedding": self.empty_text_cond,
+            "surface": self.encode_surface,
+            "surface_unconditional_embedding": self.empty_surface_cond,
+        }
+
+        # 3. diffusion model
+        self.model = instantiate_from_config(
+            denoiser_cfg, device=None, dtype=None
+        )
+
+        self.optimizer_cfg = optimizer_cfg
+
+        # 4. scheduling strategy
+        self.scheduler_cfg = scheduler_cfg
+
+        self.noise_scheduler: DDPMScheduler = instantiate_from_config(scheduler_cfg.noise)
+        self.denoise_scheduler: SchedulerType = instantiate_from_config(scheduler_cfg.denoise)
+
+        # 5. loss configures
+        self.loss_cfg = loss_cfg
+
+        self.scale_by_std = scale_by_std
+        if scale_by_std:
+            self.register_buffer("z_scale_factor", torch.tensor(z_scale_factor))
+        else:
+            self.z_scale_factor = z_scale_factor
+
+        self.ckpt_path = ckpt_path
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+
+    def instantiate_first_stage(self, config):
+        model = instantiate_from_config(config)
+        self.first_stage_model = model.eval()
+        self.first_stage_model.train = disabled_train
+        for param in self.first_stage_model.parameters():
+            param.requires_grad = False
+
+        self.first_stage_model = self.first_stage_model.to(self.device)
+
+    # def instantiate_cond_stage(self, config):
+    #     if not self.cond_stage_trainable:
+    #         if config == "__is_first_stage__":
+    #             print("Using first stage also as cond stage.")
+    #             self.cond_stage_model = self.first_stage_model
+    #         elif config == "__is_unconditional__":
+    #             print(f"Training {self.__class__.__name__} as an unconditional model.")
+    #             self.cond_stage_model = None
+    #             # self.be_unconditional = True
+    #         else:
+    #             model = instantiate_from_config(config)
+    #             self.cond_stage_model = model.eval()
+    #             self.cond_stage_model.train = disabled_train
+    #             for param in self.cond_stage_model.parameters():
+    #                 param.requires_grad = False
+    #     else:
+    #         assert config != "__is_first_stage__"
+    #         assert config != "__is_unconditional__"
+    #         model = instantiate_from_config(config)
+    #         self.cond_stage_model = model
+
+    def init_from_ckpt(self, path, ignore_keys=()):
+        state_dict = torch.load(path, map_location="cpu")["state_dict"]
+
+        keys = list(state_dict.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del state_dict[k]
+
+        missing, unexpected = self.load_state_dict(state_dict, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+            print(f"Unexpected Keys: {unexpected}")
+
+    @property
+    def zero_rank(self):
+        if self._trainer:
+            zero_rank = self.trainer.local_rank == 0
+        else:
+            zero_rank = True
+
+        return zero_rank
+
+    def configure_optimizers(self) -> Tuple[List, List]:
+
+        lr = self.learning_rate
+
+        trainable_parameters = list(self.model.parameters())
+        # if the conditional encoder is trainable
+
+        # if self.cond_stage_trainable:
+        #     conditioner_params = [p for p in self.cond_stage_model.parameters() if p.requires_grad]
+        #     trainable_parameters += conditioner_params
+        #     print(f"number of trainable conditional parameters: {len(conditioner_params)}.")
+
+        if self.optimizer_cfg is None:
+            optimizers = [torch.optim.AdamW(trainable_parameters, lr=lr, betas=(0.9, 0.99), weight_decay=1e-3)]
+            schedulers = []
+        else:
+            optimizer = instantiate_from_config(self.optimizer_cfg.optimizer, params=trainable_parameters)
+            scheduler_func = instantiate_from_config(
+                self.optimizer_cfg.scheduler,
+                max_decay_steps=self.trainer.max_steps,
+                lr_max=lr
+            )
+            scheduler = {
+                "scheduler": lr_scheduler.LambdaLR(optimizer, lr_lambda=scheduler_func.schedule),
+                "interval": "step",
+                "frequency": 1
+            }
+            optimizers = [optimizer]
+            schedulers = [scheduler]
+
+        return optimizers, schedulers
+
+    @torch.no_grad()
+    def encode_text(self, text):
+
+        b = text.shape[0]
+        text_tokens = rearrange(text, "b t l -> (b t) l")
+        text_embed = self.first_stage_model.model.encode_text_embed(text_tokens)
+        text_embed = rearrange(text_embed, "(b t) d -> b t d", b=b)
+        text_embed = text_embed.mean(dim=1)
+        text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)
+
+        return text_embed
+
+    @torch.no_grad()
+    def encode_image(self, img):
+
+        return self.first_stage_model.model.encode_image_embed(img)
+
+    @torch.no_grad()
+    def encode_surface(self, surface):
+
+        return self.first_stage_model.model.encode_shape_embed(surface, return_latents=False)
+
+    @torch.no_grad()
+    def empty_text_cond(self, cond):
+
+        return torch.zeros_like(cond, device=cond.device)
+
+    @torch.no_grad()
+    def empty_img_cond(self, cond):
+
+        return torch.zeros_like(cond, device=cond.device)
+
+    @torch.no_grad()
+    def empty_surface_cond(self, cond):
+
+        return torch.zeros_like(cond, device=cond.device)
+
+    @torch.no_grad()
+    def encode_first_stage(self, surface: torch.FloatTensor, sample_posterior=True):
+
+        z_q = self.first_stage_model.encode(surface, sample_posterior)
+        z_q = self.z_scale_factor * z_q
+
+        return z_q
+
+    @torch.no_grad()
+    def decode_first_stage(self, z_q: torch.FloatTensor, **kwargs):
+
+        z_q = 1. / self.z_scale_factor * z_q
+        latents = self.first_stage_model.decode(z_q, **kwargs)
+        return latents
+
+    @rank_zero_only
+    @torch.no_grad()
+    def on_train_batch_start(self, batch, batch_idx):
+        # only for very first batch
+        if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 \
+                and batch_idx == 0 and self.ckpt_path is None:
+            # set rescale weight to 1./std of encodings
+            print("### USING STD-RESCALING ###")
+
+            z_q = self.encode_first_stage(batch[self.first_stage_key])
+            z = z_q.detach()
+
+            del self.z_scale_factor
+            self.register_buffer("z_scale_factor", 1. / z.flatten().std())
+            print(f"setting self.z_scale_factor to {self.z_scale_factor}")
+
+            print("### USING STD-RESCALING ###")
+
+    def compute_loss(self, model_outputs, split):
+        """
+
+        Args:
+            model_outputs (dict):
+                - x_0:
+                - noise:
+                - noise_prior:
+                - noise_pred:
+                - noise_pred_prior:
+
+            split (str):
+
+        Returns:
+
+        """
+
+        pred = model_outputs["pred"]
+
+        if self.noise_scheduler.prediction_type == "epsilon":
+            target = model_outputs["noise"]
+        elif self.noise_scheduler.prediction_type == "sample":
+            target = model_outputs["x_0"]
+        else:
+            raise NotImplementedError(f"Prediction Type: {self.noise_scheduler.prediction_type} not yet supported.")
+
+        if self.loss_cfg.loss_type == "l1":
+            simple = F.l1_loss(pred, target, reduction="mean")
+        elif self.loss_cfg.loss_type in ["mse", "l2"]:
+            simple = F.mse_loss(pred, target, reduction="mean")
+        else:
+            raise NotImplementedError(f"Loss Type: {self.loss_cfg.loss_type} not yet supported.")
+
+        total_loss = simple
+
+        loss_dict = {
+            f"{split}/total_loss": total_loss.clone().detach(),
+            f"{split}/simple": simple.detach(),
+        }
+
+        return total_loss, loss_dict
+
+    def forward(self, batch):
+        """
+
+        Args:
+            batch:
+
+        Returns:
+
+        """
+
+        if self.first_stage_model is None:
+            self.instantiate_first_stage(self.first_stage_config)
+
+        latents = self.encode_first_stage(batch[self.first_stage_key])
+
+        # conditions = self.cond_stage_model.encode(batch[self.cond_stage_key])
+
+        conditions = self.cond_stage_model[self.cond_stage_key](batch[self.cond_stage_key]).unsqueeze(1)
+
+        mask = torch.rand((len(conditions), 1, 1), device=conditions.device, dtype=conditions.dtype) >= 0.1
+        conditions = conditions * mask.to(conditions)
+
+        # Sample noise that we"ll add to the latents
+        # [batch_size, n_token, latent_dim]
+        noise = torch.randn_like(latents)
+        bs = latents.shape[0]
+        # Sample a random timestep for each motion
+        timesteps = torch.randint(
+            0,
+            self.noise_scheduler.config.num_train_timesteps,
+            (bs,),
+            device=latents.device,
+        )
+        timesteps = timesteps.long()
+        # Add noise to the latents according to the noise magnitude at each timestep
+        noisy_z = self.noise_scheduler.add_noise(latents, noise, timesteps)
+
+        # diffusion model forward
+        noise_pred = self.model(noisy_z, timesteps, conditions)
+
+        diffusion_outputs = {
+            "x_0": noisy_z,
+            "noise": noise,
+            "pred": noise_pred
+        }
+
+        return diffusion_outputs
+
+    def training_step(self, batch: Dict[str, Union[torch.FloatTensor, List[str]]],
+                      batch_idx: int, optimizer_idx: int = 0) -> torch.FloatTensor:
+        """
+
+        Args:
+            batch (dict): the batch sample, and it contains:
+                - surface (torch.FloatTensor):
+                - image (torch.FloatTensor): if provide, [bs, 3, h, w], item range [0, 1]
+                - depth (torch.FloatTensor): if provide, [bs, 1, h, w], item range [-1, 1]
+                - normal (torch.FloatTensor): if provide, [bs, 3, h, w], item range [-1, 1]
+                - text (list of str):
+
+            batch_idx (int):
+
+            optimizer_idx (int):
+
+        Returns:
+            loss (torch.FloatTensor):
+
+        """
+
+        diffusion_outputs = self(batch)
+
+        loss, loss_dict = self.compute_loss(diffusion_outputs, "train")
+        self.log_dict(loss_dict, prog_bar=True, logger=True, sync_dist=False, rank_zero_only=True)
+
+        return loss
+
+    def validation_step(self, batch: Dict[str, torch.FloatTensor],
+                        batch_idx: int, optimizer_idx: int = 0) -> torch.FloatTensor:
+        """
+
+        Args:
+            batch (dict): the batch sample, and it contains:
+                - surface_pc (torch.FloatTensor): [n_pts, 4]
+                - surface_feats (torch.FloatTensor): [n_pts, c]
+                - text (list of str):
+
+            batch_idx (int):
+
+            optimizer_idx (int):
+
+        Returns:
+            loss (torch.FloatTensor):
+
+        """
+
+        diffusion_outputs = self(batch)
+
+        loss, loss_dict = self.compute_loss(diffusion_outputs, "val")
+        self.log_dict(loss_dict, prog_bar=True, logger=True, sync_dist=False, rank_zero_only=True)
+
+        return loss
+
+    @torch.no_grad()
+    def sample(self,
+               batch: Dict[str, Union[torch.FloatTensor, List[str]]],
+               sample_times: int = 1,
+               steps: Optional[int] = None,
+               guidance_scale: Optional[float] = None,
+               eta: float = 0.0,
+               return_intermediates: bool = False, **kwargs):
+
+        if self.first_stage_model is None:
+            self.instantiate_first_stage(self.first_stage_config)
+
+        if steps is None:
+            steps = self.scheduler_cfg.num_inference_steps
+
+        if guidance_scale is None:
+            guidance_scale = self.scheduler_cfg.guidance_scale
+        do_classifier_free_guidance = guidance_scale > 0
+
+        # conditional encode
+        xc = batch[self.cond_stage_key]
+        # cond = self.cond_stage_model[self.cond_stage_key](xc)
+        cond = self.cond_stage_model[self.cond_stage_key](xc).unsqueeze(1)
+
+        if do_classifier_free_guidance:
+            """
+            Note: There are two kinds of uncond for text. 
+            1: using "" as uncond text; (in SAL diffusion)
+            2: zeros_like(cond) as uncond text; (in MDM)
+            """
+            # un_cond = self.cond_stage_model.unconditional_embedding(batch_size=len(xc))
+            un_cond = self.cond_stage_model[f"{self.cond_stage_key}_unconditional_embedding"](cond)
+            # un_cond = torch.zeros_like(cond, device=cond.device)
+            cond = torch.cat([un_cond, cond], dim=0)
+
+        outputs = []
+        latents = None
+
+        if not return_intermediates:
+            for _ in range(sample_times):
+                sample_loop = ddim_sample(
+                    self.denoise_scheduler,
+                    self.model,
+                    shape=self.first_stage_model.latent_shape,
+                    cond=cond,
+                    steps=steps,
+                    guidance_scale=guidance_scale,
+                    do_classifier_free_guidance=do_classifier_free_guidance,
+                    device=self.device,
+                    eta=eta,
+                    disable_prog=not self.zero_rank
+                )
+                for sample, t in sample_loop:
+                    latents = sample
+                outputs.append(self.decode_first_stage(latents, **kwargs))
+        else:
+
+            sample_loop = ddim_sample(
+                self.denoise_scheduler,
+                self.model,
+                shape=self.first_stage_model.latent_shape,
+                cond=cond,
+                steps=steps,
+                guidance_scale=guidance_scale,
+                do_classifier_free_guidance=do_classifier_free_guidance,
+                device=self.device,
+                eta=eta,
+                disable_prog=not self.zero_rank
+            )
+
+            iter_size = steps // sample_times
+            i = 0
+            for sample, t in sample_loop:
+                latents = sample
+                if i % iter_size == 0 or i == steps - 1:
+                    outputs.append(self.decode_first_stage(latents, **kwargs))
+                i += 1
+
+        return outputs

michelangelo/models/asl_diffusion/asl_udt.py

@@ -0,0 +1,104 @@
+# -*- coding: utf-8 -*-
+
+import torch
+import torch.nn as nn
+from typing import Optional
+from diffusers.models.embeddings import Timesteps
+import math
+
+from michelangelo.models.modules.transformer_blocks import MLP
+from michelangelo.models.modules.diffusion_transformer import UNetDiffusionTransformer
+
+
+class ConditionalASLUDTDenoiser(nn.Module):
+
+    def __init__(self, *,
+                 device: Optional[torch.device],
+                 dtype: Optional[torch.dtype],
+                 input_channels: int,
+                 output_channels: int,
+                 n_ctx: int,
+                 width: int,
+                 layers: int,
+                 heads: int,
+                 context_dim: int,
+                 context_ln: bool = True,
+                 skip_ln: bool = False,
+                 init_scale: float = 0.25,
+                 flip_sin_to_cos: bool = False,
+                 use_checkpoint: bool = False):
+        super().__init__()
+
+        self.use_checkpoint = use_checkpoint
+
+        init_scale = init_scale * math.sqrt(1.0 / width)
+
+        self.backbone = UNetDiffusionTransformer(
+            device=device,
+            dtype=dtype,
+            n_ctx=n_ctx,
+            width=width,
+            layers=layers,
+            heads=heads,
+            skip_ln=skip_ln,
+            init_scale=init_scale,
+            use_checkpoint=use_checkpoint
+        )
+        self.ln_post = nn.LayerNorm(width, device=device, dtype=dtype)
+        self.input_proj = nn.Linear(input_channels, width, device=device, dtype=dtype)
+        self.output_proj = nn.Linear(width, output_channels, device=device, dtype=dtype)
+
+        # timestep embedding
+        self.time_embed = Timesteps(width, flip_sin_to_cos=flip_sin_to_cos, downscale_freq_shift=0)
+        self.time_proj = MLP(
+            device=device, dtype=dtype, width=width, init_scale=init_scale
+        )
+
+        self.context_embed = nn.Sequential(
+            nn.LayerNorm(context_dim, device=device, dtype=dtype),
+            nn.Linear(context_dim, width, device=device, dtype=dtype),
+        )
+
+        if context_ln:
+            self.context_embed = nn.Sequential(
+                nn.LayerNorm(context_dim, device=device, dtype=dtype),
+                nn.Linear(context_dim, width, device=device, dtype=dtype),
+            )
+        else:
+            self.context_embed = nn.Linear(context_dim, width, device=device, dtype=dtype)
+
+    def forward(self,
+                model_input: torch.FloatTensor,
+                timestep: torch.LongTensor,
+                context: torch.FloatTensor):
+
+        r"""
+        Args:
+            model_input (torch.FloatTensor): [bs, n_data, c]
+            timestep (torch.LongTensor): [bs,]
+            context (torch.FloatTensor): [bs, context_tokens, c]
+
+        Returns:
+            sample (torch.FloatTensor): [bs, n_data, c]
+
+        """
+
+        _, n_data, _ = model_input.shape
+
+        # 1. time
+        t_emb = self.time_proj(self.time_embed(timestep)).unsqueeze(dim=1)
+
+        # 2. conditions projector
+        context = self.context_embed(context)
+
+        # 3. denoiser
+        x = self.input_proj(model_input)
+        x = torch.cat([t_emb, context, x], dim=1)
+        x = self.backbone(x)
+        x = self.ln_post(x)
+        x = x[:, -n_data:]
+        sample = self.output_proj(x)
+
+        return sample
+
+

michelangelo/models/asl_diffusion/base.py

@@ -0,0 +1,13 @@
+# -*- coding: utf-8 -*-
+
+import torch
+import torch.nn as nn
+
+
+class BaseDenoiser(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x, t, context):
+        raise NotImplementedError

michelangelo/models/asl_diffusion/clip_asl_diffuser_pl_module.py

@@ -0,0 +1,393 @@
+# -*- coding: utf-8 -*-
+
+from omegaconf import DictConfig
+from typing import List, Tuple, Dict, Optional, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.optim import lr_scheduler
+import pytorch_lightning as pl
+from pytorch_lightning.utilities import rank_zero_only
+
+from diffusers.schedulers import (
+    DDPMScheduler,
+    DDIMScheduler,
+    KarrasVeScheduler,
+    DPMSolverMultistepScheduler
+)
+
+from michelangelo.utils import instantiate_from_config
+from michelangelo.models.tsal.tsal_base import AlignedShapeAsLatentPLModule
+from michelangelo.models.asl_diffusion.inference_utils import ddim_sample
+
+SchedulerType = Union[DDIMScheduler, KarrasVeScheduler, DPMSolverMultistepScheduler]
+
+
+def disabled_train(self, mode=True):
+    """Overwrite model.train with this function to make sure train/eval mode
+    does not change anymore."""
+    return self
+
+
+class ClipASLDiffuser(pl.LightningModule):
+    first_stage_model: Optional[AlignedShapeAsLatentPLModule]
+    cond_stage_model: Optional[Union[nn.Module, pl.LightningModule]]
+    model: nn.Module
+
+    def __init__(self, *,
+                 first_stage_config,
+                 cond_stage_config,
+                 denoiser_cfg,
+                 scheduler_cfg,
+                 optimizer_cfg,
+                 loss_cfg,
+                 first_stage_key: str = "surface",
+                 cond_stage_key: str = "image",
+                 scale_by_std: bool = False,
+                 z_scale_factor: float = 1.0,
+                 ckpt_path: Optional[str] = None,
+                 ignore_keys: Union[Tuple[str], List[str]] = ()):
+
+        super().__init__()
+
+        self.first_stage_key = first_stage_key
+        self.cond_stage_key = cond_stage_key
+
+        # 1. lazy initialize first stage
+        self.instantiate_first_stage(first_stage_config)
+
+        # 2. initialize conditional stage
+        self.instantiate_cond_stage(cond_stage_config)
+
+        # 3. diffusion model
+        self.model = instantiate_from_config(
+            denoiser_cfg, device=None, dtype=None
+        )
+
+        self.optimizer_cfg = optimizer_cfg
+
+        # 4. scheduling strategy
+        self.scheduler_cfg = scheduler_cfg
+
+        self.noise_scheduler: DDPMScheduler = instantiate_from_config(scheduler_cfg.noise)
+        self.denoise_scheduler: SchedulerType = instantiate_from_config(scheduler_cfg.denoise)
+
+        # 5. loss configures
+        self.loss_cfg = loss_cfg
+
+        self.scale_by_std = scale_by_std
+        if scale_by_std:
+            self.register_buffer("z_scale_factor", torch.tensor(z_scale_factor))
+        else:
+            self.z_scale_factor = z_scale_factor
+
+        self.ckpt_path = ckpt_path
+        if ckpt_path is not None:
+            self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys)
+
+    def instantiate_non_trainable_model(self, config):
+        model = instantiate_from_config(config)
+        model = model.eval()
+        model.train = disabled_train
+        for param in model.parameters():
+            param.requires_grad = False
+
+        return model
+
+    def instantiate_first_stage(self, first_stage_config):
+        self.first_stage_model = self.instantiate_non_trainable_model(first_stage_config)
+        self.first_stage_model.set_shape_model_only()
+
+    def instantiate_cond_stage(self, cond_stage_config):
+        self.cond_stage_model = self.instantiate_non_trainable_model(cond_stage_config)
+
+    def init_from_ckpt(self, path, ignore_keys=()):
+        state_dict = torch.load(path, map_location="cpu")["state_dict"]
+
+        keys = list(state_dict.keys())
+        for k in keys:
+            for ik in ignore_keys:
+                if k.startswith(ik):
+                    print("Deleting key {} from state_dict.".format(k))
+                    del state_dict[k]
+
+        missing, unexpected = self.load_state_dict(state_dict, strict=False)
+        print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys")
+        if len(missing) > 0:
+            print(f"Missing Keys: {missing}")
+            print(f"Unexpected Keys: {unexpected}")
+
+    @property
+    def zero_rank(self):
+        if self._trainer:
+            zero_rank = self.trainer.local_rank == 0
+        else:
+            zero_rank = True
+
+        return zero_rank
+
+    def configure_optimizers(self) -> Tuple[List, List]:
+
+        lr = self.learning_rate
+
+        trainable_parameters = list(self.model.parameters())
+        if self.optimizer_cfg is None:
+            optimizers = [torch.optim.AdamW(trainable_parameters, lr=lr, betas=(0.9, 0.99), weight_decay=1e-3)]
+            schedulers = []
+        else:
+            optimizer = instantiate_from_config(self.optimizer_cfg.optimizer, params=trainable_parameters)
+            scheduler_func = instantiate_from_config(
+                self.optimizer_cfg.scheduler,
+                max_decay_steps=self.trainer.max_steps,
+                lr_max=lr
+            )
+            scheduler = {
+                "scheduler": lr_scheduler.LambdaLR(optimizer, lr_lambda=scheduler_func.schedule),
+                "interval": "step",
+                "frequency": 1
+            }
+            optimizers = [optimizer]
+            schedulers = [scheduler]
+
+        return optimizers, schedulers
+
+    @torch.no_grad()
+    def encode_first_stage(self, surface: torch.FloatTensor, sample_posterior=True):
+
+        z_q = self.first_stage_model.encode(surface, sample_posterior)
+        z_q = self.z_scale_factor * z_q
+
+        return z_q
+
+    @torch.no_grad()
+    def decode_first_stage(self, z_q: torch.FloatTensor, **kwargs):
+
+        z_q = 1. / self.z_scale_factor * z_q
+        latents = self.first_stage_model.decode(z_q, **kwargs)
+        return latents
+
+    @rank_zero_only
+    @torch.no_grad()
+    def on_train_batch_start(self, batch, batch_idx):
+        # only for very first batch
+        if self.scale_by_std and self.current_epoch == 0 and self.global_step == 0 \
+                and batch_idx == 0 and self.ckpt_path is None:
+            # set rescale weight to 1./std of encodings
+            print("### USING STD-RESCALING ###")
+
+            z_q = self.encode_first_stage(batch[self.first_stage_key])
+            z = z_q.detach()
+
+            del self.z_scale_factor
+            self.register_buffer("z_scale_factor", 1. / z.flatten().std())
+            print(f"setting self.z_scale_factor to {self.z_scale_factor}")
+
+            print("### USING STD-RESCALING ###")
+
+    def compute_loss(self, model_outputs, split):
+        """
+
+        Args:
+            model_outputs (dict):
+                - x_0:
+                - noise:
+                - noise_prior:
+                - noise_pred:
+                - noise_pred_prior:
+
+            split (str):
+
+        Returns:
+
+        """
+
+        pred = model_outputs["pred"]
+
+        if self.noise_scheduler.prediction_type == "epsilon":
+            target = model_outputs["noise"]
+        elif self.noise_scheduler.prediction_type == "sample":
+            target = model_outputs["x_0"]
+        else:
+            raise NotImplementedError(f"Prediction Type: {self.noise_scheduler.prediction_type} not yet supported.")
+
+        if self.loss_cfg.loss_type == "l1":
+            simple = F.l1_loss(pred, target, reduction="mean")
+        elif self.loss_cfg.loss_type in ["mse", "l2"]:
+            simple = F.mse_loss(pred, target, reduction="mean")
+        else:
+            raise NotImplementedError(f"Loss Type: {self.loss_cfg.loss_type} not yet supported.")
+
+        total_loss = simple
+
+        loss_dict = {
+            f"{split}/total_loss": total_loss.clone().detach(),
+            f"{split}/simple": simple.detach(),
+        }
+
+        return total_loss, loss_dict
+
+    def forward(self, batch):
+        """
+
+        Args:
+            batch:
+
+        Returns:
+
+        """
+
+        latents = self.encode_first_stage(batch[self.first_stage_key])
+        conditions = self.cond_stage_model.encode(batch[self.cond_stage_key])
+
+        # Sample noise that we"ll add to the latents
+        # [batch_size, n_token, latent_dim]
+        noise = torch.randn_like(latents)
+        bs = latents.shape[0]
+        # Sample a random timestep for each motion
+        timesteps = torch.randint(
+            0,
+            self.noise_scheduler.config.num_train_timesteps,
+            (bs,),
+            device=latents.device,
+        )
+        timesteps = timesteps.long()
+        # Add noise to the latents according to the noise magnitude at each timestep
+        noisy_z = self.noise_scheduler.add_noise(latents, noise, timesteps)
+
+        # diffusion model forward
+        noise_pred = self.model(noisy_z, timesteps, conditions)
+
+        diffusion_outputs = {
+            "x_0": noisy_z,
+            "noise": noise,
+            "pred": noise_pred
+        }
+
+        return diffusion_outputs
+
+    def training_step(self, batch: Dict[str, Union[torch.FloatTensor, List[str]]],
+                      batch_idx: int, optimizer_idx: int = 0) -> torch.FloatTensor:
+        """
+
+        Args:
+            batch (dict): the batch sample, and it contains:
+                - surface (torch.FloatTensor):
+                - image (torch.FloatTensor): if provide, [bs, 3, h, w], item range [0, 1]
+                - depth (torch.FloatTensor): if provide, [bs, 1, h, w], item range [-1, 1]
+                - normal (torch.FloatTensor): if provide, [bs, 3, h, w], item range [-1, 1]
+                - text (list of str):
+
+            batch_idx (int):
+
+            optimizer_idx (int):
+
+        Returns:
+            loss (torch.FloatTensor):
+
+        """
+
+        diffusion_outputs = self(batch)
+
+        loss, loss_dict = self.compute_loss(diffusion_outputs, "train")
+        self.log_dict(loss_dict, prog_bar=True, logger=True, sync_dist=False, rank_zero_only=True)
+
+        return loss
+
+    def validation_step(self, batch: Dict[str, torch.FloatTensor],
+                        batch_idx: int, optimizer_idx: int = 0) -> torch.FloatTensor:
+        """
+
+        Args:
+            batch (dict): the batch sample, and it contains:
+                - surface_pc (torch.FloatTensor): [n_pts, 4]
+                - surface_feats (torch.FloatTensor): [n_pts, c]
+                - text (list of str):
+
+            batch_idx (int):
+
+            optimizer_idx (int):
+
+        Returns:
+            loss (torch.FloatTensor):
+
+        """
+
+        diffusion_outputs = self(batch)
+
+        loss, loss_dict = self.compute_loss(diffusion_outputs, "val")
+        self.log_dict(loss_dict, prog_bar=True, logger=True, sync_dist=False, rank_zero_only=True)
+
+        return loss
+
+    @torch.no_grad()
+    def sample(self,
+               batch: Dict[str, Union[torch.FloatTensor, List[str]]],
+               sample_times: int = 1,
+               steps: Optional[int] = None,
+               guidance_scale: Optional[float] = None,
+               eta: float = 0.0,
+               return_intermediates: bool = False, **kwargs):
+
+        if steps is None:
+            steps = self.scheduler_cfg.num_inference_steps
+
+        if guidance_scale is None:
+            guidance_scale = self.scheduler_cfg.guidance_scale
+        do_classifier_free_guidance = guidance_scale > 0
+
+        # conditional encode
+        xc = batch[self.cond_stage_key]
+
+        # print(self.first_stage_model.device, self.cond_stage_model.device, self.device)
+
+        cond = self.cond_stage_model(xc)
+
+        if do_classifier_free_guidance:
+            un_cond = self.cond_stage_model.unconditional_embedding(batch_size=len(xc))
+            cond = torch.cat([un_cond, cond], dim=0)
+
+        outputs = []
+        latents = None
+
+        if not return_intermediates:
+            for _ in range(sample_times):
+                sample_loop = ddim_sample(
+                    self.denoise_scheduler,
+                    self.model,
+                    shape=self.first_stage_model.latent_shape,
+                    cond=cond,
+                    steps=steps,
+                    guidance_scale=guidance_scale,
+                    do_classifier_free_guidance=do_classifier_free_guidance,
+                    device=self.device,
+                    eta=eta,
+                    disable_prog=not self.zero_rank
+                )
+                for sample, t in sample_loop:
+                    latents = sample
+                outputs.append(self.decode_first_stage(latents, **kwargs))
+        else:
+
+            sample_loop = ddim_sample(
+                self.denoise_scheduler,
+                self.model,
+                shape=self.first_stage_model.latent_shape,
+                cond=cond,
+                steps=steps,
+                guidance_scale=guidance_scale,
+                do_classifier_free_guidance=do_classifier_free_guidance,
+                device=self.device,
+                eta=eta,
+                disable_prog=not self.zero_rank
+            )
+
+            iter_size = steps // sample_times
+            i = 0
+            for sample, t in sample_loop:
+                latents = sample
+                if i % iter_size == 0 or i == steps - 1:
+                    outputs.append(self.decode_first_stage(latents, **kwargs))
+                i += 1
+
+        return outputs

michelangelo/models/asl_diffusion/inference_utils.py

@@ -0,0 +1,80 @@
+# -*- coding: utf-8 -*-
+
+import torch
+from tqdm import tqdm
+from typing import Tuple, List, Union, Optional
+from diffusers.schedulers import DDIMScheduler
+
+
+__all__ = ["ddim_sample"]
+
+
+def ddim_sample(ddim_scheduler: DDIMScheduler,
+                diffusion_model: torch.nn.Module,
+                shape: Union[List[int], Tuple[int]],
+                cond: torch.FloatTensor,
+                steps: int,
+                eta: float = 0.0,
+                guidance_scale: float = 3.0,
+                do_classifier_free_guidance: bool = True,
+                generator: Optional[torch.Generator] = None,
+                device: torch.device = "cuda:0",
+                disable_prog: bool = True):
+
+    assert steps > 0, f"{steps} must > 0."
+
+    # init latents
+    bsz = cond.shape[0]
+    if do_classifier_free_guidance:
+        bsz = bsz // 2
+
+    latents = torch.randn(
+        (bsz, *shape),
+        generator=generator,
+        device=cond.device,
+        dtype=cond.dtype,
+    )
+    # scale the initial noise by the standard deviation required by the scheduler
+    latents = latents * ddim_scheduler.init_noise_sigma
+    # set timesteps
+    ddim_scheduler.set_timesteps(steps)
+    timesteps = ddim_scheduler.timesteps.to(device)
+    # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+    # eta (η) is only used with the DDIMScheduler, and between [0, 1]
+    extra_step_kwargs = {
+        "eta": eta,
+        "generator": generator
+    }
+
+    # reverse
+    for i, t in enumerate(tqdm(timesteps, disable=disable_prog, desc="DDIM Sampling:", leave=False)):
+        # expand the latents if we are doing classifier free guidance
+        latent_model_input = (
+            torch.cat([latents] * 2)
+            if do_classifier_free_guidance
+            else latents
+        )
+        # latent_model_input = scheduler.scale_model_input(latent_model_input, t)
+        # predict the noise residual
+        timestep_tensor = torch.tensor([t], dtype=torch.long, device=device)
+        timestep_tensor = timestep_tensor.expand(latent_model_input.shape[0])
+        noise_pred = diffusion_model.forward(latent_model_input, timestep_tensor, cond)
+
+        # perform guidance
+        if do_classifier_free_guidance:
+            noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+            noise_pred = noise_pred_uncond + guidance_scale * (
+                    noise_pred_text - noise_pred_uncond
+            )
+            # text_embeddings_for_guidance = encoder_hidden_states.chunk(
+            #     2)[1] if do_classifier_free_guidance else encoder_hidden_states
+        # compute the previous noisy sample x_t -> x_t-1
+        latents = ddim_scheduler.step(
+            noise_pred, t, latents, **extra_step_kwargs
+        ).prev_sample
+
+        yield latents, t
+
+
+def karra_sample():
+    pass

michelangelo/models/conditional_encoders/__init__.py

@@ -0,0 +1,3 @@
+# -*- coding: utf-8 -*-
+
+from .clip import CLIPEncoder

michelangelo/models/conditional_encoders/clip.py

@@ -0,0 +1,89 @@
+# -*- coding: utf-8 -*-
+
+import torch
+import numpy as np
+from PIL import Image
+from dataclasses import dataclass
+from torchvision.transforms import Normalize
+from transformers import CLIPModel, CLIPTokenizer
+from transformers.utils import ModelOutput
+from typing import Iterable, Optional, Union, List
+
+
+ImageType = Union[np.ndarray, torch.Tensor, Image.Image]
+
+
+@dataclass
+class CLIPEmbedOutput(ModelOutput):
+    last_hidden_state: torch.FloatTensor = None
+    pooler_output: torch.FloatTensor = None
+    embeds: torch.FloatTensor = None
+
+
+class CLIPEncoder(torch.nn.Module):
+
+    def __init__(self, model_path="openai/clip-vit-base-patch32"):
+
+        super().__init__()
+
+        # Load the CLIP model and processor
+        self.model: CLIPModel = CLIPModel.from_pretrained(model_path)
+        self.tokenizer = CLIPTokenizer.from_pretrained(model_path)
+        self.image_preprocess = Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
+
+        self.model.training = False
+        for p in self.model.parameters():
+            p.requires_grad = False
+
+    @torch.no_grad()
+    def encode_image(self, images: Iterable[Optional[ImageType]]):
+        pixel_values = self.image_preprocess(images)
+
+        vision_outputs = self.model.vision_model(pixel_values=pixel_values)
+
+        pooler_output = vision_outputs[1]  # pooled_output
+        image_features = self.model.visual_projection(pooler_output)
+
+        visual_embeds = CLIPEmbedOutput(
+            last_hidden_state=vision_outputs.last_hidden_state,
+            pooler_output=pooler_output,
+            embeds=image_features
+        )
+
+        return visual_embeds
+
+    @torch.no_grad()
+    def encode_text(self, texts: List[str]):
+        text_inputs = self.tokenizer(texts, padding=True, return_tensors="pt")
+
+        text_outputs = self.model.text_model(input_ids=text_inputs)
+
+        pooler_output = text_outputs[1]  # pooled_output
+        text_features = self.model.text_projection(pooler_output)
+
+        text_embeds = CLIPEmbedOutput(
+            last_hidden_state=text_outputs.last_hidden_state,
+            pooler_output=pooler_output,
+            embeds=text_features
+        )
+
+        return text_embeds
+
+    def forward(self,
+                images: Iterable[Optional[ImageType]],
+                texts: List[str]):
+
+        visual_embeds = self.encode_image(images)
+        text_embeds = self.encode_text(texts)
+
+        return visual_embeds, text_embeds
+
+
+
+
+
+
+
+
+
+

michelangelo/models/conditional_encoders/encoder_factory.py

@@ -0,0 +1,562 @@
+# -*- coding: utf-8 -*-
+import os
+
+import torch
+import torch.nn as nn
+from torchvision import transforms
+from transformers import CLIPModel, CLIPTokenizer
+from collections import OrderedDict
+
+from michelangelo.data.transforms import RandomResize
+
+
+class AbstractEncoder(nn.Module):
+    embedding_dim: int
+
+    def __init__(self):
+        super().__init__()
+
+    def encode(self, *args, **kwargs):
+        raise NotImplementedError
+
+
+class ClassEmbedder(nn.Module):
+    def __init__(self, embed_dim, n_classes=1000, key="class"):
+        super().__init__()
+        self.key = key
+        self.embedding = nn.Embedding(n_classes, embed_dim)
+
+    def forward(self, batch, key=None):
+        if key is None:
+            key = self.key
+        # this is for use in crossattn
+        c = batch[key][:, None]
+        c = self.embedding(c)
+        return c
+
+
+class FrozenCLIPTextEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from Hugging Face)"""
+
+    def __init__(
+        self,
+        version="openai/clip-vit-large-patch14",
+        tokenizer_version=None,
+        device="cuda",
+        max_length=77,
+        zero_embedding_radio: float = 0.1,
+    ):
+        super().__init__()
+        self.tokenizer = CLIPTokenizer.from_pretrained(tokenizer_version or version)
+
+        self.device = device
+        self.max_length = max_length
+        self.zero_embedding_radio = zero_embedding_radio
+
+        self.clip_dict = OrderedDict()
+        self.clip_name = os.path.split(version)[-1]
+
+        transformer = CLIPModel.from_pretrained(version).text_model
+
+        for param in transformer.parameters():
+            param.requires_grad = False
+        self.clip_dict[self.clip_name] = transformer
+
+        self._move_flag = False
+
+    @property
+    def clip(self):
+        return self.clip_dict[self.clip_name]
+
+    def move(self):
+        if self._move_flag:
+            return
+
+        self.clip_dict[self.clip_name] = self.clip_dict[self.clip_name].to(self.device)
+        self._move_flag = True
+
+    def unconditional_embedding(self, batch_size):
+        empty_text = [""] * batch_size
+        empty_z = self.forward(empty_text)
+        return empty_z
+
+    def forward(self, text):
+        self.move()
+
+        batch_encoding = self.tokenizer(
+            text,
+            truncation=True,
+            max_length=self.max_length,
+            return_length=True,
+            return_overflowing_tokens=False,
+            padding="max_length",
+            return_tensors="pt",
+        )
+
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.clip(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        batch_size = len(text)
+        batch_mask = torch.rand((batch_size,))
+        for i in range(batch_size):
+            if batch_mask[i] < self.zero_embedding_radio:
+                text[i] = ""
+
+        return self(text)
+
+class FrozenAlignedCLIPTextEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from Hugging Face)"""
+
+    def __init__(
+        self,
+        version="openai/clip-vit-large-patch14",
+        tokenizer_version=None,
+        device="cuda",
+        max_length=77,
+        zero_embedding_radio: float = 0.1,
+    ):
+        super().__init__()
+        self.tokenizer = CLIPTokenizer.from_pretrained(tokenizer_version or version)
+
+        self.device = device
+        self.max_length = max_length
+        self.zero_embedding_radio = zero_embedding_radio
+
+        self.clip_dict = OrderedDict()
+        self.clip_name = os.path.split(version)[-1]
+
+        transformer = CLIPModel.from_pretrained(version).text_model
+
+        for param in transformer.parameters():
+            param.requires_grad = False
+        self.clip_dict[self.clip_name] = transformer
+
+        self._move_flag = False
+
+    @property
+    def clip(self):
+        return self.clip_dict[self.clip_name]
+
+    def move(self):
+        if self._move_flag:
+            return
+
+        self.clip_dict[self.clip_name] = self.clip_dict[self.clip_name].to(self.device)
+        self._move_flag = True
+
+    def unconditional_embedding(self, batch_size):
+        empty_text = [""] * batch_size
+        empty_z = self.forward(empty_text)
+        return empty_z
+
+    def forward(self, text):
+        self.move()
+
+        batch_encoding = self.tokenizer(
+            text,
+            truncation=True,
+            max_length=self.max_length,
+            return_length=True,
+            return_overflowing_tokens=False,
+            padding="max_length",
+            return_tensors="pt",
+        )
+
+        tokens = batch_encoding["input_ids"].to(self.device)
+        outputs = self.clip(input_ids=tokens)
+
+        z = outputs.last_hidden_state
+        return z
+
+    def encode(self, text):
+        batch_size = len(text)
+        batch_mask = torch.rand((batch_size,))
+        for i in range(batch_size):
+            if batch_mask[i] < self.zero_embedding_radio:
+                text[i] = ""
+
+        return self(text)
+
+
+class FrozenCLIPImageEmbedder(AbstractEncoder):
+    """Uses the CLIP transformer encoder for text (from Hugging Face)"""
+
+    def __init__(
+            self,
+            version="openai/clip-vit-large-patch14",
+            device="cuda",
+            zero_embedding_radio=0.1,
+            normalize_embedding=True,
+            num_projection_vector=0,
+            linear_mapping_bias=True,
+            reverse_visual_projection=False,
+    ):
+        super().__init__()
+
+        self.device = device
+
+        self.clip_dict = OrderedDict()
+        self.clip_name = os.path.split(version)[-1]
+
+        clip_model = CLIPModel.from_pretrained(version)
+        clip_model.text_model = None
+        clip_model.text_projection = None
+        clip_model = clip_model.eval()
+        for param in self.parameters():
+            param.requires_grad = False
+        self.clip_dict[self.clip_name] = clip_model
+
+        self.transform = transforms.Compose(
+            [
+                transforms.Resize(224, transforms.InterpolationMode.BICUBIC, antialias=True),
+                transforms.CenterCrop(224),  # crop a (224, 224) square
+                transforms.Normalize(
+                    mean=[0.48145466, 0.4578275, 0.40821073],
+                    std=[0.26862954, 0.26130258, 0.27577711],
+                ),
+            ]
+        )
+        self.zero_embedding_radio = zero_embedding_radio
+
+        self.num_projection_vector = num_projection_vector
+        self.reverse_visual_projection = reverse_visual_projection
+        self.normalize_embedding = normalize_embedding
+
+        embedding_dim = (
+            clip_model.visual_projection.in_features
+            if reverse_visual_projection
+            else clip_model.visual_projection.out_features
+        )
+        self.embedding_dim = embedding_dim
+        if self.num_projection_vector > 0:
+            self.projection = nn.Linear(
+                embedding_dim,
+                clip_model.visual_projection.out_features * num_projection_vector,
+                bias=linear_mapping_bias,
+            )
+            nn.init.normal_(self.projection.weight, std=embedding_dim ** -0.5)
+
+        self._move_flag = False
+
+    @property
+    def clip(self):
+        return self.clip_dict[self.clip_name]
+
+    def unconditional_embedding(self, batch_size):
+        zero = torch.zeros(
+            batch_size,
+            1,
+            self.embedding_dim,
+            device=self.device,
+            dtype=self.clip.visual_projection.weight.dtype,
+        )
+        if self.num_projection_vector > 0:
+            zero = self.projection(zero).view(batch_size, self.num_projection_vector, -1)
+        return zero
+
+    def forward(self, image, value_range=(-1, 1), zero_embedding_radio=0):
+        if value_range is not None:
+            low, high = value_range
+            image = (image - low) / (high - low)
+
+        image = image.to(self.device, dtype=self.clip.visual_projection.weight.dtype)
+
+        if self.reverse_visual_projection:
+            z = self.clip.vision_model(self.transform(image))[1]
+        else:
+            z = self.clip.get_image_features(self.transform(image))
+
+        if self.normalize_embedding:
+            z = z / z.norm(dim=-1, keepdim=True)
+        if z.ndim == 2:
+            z = z.unsqueeze(dim=-2)
+
+        if zero_embedding_radio > 0:
+            mask = torch.rand((len(image), 1, 1), device=z.device, dtype=z.dtype) < zero_embedding_radio
+            z = z * mask.to(z)
+
+        if self.num_projection_vector > 0:
+            z = self.projection(z).view(len(image), self.num_projection_vector, -1)
+
+        return z
+
+    def move(self):
+        if self._move_flag:
+            return
+
+        self.clip_dict[self.clip_name] = self.clip_dict[self.clip_name].to(self.device)
+        self._move_flag = True
+
+    def encode(self, image):
+        self.move()
+        return self(image, zero_embedding_radio=self.zero_embedding_radio)
+
+
+class FrozenCLIPImageGridEmbedder(AbstractEncoder):
+
+    def __init__(
+            self,
+            version="openai/clip-vit-large-patch14",
+            device="cuda",
+            zero_embedding_radio=0.1,
+    ):
+        super().__init__()
+
+        self.device = device
+
+        self.clip_dict = OrderedDict()
+        self.clip_name = os.path.split(version)[-1]
+
+        clip_model: CLIPModel = CLIPModel.from_pretrained(version)
+        clip_model.text_model = None
+        clip_model.text_projection = None
+        clip_model = clip_model.eval()
+        for param in self.parameters():
+            param.requires_grad = False
+        self.clip_dict[self.clip_name] = clip_model
+
+        self.transform = transforms.Compose(
+            [
+                transforms.Resize(224, transforms.InterpolationMode.BILINEAR, antialias=True),
+                transforms.CenterCrop(224),  # crop a (224, 224) square
+                transforms.Normalize(
+                    mean=[0.48145466, 0.4578275, 0.40821073],
+                    std=[0.26862954, 0.26130258, 0.27577711],
+                ),
+            ]
+        )
+        self.zero_embedding_radio = zero_embedding_radio
+        self.embedding_dim = clip_model.vision_embed_dim
+
+        self._move_flag = False
+
+    @property
+    def clip(self):
+        return self.clip_dict[self.clip_name]
+
+    def move(self):
+        if self._move_flag:
+            return
+
+        self.clip_dict[self.clip_name] = self.clip_dict[self.clip_name].to(self.device)
+        self._move_flag = True
+
+    def unconditional_embedding(self, batch_size):
+        zero = torch.zeros(
+            batch_size,
+            self.clip.vision_model.embeddings.num_positions,
+            self.embedding_dim,
+            device=self.device,
+            dtype=self.clip.visual_projection.weight.dtype,
+        )
+        return zero
+
+    def forward(self, image, value_range=(-1, 1), zero_embedding_radio=0):
+        self.move()
+
+        if value_range is not None:
+            low, high = value_range
+            image = (image - low) / (high - low)
+
+        image = image.to(self.device, dtype=self.clip.visual_projection.weight.dtype)
+
+        z = self.clip.vision_model(self.transform(image)).last_hidden_state
+
+        if zero_embedding_radio > 0:
+            mask = torch.rand((len(image), 1, 1), device=z.device, dtype=z.dtype) >= zero_embedding_radio
+            z = z * mask.to(z)
+
+        return z
+
+    def encode(self, image):
+        return self(image, zero_embedding_radio=self.zero_embedding_radio)
+
+
+class MoECLIPImageEncoder(nn.Module):
+    def __init__(
+            self,
+            versions,
+            hidden_state_dim,
+            num_projection_vector=8,
+            zero_embedding_radio=0.1,
+            device="cuda",
+            precision="fp16",
+            normalize=False,
+            clip_max=0,
+            transform_type="base",
+            argument_p=0.2,
+    ):
+        super().__init__()
+
+        self.device = torch.device(device)
+        self.hidden_state_dim = hidden_state_dim
+        self.zero_embedding_radio = zero_embedding_radio
+        self.num_projection_vector = num_projection_vector
+        self.dtype = dict(fp16=torch.float16, fp32=torch.float32, bf16=torch.bfloat16)[precision]
+        self.normalize = normalize
+        self.clip_max = clip_max
+
+        if transform_type == "base":
+            self.transform = transforms.Compose(
+                [
+                    transforms.Resize(224, transforms.InterpolationMode.BICUBIC, antialias=True),
+                    transforms.CenterCrop(224),  # crop a (224, 224) square
+                    transforms.Normalize(
+                        mean=[0.48145466, 0.4578275, 0.40821073],
+                        std=[0.26862954, 0.26130258, 0.27577711],
+                    ),
+                ]
+            )
+        elif transform_type == "crop_blur_resize":
+            self.transform = transforms.Compose(
+                [
+                    transforms.Resize(224, transforms.InterpolationMode.BICUBIC, antialias=True),
+                    transforms.CenterCrop(224),  # crop a (224, 224) square
+                    transforms.RandomApply(
+                        transforms=[
+                            transforms.RandomResizedCrop(
+                                size=224,
+                                scale=(0.8, 1.0),
+                                ratio=(0.99, 1.01),
+                                interpolation=transforms.InterpolationMode.BICUBIC,
+                            ),
+                        ],
+                        p=argument_p,
+                    ),
+                    transforms.RandomApply(
+                        transforms=[
+                            transforms.GaussianBlur(kernel_size=9, sigma=(0.1, 5)),
+                        ],
+                        p=argument_p,
+                    ),
+                    transforms.RandomApply(
+                        transforms=[
+                            RandomResize(size=224, resize_radio=(0.2, 1)),
+                        ],
+                        p=argument_p,
+                    ),
+                    transforms.Normalize(
+                        mean=[0.48145466, 0.4578275, 0.40821073],
+                        std=[0.26862954, 0.26130258, 0.27577711],
+                    ),
+                ]
+            )
+        else:
+            raise ValueError(f"invalid {transform_type=}")
+
+        if isinstance(versions, str):
+            versions = (versions,)
+
+        # 如果直接把clips定位为当前类的子module，1. 会在保存ckp时存无用的多个权重。 2. pl会调用to，导致layer_norm的权重也被转换成fp16
+        clips = OrderedDict()
+
+        for v in versions:
+            # 因为clips不是子module，直接指定device="cuda"会错误地导致clip模型权重都被放到cuda:0上。
+            clips[v], _ = clip.load(name=v, device="cpu", jit=False, download_root=None)
+            delattr(clips[v], "transformer")
+            clips[v].eval()
+            clips[v].requires_grad_(False)
+
+        self.clips_hidden_dim = sum(clips[v].ln_final.weight.size(0) for v in clips)
+
+        if self.num_projection_vector == 0:
+            self.projection = nn.Identity()
+        else:
+            self.projection = nn.Linear(self.clips_hidden_dim, hidden_state_dim * self.num_projection_vector, bias=True)
+            self.projection.to(dtype=self.dtype)
+            nn.init.normal_(self.projection.weight, std=self.clips_hidden_dim ** -0.5)
+
+        self.clips = clips
+
+        self._move_flag = False
+
+    def move(self):
+        if self._move_flag:
+            return
+
+        def convert_weights(model: nn.Module):
+            """Convert applicable model parameters to fp16"""
+
+            def _convert_weights_to_fp16(l):
+                if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):
+                    l.weight.data = l.weight.data.type(self.dtype)
+                    if l.bias is not None:
+                        l.bias.data = l.bias.data.type(self.dtype)
+
+                if isinstance(l, nn.MultiheadAttention):
+                    for attr in [
+                        *[f"{s}_proj_weight" for s in ["in", "q", "k", "v"]],
+                        "in_proj_bias",
+                        "bias_k",
+                        "bias_v",
+                    ]:
+                        tensor = getattr(l, attr)
+                        if tensor is not None:
+                            tensor.data = tensor.data.type(self.dtype)
+
+                for name in ["text_projection", "proj"]:
+                    if hasattr(l, name):
+                        attr = getattr(l, name)
+                        if attr is not None:
+                            attr.data = attr.data.type(self.dtype)
+
+            model.apply(_convert_weights_to_fp16)
+
+        for k in self.clips:
+            self.clips[k].to(self.device)
+            convert_weights(self.clips[k])  # fp32 -> self.dtype
+        self._move_flag = True
+
+    def unconditional_embedding(self, batch_size=None):
+        zero = torch.zeros(
+            batch_size,
+            self.clips_hidden_dim,
+            device=self.device,
+            dtype=self.dtype,
+        )
+        if self.num_projection_vector > 0:
+            zero = self.projection(zero).view(batch_size, self.num_projection_vector, -1)
+        return zero
+
+    def convert_embedding(self, z):
+        if self.num_projection_vector > 0:
+            z = self.projection(z.type(self.projection.weight.dtype)).view(len(z), self.num_projection_vector, -1)
+        return z
+
+    def forward(self, image, value_range=(-1, 1), zero_embedding_radio=0):
+        if value_range is not None:
+            low, high = value_range
+            image = (image - low) / (high - low)
+
+        image = self.transform(image)
+
+        with torch.no_grad():
+            embs = []
+            for v in self.clips:
+                x = self.clips[v].encode_image(image)
+                if self.normalize:
+                    x = x / x.norm(p=2, dim=-1, keepdim=True) * (x.size(-1) ** 0.5)
+                    # clip_max only works with normalization
+                    if self.clip_max > 0:
+                        x = x.clamp(-self.clip_max, self.clip_max)
+                embs.append(x)
+
+            z = torch.cat(embs, dim=-1)
+            if self.normalize:
+                z /= z.size(-1) ** 0.5
+
+        if zero_embedding_radio > 0:
+            mask = torch.rand((len(image), 1, 1), device=z.device, dtype=z.dtype) >= zero_embedding_radio
+            z = z + mask.to(z)
+
+        if self.num_projection_vector > 0:
+            z = self.projection(z).view(len(image), self.num_projection_vector, -1)
+        return z
+
+    def encode(self, image):
+        self.move()
+        return self(image, zero_embedding_radio=self.zero_embedding_radio)

michelangelo/models/modules/__init__.py

@@ -0,0 +1,3 @@
+# -*- coding: utf-8 -*-
+
+from .checkpoint import checkpoint

michelangelo/models/modules/checkpoint.py

@@ -0,0 +1,69 @@
+# -*- coding: utf-8 -*-
+"""
+Adapted from: https://github.com/openai/guided-diffusion/blob/22e0df8183507e13a7813f8d38d51b072ca1e67c/guided_diffusion/nn.py#L124
+"""
+
+import torch
+from typing import Callable, Iterable, Sequence, Union
+
+
+def checkpoint(
+    func: Callable[..., Union[torch.Tensor, Sequence[torch.Tensor]]],
+    inputs: Sequence[torch.Tensor],
+    params: Iterable[torch.Tensor],
+    flag: bool,
+    use_deepspeed: bool = False
+):
+    """
+    Evaluate a function without caching intermediate activations, allowing for
+    reduced memory at the expense of extra compute in the backward pass.
+    :param func: the function to evaluate.
+    :param inputs: the argument sequence to pass to `func`.
+    :param params: a sequence of parameters `func` depends on but does not
+                   explicitly take as arguments.
+    :param flag: if False, disable gradient checkpointing.
+    :param use_deepspeed: if True, use deepspeed
+    """
+    if flag:
+        if use_deepspeed:
+            import deepspeed
+            return deepspeed.checkpointing.checkpoint(func, *inputs)
+
+        args = tuple(inputs) + tuple(params)
+        return CheckpointFunction.apply(func, len(inputs), *args)
+    else:
+        return func(*inputs)
+
+
+class CheckpointFunction(torch.autograd.Function):
+    @staticmethod
+    @torch.cuda.amp.custom_fwd
+    def forward(ctx, run_function, length, *args):
+        ctx.run_function = run_function
+        ctx.input_tensors = list(args[:length])
+        ctx.input_params = list(args[length:])
+
+        with torch.no_grad():
+            output_tensors = ctx.run_function(*ctx.input_tensors)
+        return output_tensors
+
+    @staticmethod
+    @torch.cuda.amp.custom_bwd
+    def backward(ctx, *output_grads):
+        ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors]
+        with torch.enable_grad():
+            # Fixes a bug where the first op in run_function modifies the
+            # Tensor storage in place, which is not allowed for detach()'d
+            # Tensors.
+            shallow_copies = [x.view_as(x) for x in ctx.input_tensors]
+            output_tensors = ctx.run_function(*shallow_copies)
+        input_grads = torch.autograd.grad(
+            output_tensors,
+            ctx.input_tensors + ctx.input_params,
+            output_grads,
+            allow_unused=True,
+        )
+        del ctx.input_tensors
+        del ctx.input_params
+        del output_tensors
+        return (None, None) + input_grads

michelangelo/models/modules/diffusion_transformer.py

@@ -0,0 +1,218 @@
+# -*- coding: utf-8 -*-
+
+import math
+import torch
+import torch.nn as nn
+from typing import Optional
+
+from michelangelo.models.modules.checkpoint import checkpoint
+from michelangelo.models.modules.transformer_blocks import (
+    init_linear,
+    MLP,
+    MultiheadCrossAttention,
+    MultiheadAttention,
+    ResidualAttentionBlock
+)
+
+
+class AdaLayerNorm(nn.Module):
+    def __init__(self,
+                 device: torch.device,
+                 dtype: torch.dtype,
+                 width: int):
+
+        super().__init__()
+
+        self.silu = nn.SiLU(inplace=True)
+        self.linear = nn.Linear(width, width * 2, device=device, dtype=dtype)
+        self.layernorm = nn.LayerNorm(width, elementwise_affine=False, device=device, dtype=dtype)
+
+    def forward(self, x, timestep):
+        emb = self.linear(timestep)
+        scale, shift = torch.chunk(emb, 2, dim=2)
+        x = self.layernorm(x) * (1 + scale) + shift
+        return x
+
+
+class DitBlock(nn.Module):
+    def __init__(
+            self,
+            *,
+            device: torch.device,
+            dtype: torch.dtype,
+            n_ctx: int,
+            width: int,
+            heads: int,
+            context_dim: int,
+            qkv_bias: bool = False,
+            init_scale: float = 1.0,
+            use_checkpoint: bool = False
+    ):
+        super().__init__()
+
+        self.use_checkpoint = use_checkpoint
+
+        self.attn = MultiheadAttention(
+            device=device,
+            dtype=dtype,
+            n_ctx=n_ctx,
+            width=width,
+            heads=heads,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias
+        )
+        self.ln_1 = AdaLayerNorm(device, dtype, width)
+
+        if context_dim is not None:
+            self.ln_2 = AdaLayerNorm(device, dtype, width)
+            self.cross_attn = MultiheadCrossAttention(
+                device=device,
+                dtype=dtype,
+                width=width,
+                heads=heads,
+                data_width=context_dim,
+                init_scale=init_scale,
+                qkv_bias=qkv_bias
+            )
+
+        self.mlp = MLP(device=device, dtype=dtype, width=width, init_scale=init_scale)
+        self.ln_3 = AdaLayerNorm(device, dtype, width)
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor, context: Optional[torch.Tensor] = None):
+        return checkpoint(self._forward, (x, t, context), self.parameters(), self.use_checkpoint)
+
+    def _forward(self, x: torch.Tensor, t: torch.Tensor, context: Optional[torch.Tensor] = None):
+        x = x + self.attn(self.ln_1(x, t))
+        if context is not None:
+            x = x + self.cross_attn(self.ln_2(x, t), context)
+        x = x + self.mlp(self.ln_3(x, t))
+        return x
+
+
+class DiT(nn.Module):
+    def __init__(
+            self,
+            *,
+            device: Optional[torch.device],
+            dtype: Optional[torch.dtype],
+            n_ctx: int,
+            width: int,
+            layers: int,
+            heads: int,
+            context_dim: int,
+            init_scale: float = 0.25,
+            qkv_bias: bool = False,
+            use_checkpoint: bool = False
+    ):
+        super().__init__()
+        self.n_ctx = n_ctx
+        self.width = width
+        self.layers = layers
+
+        self.resblocks = nn.ModuleList(
+            [
+                DitBlock(
+                    device=device,
+                    dtype=dtype,
+                    n_ctx=n_ctx,
+                    width=width,
+                    heads=heads,
+                    context_dim=context_dim,
+                    qkv_bias=qkv_bias,
+                    init_scale=init_scale,
+                    use_checkpoint=use_checkpoint
+                )
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor, context: Optional[torch.Tensor] = None):
+        for block in self.resblocks:
+            x = block(x, t, context)
+        return x
+
+
+class UNetDiffusionTransformer(nn.Module):
+    def __init__(
+            self,
+            *,
+            device: Optional[torch.device],
+            dtype: Optional[torch.dtype],
+            n_ctx: int,
+            width: int,
+            layers: int,
+            heads: int,
+            init_scale: float = 0.25,
+            qkv_bias: bool = False,
+            skip_ln: bool = False,
+            use_checkpoint: bool = False
+    ):
+        super().__init__()
+
+        self.n_ctx = n_ctx
+        self.width = width
+        self.layers = layers
+
+        self.encoder = nn.ModuleList()
+        for _ in range(layers):
+            resblock = ResidualAttentionBlock(
+                device=device,
+                dtype=dtype,
+                n_ctx=n_ctx,
+                width=width,
+                heads=heads,
+                init_scale=init_scale,
+                qkv_bias=qkv_bias,
+                use_checkpoint=use_checkpoint
+            )
+            self.encoder.append(resblock)
+
+        self.middle_block = ResidualAttentionBlock(
+            device=device,
+            dtype=dtype,
+            n_ctx=n_ctx,
+            width=width,
+            heads=heads,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            use_checkpoint=use_checkpoint
+        )
+
+        self.decoder = nn.ModuleList()
+        for _ in range(layers):
+            resblock = ResidualAttentionBlock(
+                device=device,
+                dtype=dtype,
+                n_ctx=n_ctx,
+                width=width,
+                heads=heads,
+                init_scale=init_scale,
+                qkv_bias=qkv_bias,
+                use_checkpoint=use_checkpoint
+            )
+            linear = nn.Linear(width * 2, width, device=device, dtype=dtype)
+            init_linear(linear, init_scale)
+
+            layer_norm = nn.LayerNorm(width, device=device, dtype=dtype) if skip_ln else None
+
+            self.decoder.append(nn.ModuleList([resblock, linear, layer_norm]))
+
+    def forward(self, x: torch.Tensor):
+
+        enc_outputs = []
+        for block in self.encoder:
+            x = block(x)
+            enc_outputs.append(x)
+
+        x = self.middle_block(x)
+
+        for i, (resblock, linear, layer_norm) in enumerate(self.decoder):
+            x = torch.cat([enc_outputs.pop(), x], dim=-1)
+            x = linear(x)
+
+            if layer_norm is not None:
+                x = layer_norm(x)
+
+            x = resblock(x)
+
+        return x

michelangelo/models/modules/distributions.py

@@ -0,0 +1,100 @@
+import torch
+import numpy as np
+from typing import Union, List
+
+
+class AbstractDistribution(object):
+    def sample(self):
+        raise NotImplementedError()
+
+    def mode(self):
+        raise NotImplementedError()
+
+
+class DiracDistribution(AbstractDistribution):
+    def __init__(self, value):
+        self.value = value
+
+    def sample(self):
+        return self.value
+
+    def mode(self):
+        return self.value
+
+
+class DiagonalGaussianDistribution(object):
+    def __init__(self, parameters: Union[torch.Tensor, List[torch.Tensor]], deterministic=False, feat_dim=1):
+        self.feat_dim = feat_dim
+        self.parameters = parameters
+
+        if isinstance(parameters, list):
+            self.mean = parameters[0]
+            self.logvar = parameters[1]
+        else:
+            self.mean, self.logvar = torch.chunk(parameters, 2, dim=feat_dim)
+
+        self.logvar = torch.clamp(self.logvar, -30.0, 20.0)
+        self.deterministic = deterministic
+        self.std = torch.exp(0.5 * self.logvar)
+        self.var = torch.exp(self.logvar)
+        if self.deterministic:
+            self.var = self.std = torch.zeros_like(self.mean)
+
+    def sample(self):
+        x = self.mean + self.std * torch.randn_like(self.mean)
+        return x
+
+    def kl(self, other=None, dims=(1, 2, 3)):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        else:
+            if other is None:
+                return 0.5 * torch.mean(torch.pow(self.mean, 2)
+                                        + self.var - 1.0 - self.logvar,
+                                        dim=dims)
+            else:
+                return 0.5 * torch.mean(
+                    torch.pow(self.mean - other.mean, 2) / other.var
+                    + self.var / other.var - 1.0 - self.logvar + other.logvar,
+                    dim=dims)
+
+    def nll(self, sample, dims=(1, 2, 3)):
+        if self.deterministic:
+            return torch.Tensor([0.])
+        logtwopi = np.log(2.0 * np.pi)
+        return 0.5 * torch.sum(
+            logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var,
+            dim=dims)
+
+    def mode(self):
+        return self.mean
+
+
+def normal_kl(mean1, logvar1, mean2, logvar2):
+    """
+    source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12
+    Compute the KL divergence between two gaussians.
+    Shapes are automatically broadcasted, so batches can be compared to
+    scalars, among other use cases.
+    """
+    tensor = None
+    for obj in (mean1, logvar1, mean2, logvar2):
+        if isinstance(obj, torch.Tensor):
+            tensor = obj
+            break
+    assert tensor is not None, "at least one argument must be a Tensor"
+
+    # Force variances to be Tensors. Broadcasting helps convert scalars to
+    # Tensors, but it does not work for torch.exp().
+    logvar1, logvar2 = [
+        x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor)
+        for x in (logvar1, logvar2)
+    ]
+
+    return 0.5 * (
+            -1.0
+            + logvar2
+            - logvar1
+            + torch.exp(logvar1 - logvar2)
+            + ((mean1 - mean2) ** 2) * torch.exp(-logvar2)
+    )

michelangelo/models/modules/embedder.py

@@ -0,0 +1,213 @@
+# -*- coding: utf-8 -*-
+
+import numpy as np
+import torch
+import torch.nn as nn
+import math
+
+VALID_EMBED_TYPES = ["identity", "fourier", "hashgrid", "sphere_harmonic", "triplane_fourier"]
+
+
+class FourierEmbedder(nn.Module):
+    """The sin/cosine positional embedding. Given an input tensor `x` of shape [n_batch, ..., c_dim], it converts
+    each feature dimension of `x[..., i]` into:
+        [
+            sin(x[..., i]),
+            sin(f_1*x[..., i]),
+            sin(f_2*x[..., i]),
+            ...
+            sin(f_N * x[..., i]),
+            cos(x[..., i]),
+            cos(f_1*x[..., i]),
+            cos(f_2*x[..., i]),
+            ...
+            cos(f_N * x[..., i]),
+            x[..., i]     # only present if include_input is True.
+        ], here f_i is the frequency.
+
+    Denote the space is [0 / num_freqs, 1 / num_freqs, 2 / num_freqs, 3 / num_freqs, ..., (num_freqs - 1) / num_freqs].
+    If logspace is True, then the frequency f_i is [2^(0 / num_freqs), ..., 2^(i / num_freqs), ...];
+    Otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)].
+
+    Args:
+        num_freqs (int): the number of frequencies, default is 6;
+        logspace (bool): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+            otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1)];
+        input_dim (int): the input dimension, default is 3;
+        include_input (bool): include the input tensor or not, default is True.
+
+    Attributes:
+        frequencies (torch.Tensor): If logspace is True, then the frequency f_i is [..., 2^(i / num_freqs), ...],
+                otherwise, the frequencies are linearly spaced between [1.0, 2^(num_freqs - 1);
+
+        out_dim (int): the embedding size, if include_input is True, it is input_dim * (num_freqs * 2 + 1),
+            otherwise, it is input_dim * num_freqs * 2.
+
+    """
+
+    def __init__(self,
+                 num_freqs: int = 6,
+                 logspace: bool = True,
+                 input_dim: int = 3,
+                 include_input: bool = True,
+                 include_pi: bool = True) -> None:
+
+        """The initialization"""
+
+        super().__init__()
+
+        if logspace:
+            frequencies = 2.0 ** torch.arange(
+                num_freqs,
+                dtype=torch.float32
+            )
+        else:
+            frequencies = torch.linspace(
+                1.0,
+                2.0 ** (num_freqs - 1),
+                num_freqs,
+                dtype=torch.float32
+            )
+
+        if include_pi:
+            frequencies *= torch.pi
+
+        self.register_buffer("frequencies", frequencies, persistent=False)
+        self.include_input = include_input
+        self.num_freqs = num_freqs
+
+        self.out_dim = self.get_dims(input_dim)
+
+    def get_dims(self, input_dim):
+        temp = 1 if self.include_input or self.num_freqs == 0 else 0
+        out_dim = input_dim * (self.num_freqs * 2 + temp)
+
+        return out_dim
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """ Forward process.
+
+        Args:
+            x: tensor of shape [..., dim]
+
+        Returns:
+            embedding: an embedding of `x` of shape [..., dim * (num_freqs * 2 + temp)]
+                where temp is 1 if include_input is True and 0 otherwise.
+        """
+
+        if self.num_freqs > 0:
+            embed = (x[..., None].contiguous() * self.frequencies).view(*x.shape[:-1], -1)
+            if self.include_input:
+                return torch.cat((x, embed.sin(), embed.cos()), dim=-1)
+            else:
+                return torch.cat((embed.sin(), embed.cos()), dim=-1)
+        else:
+            return x
+
+
+class LearnedFourierEmbedder(nn.Module):
+    """ following @crowsonkb "s lead with learned sinusoidal pos emb """
+    """ https://github.com/crowsonkb/v-diffusion-jax/blob/master/diffusion/models/danbooru_128.py#L8 """
+
+    def __init__(self, in_channels, dim):
+        super().__init__()
+        assert (dim % 2) == 0
+        half_dim = dim // 2
+        per_channel_dim = half_dim // in_channels
+        self.weights = nn.Parameter(torch.randn(per_channel_dim))
+
+    def forward(self, x):
+        """
+
+        Args:
+            x (torch.FloatTensor): [..., c]
+
+        Returns:
+            x (torch.FloatTensor): [..., d]
+        """
+
+        # [b, t, c, 1] * [1, d] = [b, t, c, d] -> [b, t, c * d]
+        freqs = (x[..., None] * self.weights[None] * 2 * np.pi).view(*x.shape[:-1], -1)
+        fouriered = torch.cat((x, freqs.sin(), freqs.cos()), dim=-1)
+        return fouriered
+
+
+class TriplaneLearnedFourierEmbedder(nn.Module):
+    def __init__(self, in_channels, dim):
+        super().__init__()
+
+        self.yz_plane_embedder = LearnedFourierEmbedder(in_channels, dim)
+        self.xz_plane_embedder = LearnedFourierEmbedder(in_channels, dim)
+        self.xy_plane_embedder = LearnedFourierEmbedder(in_channels, dim)
+
+        self.out_dim = in_channels + dim
+
+    def forward(self, x):
+
+        yz_embed = self.yz_plane_embedder(x)
+        xz_embed = self.xz_plane_embedder(x)
+        xy_embed = self.xy_plane_embedder(x)
+
+        embed = yz_embed + xz_embed + xy_embed
+
+        return embed
+
+
+def sequential_pos_embed(num_len, embed_dim):
+    assert embed_dim % 2 == 0
+
+    pos = torch.arange(num_len, dtype=torch.float32)
+    omega = torch.arange(embed_dim // 2, dtype=torch.float32)
+    omega /= embed_dim / 2.
+    omega = 1. / 10000 ** omega  # (D/2,)
+
+    pos = pos.reshape(-1)  # (M,)
+    out = torch.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+
+    emb_sin = torch.sin(out)  # (M, D/2)
+    emb_cos = torch.cos(out)  # (M, D/2)
+
+    embeddings = torch.cat([emb_sin, emb_cos], dim=1)  # (M, D)
+
+    return embeddings
+
+
+def timestep_embedding(timesteps, dim, max_period=10000):
+    """
+    Create sinusoidal timestep embeddings.
+    :param timesteps: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an [N x dim] Tensor of positional embeddings.
+    """
+    half = dim // 2
+    freqs = torch.exp(
+        -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+    ).to(device=timesteps.device)
+    args = timesteps[:, None].to(timesteps.dtype) * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+def get_embedder(embed_type="fourier", num_freqs=-1, input_dim=3, degree=4,
+                 num_levels=16, level_dim=2, per_level_scale=2, base_resolution=16,
+                 log2_hashmap_size=19, desired_resolution=None):
+    if embed_type == "identity" or (embed_type == "fourier" and num_freqs == -1):
+        return nn.Identity(), input_dim
+
+    elif embed_type == "fourier":
+        embedder_obj = FourierEmbedder(num_freqs=num_freqs, input_dim=input_dim,
+                                       logspace=True, include_input=True)
+        return embedder_obj, embedder_obj.out_dim
+
+    elif embed_type == "hashgrid":
+        raise NotImplementedError
+
+    elif embed_type == "sphere_harmonic":
+        raise NotImplementedError
+
+    else:
+        raise ValueError(f"{embed_type} is not valid. Currently only supprts {VALID_EMBED_TYPES}")

michelangelo/models/modules/transformer_blocks.py

@@ -0,0 +1,286 @@
+# -*- coding: utf-8 -*-
+
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from typing import Optional
+
+from michelangelo.models.modules.checkpoint import checkpoint
+
+
+def init_linear(l, stddev):
+    nn.init.normal_(l.weight, std=stddev)
+    if l.bias is not None:
+        nn.init.constant_(l.bias, 0.0)
+
+
+class MultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        n_ctx: int,
+        width: int,
+        heads: int,
+        init_scale: float,
+        qkv_bias: bool,
+        flash: bool = False
+    ):
+        super().__init__()
+        self.n_ctx = n_ctx
+        self.width = width
+        self.heads = heads
+        self.c_qkv = nn.Linear(width, width * 3, bias=qkv_bias, device=device, dtype=dtype)
+        self.c_proj = nn.Linear(width, width, device=device, dtype=dtype)
+        self.attention = QKVMultiheadAttention(device=device, dtype=dtype, heads=heads, n_ctx=n_ctx, flash=flash)
+        init_linear(self.c_qkv, init_scale)
+        init_linear(self.c_proj, init_scale)
+
+    def forward(self, x):
+        x = self.c_qkv(x)
+        x = checkpoint(self.attention, (x,), (), True)
+        x = self.c_proj(x)
+        return x
+
+
+class QKVMultiheadAttention(nn.Module):
+    def __init__(self, *, device: torch.device, dtype: torch.dtype, heads: int, n_ctx: int, flash: bool = False):
+        super().__init__()
+        self.device = device
+        self.dtype = dtype
+        self.heads = heads
+        self.n_ctx = n_ctx
+        self.flash = flash
+
+    def forward(self, qkv):
+        bs, n_ctx, width = qkv.shape
+        attn_ch = width // self.heads // 3
+        scale = 1 / math.sqrt(math.sqrt(attn_ch))
+        qkv = qkv.view(bs, n_ctx, self.heads, -1)
+        q, k, v = torch.split(qkv, attn_ch, dim=-1)
+
+        if self.flash:
+            out = F.scaled_dot_product_attention(q, k, v)
+        else:
+            weight = torch.einsum(
+                "bthc,bshc->bhts", q * scale, k * scale
+            )  # More stable with f16 than dividing afterwards
+            wdtype = weight.dtype
+            weight = torch.softmax(weight.float(), dim=-1).type(wdtype)
+            out = torch.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1)
+
+        return out
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        n_ctx: int,
+        width: int,
+        heads: int,
+        init_scale: float = 1.0,
+        qkv_bias: bool = True,
+        flash: bool = False,
+        use_checkpoint: bool = False
+    ):
+        super().__init__()
+
+        self.use_checkpoint = use_checkpoint
+
+        self.attn = MultiheadAttention(
+            device=device,
+            dtype=dtype,
+            n_ctx=n_ctx,
+            width=width,
+            heads=heads,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            flash=flash
+        )
+        self.ln_1 = nn.LayerNorm(width, device=device, dtype=dtype)
+        self.mlp = MLP(device=device, dtype=dtype, width=width, init_scale=init_scale)
+        self.ln_2 = nn.LayerNorm(width, device=device, dtype=dtype)
+
+    def _forward(self, x: torch.Tensor):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+    def forward(self, x: torch.Tensor):
+        return checkpoint(self._forward, (x,), self.parameters(), self.use_checkpoint)
+
+
+class MultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        width: int,
+        heads: int,
+        init_scale: float,
+        qkv_bias: bool = True,
+        flash: bool = False,
+        n_data: Optional[int] = None,
+        data_width: Optional[int] = None,
+    ):
+        super().__init__()
+        self.n_data = n_data
+        self.width = width
+        self.heads = heads
+        self.data_width = width if data_width is None else data_width
+        self.c_q = nn.Linear(width, width, bias=qkv_bias, device=device, dtype=dtype)
+        self.c_kv = nn.Linear(self.data_width, width * 2, bias=qkv_bias, device=device, dtype=dtype)
+        self.c_proj = nn.Linear(width, width, device=device, dtype=dtype)
+        self.attention = QKVMultiheadCrossAttention(
+            device=device, dtype=dtype, heads=heads, n_data=n_data, flash=flash
+        )
+        init_linear(self.c_q, init_scale)
+        init_linear(self.c_kv, init_scale)
+        init_linear(self.c_proj, init_scale)
+
+    def forward(self, x, data):
+        x = self.c_q(x)
+        data = self.c_kv(data)
+        x = checkpoint(self.attention, (x, data), (), True)
+        x = self.c_proj(x)
+        return x
+
+
+class QKVMultiheadCrossAttention(nn.Module):
+    def __init__(self, *, device: torch.device, dtype: torch.dtype, heads: int,
+                 flash: bool = False, n_data: Optional[int] = None):
+
+        super().__init__()
+        self.device = device
+        self.dtype = dtype
+        self.heads = heads
+        self.n_data = n_data
+        self.flash = flash
+
+    def forward(self, q, kv):
+        _, n_ctx, _ = q.shape
+        bs, n_data, width = kv.shape
+        attn_ch = width // self.heads // 2
+        scale = 1 / math.sqrt(math.sqrt(attn_ch))
+        q = q.view(bs, n_ctx, self.heads, -1)
+        kv = kv.view(bs, n_data, self.heads, -1)
+        k, v = torch.split(kv, attn_ch, dim=-1)
+
+        if self.flash:
+            out = F.scaled_dot_product_attention(q, k, v)
+        else:
+            weight = torch.einsum(
+                "bthc,bshc->bhts", q * scale, k * scale
+            )  # More stable with f16 than dividing afterwards
+            wdtype = weight.dtype
+            weight = torch.softmax(weight.float(), dim=-1).type(wdtype)
+            out = torch.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1)
+
+        return out
+
+
+class ResidualCrossAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: Optional[torch.device],
+        dtype: Optional[torch.dtype],
+        n_data: Optional[int] = None,
+        width: int,
+        heads: int,
+        data_width: Optional[int] = None,
+        init_scale: float = 0.25,
+        qkv_bias: bool = True,
+        flash: bool = False
+    ):
+        super().__init__()
+
+        if data_width is None:
+            data_width = width
+
+        self.attn = MultiheadCrossAttention(
+            device=device,
+            dtype=dtype,
+            n_data=n_data,
+            width=width,
+            heads=heads,
+            data_width=data_width,
+            init_scale=init_scale,
+            qkv_bias=qkv_bias,
+            flash=flash,
+        )
+        self.ln_1 = nn.LayerNorm(width, device=device, dtype=dtype)
+        self.ln_2 = nn.LayerNorm(data_width, device=device, dtype=dtype)
+        self.mlp = MLP(device=device, dtype=dtype, width=width, init_scale=init_scale)
+        self.ln_3 = nn.LayerNorm(width, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, data: torch.Tensor):
+        x = x + self.attn(self.ln_1(x), self.ln_2(data))
+        x = x + self.mlp(self.ln_3(x))
+        return x
+
+
+class MLP(nn.Module):
+    def __init__(self, *,
+                 device: Optional[torch.device],
+                 dtype: Optional[torch.dtype],
+                 width: int,
+                 init_scale: float):
+        super().__init__()
+        self.width = width
+        self.c_fc = nn.Linear(width, width * 4, device=device, dtype=dtype)
+        self.c_proj = nn.Linear(width * 4, width, device=device, dtype=dtype)
+        self.gelu = nn.GELU()
+        init_linear(self.c_fc, init_scale)
+        init_linear(self.c_proj, init_scale)
+
+    def forward(self, x):
+        return self.c_proj(self.gelu(self.c_fc(x)))
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: Optional[torch.device],
+        dtype: Optional[torch.dtype],
+        n_ctx: int,
+        width: int,
+        layers: int,
+        heads: int,
+        init_scale: float = 0.25,
+        qkv_bias: bool = True,
+        flash: bool = False,
+        use_checkpoint: bool = False
+    ):
+        super().__init__()
+        self.n_ctx = n_ctx
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(
+                    device=device,
+                    dtype=dtype,
+                    n_ctx=n_ctx,
+                    width=width,
+                    heads=heads,
+                    init_scale=init_scale,
+                    qkv_bias=qkv_bias,
+                    flash=flash,
+                    use_checkpoint=use_checkpoint
+                )
+                for _ in range(layers)
+            ]
+        )
+
+    def forward(self, x: torch.Tensor):
+        for block in self.resblocks:
+            x = block(x)
+        return x

michelangelo/models/modules/transformer_vit.py

@@ -0,0 +1,308 @@
+# -*- coding: utf-8 -*-
+
+import math
+import torch
+import torch.nn as nn
+from typing import Optional
+import warnings
+
+from michelangelo.models.modules.checkpoint import checkpoint
+
+
+def _trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1. + math.erf(x / math.sqrt(2.))) / 2.
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+                      "The distribution of values may be incorrect.",
+                      stacklevel=2)
+
+    # Values are generated by using a truncated uniform distribution and
+    # then using the inverse CDF for the normal distribution.
+    # Get upper and lower cdf values
+    l = norm_cdf((a - mean) / std)
+    u = norm_cdf((b - mean) / std)
+
+    # Uniformly fill tensor with values from [l, u], then translate to
+    # [2l-1, 2u-1].
+    tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+    # Use inverse cdf transform for normal distribution to get truncated
+    # standard normal
+    tensor.erfinv_()
+
+    # Transform to proper mean, std
+    tensor.mul_(std * math.sqrt(2.))
+    tensor.add_(mean)
+
+    # Clamp to ensure it's in the proper range
+    tensor.clamp_(min=a, max=b)
+    return tensor
+
+
+def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
+    # type: (Tensor | nn.Parameter, float, float, float, float) -> Tensor
+    r"""Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \leq \text{mean} \leq b`.
+    NOTE: this impl is similar to the PyTorch trunc_normal_, the bounds [a, b] are
+    applied while sampling the normal with mean/std applied, therefore a, b args
+    should be adjusted to match the range of mean, std args.
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    Examples:
+        >>> w = torch.empty(3, 5)
+        >>> nn.init.trunc_normal_(w)
+    """
+    with torch.no_grad():
+        return _trunc_normal_(tensor, mean, std, a, b)
+
+
+def init_weights(m):
+    if isinstance(m, nn.Linear):
+        trunc_normal_(m.weight, std=.02)
+        if isinstance(m, nn.Linear) and m.bias is not None:
+            nn.init.constant_(m.bias, 0)
+    elif isinstance(m, nn.LayerNorm):
+        nn.init.constant_(m.bias, 0)
+        nn.init.constant_(m.weight, 1.0)
+
+
+class MultiheadAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        n_ctx: int,
+        width: int,
+        heads: int,
+        qkv_bias: bool
+    ):
+        super().__init__()
+        self.n_ctx = n_ctx
+        self.width = width
+        self.heads = heads
+        self.c_qkv = nn.Linear(width, width * 3, bias=qkv_bias, device=device, dtype=dtype)
+        self.c_proj = nn.Linear(width, width, device=device, dtype=dtype)
+        self.attention = QKVMultiheadAttention(device=device, dtype=dtype, heads=heads, n_ctx=n_ctx)
+
+    def forward(self, x):
+        x = self.c_qkv(x)
+        x = checkpoint(self.attention, (x,), (), True)
+        x = self.c_proj(x)
+        return x
+
+
+class QKVMultiheadAttention(nn.Module):
+    def __init__(self, *, device: torch.device, dtype: torch.dtype, heads: int, n_ctx: int):
+        super().__init__()
+        self.device = device
+        self.dtype = dtype
+        self.heads = heads
+        self.n_ctx = n_ctx
+
+    def forward(self, qkv):
+        bs, n_ctx, width = qkv.shape
+        attn_ch = width // self.heads // 3
+        scale = 1 / math.sqrt(attn_ch)
+        qkv = qkv.view(bs, n_ctx, self.heads, -1)
+        q, k, v = torch.split(qkv, attn_ch, dim=-1)
+        weight = torch.einsum("bthc,bshc->bhts", q, k) * scale
+        wdtype = weight.dtype
+        weight = torch.softmax(weight.float(), dim=-1).type(wdtype)
+        return torch.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1)
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        n_ctx: int,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        use_checkpoint: bool = False
+    ):
+        super().__init__()
+
+        self.use_checkpoint = use_checkpoint
+
+        self.attn = MultiheadAttention(
+            device=device,
+            dtype=dtype,
+            n_ctx=n_ctx,
+            width=width,
+            heads=heads,
+            qkv_bias=qkv_bias
+        )
+        self.ln_1 = nn.LayerNorm(width, device=device, dtype=dtype)
+        self.mlp = MLP(device=device, dtype=dtype, width=width)
+        self.ln_2 = nn.LayerNorm(width, device=device, dtype=dtype)
+
+    def _forward(self, x: torch.Tensor):
+        x = x + self.attn(self.ln_1(x))
+        x = x + self.mlp(self.ln_2(x))
+        return x
+
+    def forward(self, x: torch.Tensor):
+        return checkpoint(self._forward, (x,), self.parameters(), self.use_checkpoint)
+
+
+class MultiheadCrossAttention(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: torch.device,
+        dtype: torch.dtype,
+        width: int,
+        heads: int,
+        qkv_bias: bool = True,
+        n_data: Optional[int] = None,
+        data_width: Optional[int] = None,
+    ):
+        super().__init__()
+        self.n_data = n_data
+        self.width = width
+        self.heads = heads
+        self.data_width = width if data_width is None else data_width
+        self.c_q = nn.Linear(width, width, bias=qkv_bias, device=device, dtype=dtype)
+        self.c_kv = nn.Linear(self.data_width, width * 2, bias=qkv_bias, device=device, dtype=dtype)
+        self.c_proj = nn.Linear(width, width, device=device, dtype=dtype)
+        self.attention = QKVMultiheadCrossAttention(
+            device=device, dtype=dtype, heads=heads, n_data=n_data
+        )
+
+    def forward(self, x, data):
+        x = self.c_q(x)
+        data = self.c_kv(data)
+        x = checkpoint(self.attention, (x, data), (), True)
+        x = self.c_proj(x)
+        return x
+
+
+class QKVMultiheadCrossAttention(nn.Module):
+    def __init__(self, *, device: torch.device, dtype: torch.dtype, heads: int, n_data: Optional[int] = None):
+        super().__init__()
+        self.device = device
+        self.dtype = dtype
+        self.heads = heads
+        self.n_data = n_data
+
+    def forward(self, q, kv):
+        _, n_ctx, _ = q.shape
+        bs, n_data, width = kv.shape
+        attn_ch = width // self.heads // 2
+        scale = 1 / math.sqrt(attn_ch)
+        q = q.view(bs, n_ctx, self.heads, -1)
+        kv = kv.view(bs, n_data, self.heads, -1)
+        k, v = torch.split(kv, attn_ch, dim=-1)
+        weight = torch.einsum("bthc,bshc->bhts", q, k) * scale
+        wdtype = weight.dtype
+        weight = torch.softmax(weight.float(), dim=-1).type(wdtype)
+        return torch.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1)
+
+
+class ResidualCrossAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: Optional[torch.device],
+        dtype: Optional[torch.dtype],
+        n_data: Optional[int] = None,
+        width: int,
+        heads: int,
+        data_width: Optional[int] = None,
+        qkv_bias: bool = True
+    ):
+        super().__init__()
+
+        if data_width is None:
+            data_width = width
+
+        self.attn = MultiheadCrossAttention(
+            device=device,
+            dtype=dtype,
+            n_data=n_data,
+            width=width,
+            heads=heads,
+            data_width=data_width,
+            qkv_bias=qkv_bias
+        )
+        self.ln_1 = nn.LayerNorm(width, device=device, dtype=dtype)
+        self.ln_2 = nn.LayerNorm(data_width, device=device, dtype=dtype)
+        self.mlp = MLP(device=device, dtype=dtype, width=width)
+        self.ln_3 = nn.LayerNorm(width, device=device, dtype=dtype)
+
+    def forward(self, x: torch.Tensor, data: torch.Tensor):
+        x = x + self.attn(self.ln_1(x), self.ln_2(data))
+        x = x + self.mlp(self.ln_3(x))
+        return x
+
+
+class MLP(nn.Module):
+    def __init__(self, *,
+                 device: Optional[torch.device],
+                 dtype: Optional[torch.dtype],
+                 width: int):
+        super().__init__()
+        self.width = width
+        self.c_fc = nn.Linear(width, width * 4, device=device, dtype=dtype)
+        self.c_proj = nn.Linear(width * 4, width, device=device, dtype=dtype)
+        self.gelu = nn.GELU()
+
+    def forward(self, x):
+        return self.c_proj(self.gelu(self.c_fc(x)))
+
+
+class Transformer(nn.Module):
+    def __init__(
+        self,
+        *,
+        device: Optional[torch.device],
+        dtype: Optional[torch.dtype],
+        n_ctx: int,
+        width: int,
+        layers: int,
+        heads: int,
+        qkv_bias: bool = True,
+        use_checkpoint: bool = False
+    ):
+        super().__init__()
+        self.n_ctx = n_ctx
+        self.width = width
+        self.layers = layers
+        self.resblocks = nn.ModuleList(
+            [
+                ResidualAttentionBlock(
+                    device=device,
+                    dtype=dtype,
+                    n_ctx=n_ctx,
+                    width=width,
+                    heads=heads,
+                    qkv_bias=qkv_bias,
+                    use_checkpoint=use_checkpoint
+                )
+                for _ in range(layers)
+            ]
+        )
+
+        self.apply(init_weights)
+
+    def forward(self, x: torch.Tensor):
+        for block in self.resblocks:
+            x = block(x)
+        return x