Upload python file
Browse files- data_preprocess.py +543 -0
- modeling_tinyllava_phi.py +55 -0
data_preprocess.py
ADDED
|
@@ -0,0 +1,543 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
import requests
|
| 3 |
+
from PIL import Image
|
| 4 |
+
import torch
|
| 5 |
+
from io import BytesIO
|
| 6 |
+
import base64
|
| 7 |
+
import time
|
| 8 |
+
import torch
|
| 9 |
+
from transformers import StoppingCriteria
|
| 10 |
+
|
| 11 |
+
import math
|
| 12 |
+
import ast
|
| 13 |
+
|
| 14 |
+
# Model Constants
|
| 15 |
+
IGNORE_INDEX = -100
|
| 16 |
+
IMAGE_TOKEN_INDEX = -200
|
| 17 |
+
DEFAULT_IMAGE_TOKEN = "<image>"
|
| 18 |
+
DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
|
| 19 |
+
DEFAULT_IM_START_TOKEN = "<im_start>"
|
| 20 |
+
DEFAULT_IM_END_TOKEN = "<im_end>"
|
| 21 |
+
IMAGE_PLACEHOLDER = "<image-placeholder>"
|
| 22 |
+
import dataclasses
|
| 23 |
+
from enum import auto, Enum
|
| 24 |
+
from typing import List, Tuple
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
class SeparatorStyle(Enum):
|
| 28 |
+
"""Different separator style."""
|
| 29 |
+
SINGLE = auto()
|
| 30 |
+
TWO = auto()
|
| 31 |
+
MPT = auto()
|
| 32 |
+
PLAIN = auto()
|
| 33 |
+
LLAMA_2 = auto()
|
| 34 |
+
TINY_LLAMA = auto()
|
| 35 |
+
QWEN_2 = auto()
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
@dataclasses.dataclass
|
| 39 |
+
class Conversation:
|
| 40 |
+
"""A class that keeps all conversation history."""
|
| 41 |
+
system: str
|
| 42 |
+
roles: List[str]
|
| 43 |
+
messages: List[List[str]]
|
| 44 |
+
offset: int
|
| 45 |
+
sep_style: SeparatorStyle = SeparatorStyle.SINGLE
|
| 46 |
+
sep: str = "###"
|
| 47 |
+
sep2: str = None
|
| 48 |
+
version: str = "Unknown"
|
| 49 |
+
|
| 50 |
+
skip_next: bool = False
|
| 51 |
+
|
| 52 |
+
def get_prompt(self):
|
| 53 |
+
messages = self.messages
|
| 54 |
+
if len(messages) > 0 and type(messages[0][1]) is tuple:
|
| 55 |
+
messages = self.messages.copy()
|
| 56 |
+
init_role, init_msg = messages[0].copy()
|
| 57 |
+
init_msg = init_msg[0].replace("<image>", "").strip()
|
| 58 |
+
if 'mmtag' in self.version:
|
| 59 |
+
messages[0] = (init_role, init_msg)
|
| 60 |
+
messages.insert(0, (self.roles[0], "<Image><image></Image>"))
|
| 61 |
+
messages.insert(1, (self.roles[1], "Received."))
|
| 62 |
+
else:
|
| 63 |
+
messages[0] = (init_role, "<image>\n" + init_msg)
|
| 64 |
+
|
| 65 |
+
if self.sep_style == SeparatorStyle.SINGLE:
|
| 66 |
+
ret = self.system + self.sep
|
| 67 |
+
for role, message in messages:
|
| 68 |
+
if message:
|
| 69 |
+
if type(message) is tuple:
|
| 70 |
+
message, _, _ = message
|
| 71 |
+
ret += role + ": " + message + self.sep
|
| 72 |
+
else:
|
| 73 |
+
ret += role + ":"
|
| 74 |
+
elif self.sep_style == SeparatorStyle.TWO:
|
| 75 |
+
seps = [self.sep, self.sep2]
|
| 76 |
+
ret = self.system + seps[0]
|
| 77 |
+
for i, (role, message) in enumerate(messages):
|
| 78 |
+
if message:
|
| 79 |
+
if type(message) is tuple:
|
| 80 |
+
message, _, _ = message
|
| 81 |
+
ret += role + ": " + message + seps[i % 2]
|
| 82 |
+
else:
|
| 83 |
+
ret += role + ":"
|
| 84 |
+
elif self.sep_style == SeparatorStyle.MPT:
|
| 85 |
+
ret = self.system + self.sep
|
| 86 |
+
for role, message in messages:
|
| 87 |
+
if message:
|
| 88 |
+
if type(message) is tuple:
|
| 89 |
+
message, _, _ = message
|
| 90 |
+
ret += role + message + self.sep
|
| 91 |
+
else:
|
| 92 |
+
ret += role
|
| 93 |
+
elif self.sep_style == SeparatorStyle.LLAMA_2:
|
| 94 |
+
wrap_sys = lambda msg: f"<<SYS>>\n{msg}\n<</SYS>>\n\n" if len(msg) > 0 else msg
|
| 95 |
+
wrap_inst = lambda msg: f"[INST] {msg} [/INST]"
|
| 96 |
+
ret = ""
|
| 97 |
+
|
| 98 |
+
for i, (role, message) in enumerate(messages):
|
| 99 |
+
if i == 0:
|
| 100 |
+
assert message, "first message should not be none"
|
| 101 |
+
assert role == self.roles[0], "first message should come from user"
|
| 102 |
+
if message:
|
| 103 |
+
if type(message) is tuple:
|
| 104 |
+
message, _, _ = message
|
| 105 |
+
if i == 0: message = wrap_sys(self.system) + message
|
| 106 |
+
if i % 2 == 0:
|
| 107 |
+
message = wrap_inst(message)
|
| 108 |
+
ret += self.sep + message
|
| 109 |
+
else:
|
| 110 |
+
ret += " " + message + " " + self.sep2
|
| 111 |
+
else:
|
| 112 |
+
ret += ""
|
| 113 |
+
ret = ret.lstrip(self.sep)
|
| 114 |
+
elif self.sep_style == SeparatorStyle.TINY_LLAMA:
|
| 115 |
+
sep = "</s>"
|
| 116 |
+
wrap_sys = lambda msg: f"<|system|>\n{msg}\n"
|
| 117 |
+
wrap_user = lambda msg: f"<|user|>\n{msg}\n"
|
| 118 |
+
wrap_assistant = lambda msg: f"<|assistant|>\n{msg}"
|
| 119 |
+
ret = ""
|
| 120 |
+
|
| 121 |
+
for i, (role, message) in enumerate(messages):
|
| 122 |
+
if i == 0:
|
| 123 |
+
assert message, "first message should not be none"
|
| 124 |
+
assert role == self.roles[0], "first message should come from user"
|
| 125 |
+
if message:
|
| 126 |
+
if type(message) is tuple:
|
| 127 |
+
message, _, _ = message
|
| 128 |
+
if i % 2 == 0:
|
| 129 |
+
message = wrap_user(message)
|
| 130 |
+
if i == 0:
|
| 131 |
+
message = wrap_sys(self.system) + message
|
| 132 |
+
ret += self.sep + message
|
| 133 |
+
else:
|
| 134 |
+
message = wrap_assistant(message) + self.sep2
|
| 135 |
+
ret += message
|
| 136 |
+
else:
|
| 137 |
+
ret += "<|assistant|>\n"
|
| 138 |
+
ret = ret.lstrip(self.sep)
|
| 139 |
+
elif self.sep_style == SeparatorStyle.QWEN_2:
|
| 140 |
+
ret = self.system + self.sep
|
| 141 |
+
for role, message in messages:
|
| 142 |
+
if message:
|
| 143 |
+
if type(message) is tuple:
|
| 144 |
+
message, _, _ = message
|
| 145 |
+
ret += role + message + self.sep
|
| 146 |
+
else:
|
| 147 |
+
ret += role
|
| 148 |
+
elif self.sep_style == SeparatorStyle.PLAIN:
|
| 149 |
+
seps = [self.sep, self.sep2]
|
| 150 |
+
ret = self.system
|
| 151 |
+
for i, (role, message) in enumerate(messages):
|
| 152 |
+
if message:
|
| 153 |
+
if type(message) is tuple:
|
| 154 |
+
message, _, _ = message
|
| 155 |
+
ret += message + seps[i % 2]
|
| 156 |
+
else:
|
| 157 |
+
ret += ""
|
| 158 |
+
else:
|
| 159 |
+
raise ValueError(f"Invalid style: {self.sep_style}")
|
| 160 |
+
|
| 161 |
+
return ret
|
| 162 |
+
|
| 163 |
+
def append_message(self, role, message):
|
| 164 |
+
self.messages.append([role, message])
|
| 165 |
+
|
| 166 |
+
def get_images(self, return_pil=False):
|
| 167 |
+
images = []
|
| 168 |
+
for i, (role, msg) in enumerate(self.messages[self.offset:]):
|
| 169 |
+
if i % 2 == 0:
|
| 170 |
+
if type(msg) is tuple:
|
| 171 |
+
import base64
|
| 172 |
+
from io import BytesIO
|
| 173 |
+
from PIL import Image
|
| 174 |
+
msg, image, image_process_mode = msg
|
| 175 |
+
if image_process_mode == "Pad":
|
| 176 |
+
def expand2square(pil_img, background_color=(122, 116, 104)):
|
| 177 |
+
width, height = pil_img.size
|
| 178 |
+
if width == height:
|
| 179 |
+
return pil_img
|
| 180 |
+
elif width > height:
|
| 181 |
+
result = Image.new(pil_img.mode, (width, width), background_color)
|
| 182 |
+
result.paste(pil_img, (0, (width - height) // 2))
|
| 183 |
+
return result
|
| 184 |
+
else:
|
| 185 |
+
result = Image.new(pil_img.mode, (height, height), background_color)
|
| 186 |
+
result.paste(pil_img, ((height - width) // 2, 0))
|
| 187 |
+
return result
|
| 188 |
+
image = expand2square(image)
|
| 189 |
+
elif image_process_mode in ["Default", "Crop"]:
|
| 190 |
+
pass
|
| 191 |
+
elif image_process_mode == "Resize":
|
| 192 |
+
image = image.resize((336, 336))
|
| 193 |
+
else:
|
| 194 |
+
raise ValueError(f"Invalid image_process_mode: {image_process_mode}")
|
| 195 |
+
max_hw, min_hw = max(image.size), min(image.size)
|
| 196 |
+
aspect_ratio = max_hw / min_hw
|
| 197 |
+
max_len, min_len = 800, 400
|
| 198 |
+
shortest_edge = int(min(max_len / aspect_ratio, min_len, min_hw))
|
| 199 |
+
longest_edge = int(shortest_edge * aspect_ratio)
|
| 200 |
+
W, H = image.size
|
| 201 |
+
if longest_edge != max(image.size):
|
| 202 |
+
if H > W:
|
| 203 |
+
H, W = longest_edge, shortest_edge
|
| 204 |
+
else:
|
| 205 |
+
H, W = shortest_edge, longest_edge
|
| 206 |
+
image = image.resize((W, H))
|
| 207 |
+
if return_pil:
|
| 208 |
+
images.append(image)
|
| 209 |
+
else:
|
| 210 |
+
buffered = BytesIO()
|
| 211 |
+
image.save(buffered, format="PNG")
|
| 212 |
+
img_b64_str = base64.b64encode(buffered.getvalue()).decode()
|
| 213 |
+
images.append(img_b64_str)
|
| 214 |
+
return images
|
| 215 |
+
|
| 216 |
+
def to_gradio_chatbot(self):
|
| 217 |
+
ret = []
|
| 218 |
+
for i, (role, msg) in enumerate(self.messages[self.offset:]):
|
| 219 |
+
if i % 2 == 0:
|
| 220 |
+
if type(msg) is tuple:
|
| 221 |
+
import base64
|
| 222 |
+
from io import BytesIO
|
| 223 |
+
msg, image, image_process_mode = msg
|
| 224 |
+
max_hw, min_hw = max(image.size), min(image.size)
|
| 225 |
+
aspect_ratio = max_hw / min_hw
|
| 226 |
+
max_len, min_len = 800, 400
|
| 227 |
+
shortest_edge = int(min(max_len / aspect_ratio, min_len, min_hw))
|
| 228 |
+
longest_edge = int(shortest_edge * aspect_ratio)
|
| 229 |
+
W, H = image.size
|
| 230 |
+
if H > W:
|
| 231 |
+
H, W = longest_edge, shortest_edge
|
| 232 |
+
else:
|
| 233 |
+
H, W = shortest_edge, longest_edge
|
| 234 |
+
image = image.resize((W, H))
|
| 235 |
+
buffered = BytesIO()
|
| 236 |
+
image.save(buffered, format="JPEG")
|
| 237 |
+
img_b64_str = base64.b64encode(buffered.getvalue()).decode()
|
| 238 |
+
img_str = f'<img src="data:image/png;base64,{img_b64_str}" alt="user upload image" />'
|
| 239 |
+
msg = img_str + msg.replace('<image>', '').strip()
|
| 240 |
+
ret.append([msg, None])
|
| 241 |
+
else:
|
| 242 |
+
ret.append([msg, None])
|
| 243 |
+
else:
|
| 244 |
+
ret[-1][-1] = msg
|
| 245 |
+
return ret
|
| 246 |
+
|
| 247 |
+
def copy(self):
|
| 248 |
+
return Conversation(
|
| 249 |
+
system=self.system,
|
| 250 |
+
roles=self.roles,
|
| 251 |
+
messages=[[x, y] for x, y in self.messages],
|
| 252 |
+
offset=self.offset,
|
| 253 |
+
sep_style=self.sep_style,
|
| 254 |
+
sep=self.sep,
|
| 255 |
+
sep2=self.sep2,
|
| 256 |
+
version=self.version)
|
| 257 |
+
|
| 258 |
+
def dict(self):
|
| 259 |
+
if len(self.get_images()) > 0:
|
| 260 |
+
return {
|
| 261 |
+
"system": self.system,
|
| 262 |
+
"roles": self.roles,
|
| 263 |
+
"messages": [[x, y[0] if type(y) is tuple else y] for x, y in self.messages],
|
| 264 |
+
"offset": self.offset,
|
| 265 |
+
"sep": self.sep,
|
| 266 |
+
"sep2": self.sep2,
|
| 267 |
+
}
|
| 268 |
+
return {
|
| 269 |
+
"system": self.system,
|
| 270 |
+
"roles": self.roles,
|
| 271 |
+
"messages": self.messages,
|
| 272 |
+
"offset": self.offset,
|
| 273 |
+
"sep": self.sep,
|
| 274 |
+
"sep2": self.sep2,
|
| 275 |
+
}
|
| 276 |
+
|
| 277 |
+
|
| 278 |
+
|
| 279 |
+
|
| 280 |
+
conv_phi_v0 = Conversation(
|
| 281 |
+
system="A chat between a curious user and an artificial intelligence assistant. "
|
| 282 |
+
"The assistant gives helpful, detailed, and polite answers to the user's questions.",
|
| 283 |
+
roles=("USER", "ASSISTANT"),
|
| 284 |
+
version="phi",
|
| 285 |
+
messages=(),
|
| 286 |
+
offset=0,
|
| 287 |
+
sep_style=SeparatorStyle.TWO,
|
| 288 |
+
sep=" ",
|
| 289 |
+
sep2="<|endoftext|>",
|
| 290 |
+
)
|
| 291 |
+
|
| 292 |
+
|
| 293 |
+
|
| 294 |
+
def select_best_resolution(original_size, possible_resolutions):
|
| 295 |
+
"""
|
| 296 |
+
Selects the best resolution from a list of possible resolutions based on the original size.
|
| 297 |
+
|
| 298 |
+
Args:
|
| 299 |
+
original_size (tuple): The original size of the image in the format (width, height).
|
| 300 |
+
possible_resolutions (list): A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
|
| 301 |
+
|
| 302 |
+
Returns:
|
| 303 |
+
tuple: The best fit resolution in the format (width, height).
|
| 304 |
+
"""
|
| 305 |
+
original_width, original_height = original_size
|
| 306 |
+
best_fit = None
|
| 307 |
+
max_effective_resolution = 0
|
| 308 |
+
min_wasted_resolution = float('inf')
|
| 309 |
+
|
| 310 |
+
for width, height in possible_resolutions:
|
| 311 |
+
scale = min(width / original_width, height / original_height)
|
| 312 |
+
downscaled_width, downscaled_height = int(original_width * scale), int(original_height * scale)
|
| 313 |
+
effective_resolution = min(downscaled_width * downscaled_height, original_width * original_height)
|
| 314 |
+
wasted_resolution = (width * height) - effective_resolution
|
| 315 |
+
|
| 316 |
+
if effective_resolution > max_effective_resolution or (effective_resolution == max_effective_resolution and wasted_resolution < min_wasted_resolution):
|
| 317 |
+
max_effective_resolution = effective_resolution
|
| 318 |
+
min_wasted_resolution = wasted_resolution
|
| 319 |
+
best_fit = (width, height)
|
| 320 |
+
|
| 321 |
+
return best_fit
|
| 322 |
+
|
| 323 |
+
|
| 324 |
+
## added by llava-1.6
|
| 325 |
+
def resize_and_pad_image(image, target_resolution):
|
| 326 |
+
"""
|
| 327 |
+
Resize and pad an image to a target resolution while maintaining aspect ratio.
|
| 328 |
+
|
| 329 |
+
Args:
|
| 330 |
+
image (PIL.Image.Image): The input image.
|
| 331 |
+
target_resolution (tuple): The target resolution (width, height) of the image.
|
| 332 |
+
|
| 333 |
+
Returns:
|
| 334 |
+
PIL.Image.Image: The resized and padded image.
|
| 335 |
+
"""
|
| 336 |
+
original_width, original_height = image.size
|
| 337 |
+
target_width, target_height = target_resolution
|
| 338 |
+
|
| 339 |
+
scale_w = target_width / original_width
|
| 340 |
+
scale_h = target_height / original_height
|
| 341 |
+
|
| 342 |
+
if scale_w < scale_h:
|
| 343 |
+
new_width = target_width
|
| 344 |
+
new_height = min(math.ceil(original_height * scale_w), target_height)
|
| 345 |
+
else:
|
| 346 |
+
new_height = target_height
|
| 347 |
+
new_width = min(math.ceil(original_width * scale_h), target_width)
|
| 348 |
+
|
| 349 |
+
# Resize the image
|
| 350 |
+
resized_image = image.resize((new_width, new_height))
|
| 351 |
+
|
| 352 |
+
new_image = Image.new('RGB', (target_width, target_height), (0, 0, 0))
|
| 353 |
+
paste_x = (target_width - new_width) // 2
|
| 354 |
+
paste_y = (target_height - new_height) // 2
|
| 355 |
+
new_image.paste(resized_image, (paste_x, paste_y))
|
| 356 |
+
|
| 357 |
+
return new_image
|
| 358 |
+
|
| 359 |
+
|
| 360 |
+
## added by llava-1.6
|
| 361 |
+
def divide_to_patches(image, patch_size):
|
| 362 |
+
"""
|
| 363 |
+
Divides an image into patches of a specified size.
|
| 364 |
+
|
| 365 |
+
Args:
|
| 366 |
+
image (PIL.Image.Image): The input image.
|
| 367 |
+
patch_size (int): The size of each patch.
|
| 368 |
+
|
| 369 |
+
Returns:
|
| 370 |
+
list: A list of PIL.Image.Image objects representing the patches.
|
| 371 |
+
"""
|
| 372 |
+
patches = []
|
| 373 |
+
width, height = image.size
|
| 374 |
+
for i in range(0, height, patch_size):
|
| 375 |
+
for j in range(0, width, patch_size):
|
| 376 |
+
box = (j, i, j + patch_size, i + patch_size)
|
| 377 |
+
patch = image.crop(box)
|
| 378 |
+
patches.append(patch)
|
| 379 |
+
|
| 380 |
+
return patches
|
| 381 |
+
|
| 382 |
+
|
| 383 |
+
## added by llava-1.6
|
| 384 |
+
def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size):
|
| 385 |
+
"""
|
| 386 |
+
Calculate the shape of the image patch grid after the preprocessing for images of any resolution.
|
| 387 |
+
|
| 388 |
+
Args:
|
| 389 |
+
image_size (tuple): The size of the input image in the format (width, height).
|
| 390 |
+
grid_pinpoints (str): A string representation of a list of possible resolutions.
|
| 391 |
+
patch_size (int): The size of each image patch.
|
| 392 |
+
|
| 393 |
+
Returns:
|
| 394 |
+
tuple: The shape of the image patch grid in the format (width, height).
|
| 395 |
+
"""
|
| 396 |
+
if type(grid_pinpoints) is list:
|
| 397 |
+
possible_resolutions = grid_pinpoints
|
| 398 |
+
else:
|
| 399 |
+
possible_resolutions = ast.literal_eval(grid_pinpoints)
|
| 400 |
+
width, height = select_best_resolution(image_size, possible_resolutions)
|
| 401 |
+
return width // patch_size, height // patch_size
|
| 402 |
+
|
| 403 |
+
|
| 404 |
+
## added by llava-1.6
|
| 405 |
+
def process_anyres_image(image, processor, grid_pinpoints):
|
| 406 |
+
"""
|
| 407 |
+
Process an image with variable resolutions.
|
| 408 |
+
|
| 409 |
+
Args:
|
| 410 |
+
image (PIL.Image.Image): The input image to be processed.
|
| 411 |
+
processor: The image processor object.
|
| 412 |
+
grid_pinpoints (str): A string representation of a list of possible resolutions.
|
| 413 |
+
|
| 414 |
+
Returns:
|
| 415 |
+
torch.Tensor: A tensor containing the processed image patches.
|
| 416 |
+
"""
|
| 417 |
+
if type(grid_pinpoints) is list:
|
| 418 |
+
possible_resolutions = grid_pinpoints
|
| 419 |
+
else:
|
| 420 |
+
possible_resolutions = ast.literal_eval(grid_pinpoints)
|
| 421 |
+
best_resolution = select_best_resolution(image.size, possible_resolutions)
|
| 422 |
+
image_padded = resize_and_pad_image(image, best_resolution)
|
| 423 |
+
|
| 424 |
+
patches = divide_to_patches(image_padded, processor.crop_size['height'])
|
| 425 |
+
|
| 426 |
+
image_original_resize = image.resize((processor.size['shortest_edge'], processor.size['shortest_edge']))
|
| 427 |
+
|
| 428 |
+
image_patches = [image_original_resize] + patches
|
| 429 |
+
image_patches = [processor.preprocess(image_patch, return_tensors='pt')['pixel_values'][0]
|
| 430 |
+
for image_patch in image_patches]
|
| 431 |
+
return torch.stack(image_patches, dim=0)
|
| 432 |
+
|
| 433 |
+
|
| 434 |
+
def load_image_from_base64(image):
|
| 435 |
+
return Image.open(BytesIO(base64.b64decode(image)))
|
| 436 |
+
|
| 437 |
+
|
| 438 |
+
def expand2square(pil_img, background_color):
|
| 439 |
+
width, height = pil_img.size
|
| 440 |
+
if width == height:
|
| 441 |
+
return pil_img
|
| 442 |
+
elif width > height:
|
| 443 |
+
result = Image.new(pil_img.mode, (width, width), background_color)
|
| 444 |
+
result.paste(pil_img, (0, (width - height) // 2))
|
| 445 |
+
return result
|
| 446 |
+
else:
|
| 447 |
+
result = Image.new(pil_img.mode, (height, height), background_color)
|
| 448 |
+
result.paste(pil_img, ((height - width) // 2, 0))
|
| 449 |
+
return result
|
| 450 |
+
|
| 451 |
+
|
| 452 |
+
def process_images(images, image_processor, model_cfg):
|
| 453 |
+
image_aspect_ratio = getattr(model_cfg, "image_aspect_ratio", None)
|
| 454 |
+
new_images = []
|
| 455 |
+
if image_aspect_ratio == 'pad':
|
| 456 |
+
for image in images:
|
| 457 |
+
image = expand2square(image, tuple(int(x*255) for x in image_processor.image_mean))
|
| 458 |
+
image = image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
|
| 459 |
+
new_images.append(image)
|
| 460 |
+
elif image_aspect_ratio == "anyres":
|
| 461 |
+
for image in images:
|
| 462 |
+
image = process_anyres_image(image, image_processor, model_cfg.image_grid_pinpoints)
|
| 463 |
+
new_images.append(image)
|
| 464 |
+
else:
|
| 465 |
+
return image_processor(images, return_tensors='pt')['pixel_values']
|
| 466 |
+
if all(x.shape == new_images[0].shape for x in new_images):
|
| 467 |
+
new_images = torch.stack(new_images, dim=0)
|
| 468 |
+
return new_images
|
| 469 |
+
|
| 470 |
+
|
| 471 |
+
def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
|
| 472 |
+
prompt_chunks = [tokenizer(chunk).input_ids for chunk in prompt.split('<image>')]
|
| 473 |
+
|
| 474 |
+
def insert_separator(X, sep):
|
| 475 |
+
return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]
|
| 476 |
+
|
| 477 |
+
input_ids = []
|
| 478 |
+
offset = 0
|
| 479 |
+
if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
|
| 480 |
+
offset = 1
|
| 481 |
+
input_ids.append(prompt_chunks[0][0])
|
| 482 |
+
|
| 483 |
+
for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
|
| 484 |
+
input_ids.extend(x[offset:])
|
| 485 |
+
|
| 486 |
+
if return_tensors is not None:
|
| 487 |
+
if return_tensors == 'pt':
|
| 488 |
+
return torch.tensor(input_ids, dtype=torch.long)
|
| 489 |
+
raise ValueError(f'Unsupported tensor type: {return_tensors}')
|
| 490 |
+
return input_ids
|
| 491 |
+
|
| 492 |
+
|
| 493 |
+
def get_model_name_from_path(model_path):
|
| 494 |
+
model_path = model_path.strip("/")
|
| 495 |
+
model_paths = model_path.split("/")
|
| 496 |
+
if model_paths[-1].startswith('checkpoint-'):
|
| 497 |
+
return model_paths[-2] + "_" + model_paths[-1]
|
| 498 |
+
else:
|
| 499 |
+
return model_paths[-1]
|
| 500 |
+
|
| 501 |
+
|
| 502 |
+
class KeywordsStoppingCriteria(StoppingCriteria):
|
| 503 |
+
def __init__(self, keywords, tokenizer, input_ids):
|
| 504 |
+
self.keywords = keywords
|
| 505 |
+
self.keyword_ids = []
|
| 506 |
+
self.max_keyword_len = 0
|
| 507 |
+
for keyword in keywords:
|
| 508 |
+
cur_keyword_ids = tokenizer(keyword).input_ids
|
| 509 |
+
if len(cur_keyword_ids) > 1 and cur_keyword_ids[0] == tokenizer.bos_token_id:
|
| 510 |
+
cur_keyword_ids = cur_keyword_ids[1:]
|
| 511 |
+
if len(cur_keyword_ids) > self.max_keyword_len:
|
| 512 |
+
self.max_keyword_len = len(cur_keyword_ids)
|
| 513 |
+
self.keyword_ids.append(torch.tensor(cur_keyword_ids))
|
| 514 |
+
self.tokenizer = tokenizer
|
| 515 |
+
self.start_len = input_ids.shape[1]
|
| 516 |
+
|
| 517 |
+
def call_for_batch(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
|
| 518 |
+
offset = min(output_ids.shape[1] - self.start_len, self.max_keyword_len)
|
| 519 |
+
self.keyword_ids = [keyword_id.to(output_ids.device) for keyword_id in self.keyword_ids]
|
| 520 |
+
for keyword_id in self.keyword_ids:
|
| 521 |
+
if (output_ids[0, -keyword_id.shape[0]:] == keyword_id).all():
|
| 522 |
+
return True
|
| 523 |
+
outputs = self.tokenizer.batch_decode(output_ids[:, -offset:], skip_special_tokens=True)[0]
|
| 524 |
+
for keyword in self.keywords:
|
| 525 |
+
if keyword in outputs:
|
| 526 |
+
return True
|
| 527 |
+
return False
|
| 528 |
+
|
| 529 |
+
def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
|
| 530 |
+
outputs = []
|
| 531 |
+
for i in range(output_ids.shape[0]):
|
| 532 |
+
outputs.append(self.call_for_batch(output_ids[i].unsqueeze(0), scores))
|
| 533 |
+
return all(outputs)
|
| 534 |
+
|
| 535 |
+
|
| 536 |
+
|
| 537 |
+
def load_image(image_file):
|
| 538 |
+
if image_file.startswith("http") or image_file.startswith("https"):
|
| 539 |
+
response = requests.get(image_file)
|
| 540 |
+
image = Image.open(BytesIO(response.content)).convert("RGB")
|
| 541 |
+
else:
|
| 542 |
+
image = Image.open(image_file).convert("RGB")
|
| 543 |
+
return image
|
modeling_tinyllava_phi.py
CHANGED
|
@@ -16,6 +16,7 @@ from transformers import CLIPVisionModel, CLIPImageProcessor, SiglipVisionModel,
|
|
| 16 |
from .configuration import TinyLlavaConfig, IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN
|
| 17 |
|
| 18 |
from transformers import AutoConfig, AutoModelForCausalLM, PhiForCausalLM
|
|
|
|
| 19 |
|
| 20 |
# from tinyllava.utils.data_utils import get_value_from_kwargs
|
| 21 |
CONTROLLER_HEART_BEAT_EXPIRATION = 30
|
|
@@ -414,6 +415,60 @@ class TinyLlavaForConditionalGeneration(TinyLlavaPreTrainedModel):
|
|
| 414 |
position_ids = None
|
| 415 |
|
| 416 |
return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 417 |
|
| 418 |
AutoConfig.register("tinyllava", TinyLlavaConfig)
|
| 419 |
AutoModelForCausalLM.register(TinyLlavaConfig, TinyLlavaForConditionalGeneration)
|
|
|
|
| 16 |
from .configuration import TinyLlavaConfig, IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN
|
| 17 |
|
| 18 |
from transformers import AutoConfig, AutoModelForCausalLM, PhiForCausalLM
|
| 19 |
+
from data_preprocess import *
|
| 20 |
|
| 21 |
# from tinyllava.utils.data_utils import get_value_from_kwargs
|
| 22 |
CONTROLLER_HEART_BEAT_EXPIRATION = 30
|
|
|
|
| 415 |
position_ids = None
|
| 416 |
|
| 417 |
return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels
|
| 418 |
+
|
| 419 |
+
def chat(
|
| 420 |
+
self,
|
| 421 |
+
prompt: str,
|
| 422 |
+
tokenizer = None,
|
| 423 |
+
image: str = None,
|
| 424 |
+
max_new_tokens: int = 512,
|
| 425 |
+
num_beams = 1,
|
| 426 |
+
top_p=None,
|
| 427 |
+
temperature=0
|
| 428 |
+
):
|
| 429 |
+
image_processor = self.vision_tower._image_processor
|
| 430 |
+
|
| 431 |
+
if image is not None:
|
| 432 |
+
prompt = DEFAULT_IMAGE_TOKEN + '\n' + prompt
|
| 433 |
+
conv = conv_phi_v0.copy()
|
| 434 |
+
conv.append_message(conv.roles[0], prompt)
|
| 435 |
+
conv.append_message(conv.roles[1], None)
|
| 436 |
+
prompt = conv.get_prompt()
|
| 437 |
+
if image is not None:
|
| 438 |
+
image = load_image(image)
|
| 439 |
+
image_tensor = process_images(image, image_processor, self.config).to(self.device)
|
| 440 |
+
|
| 441 |
+
input_ids = (
|
| 442 |
+
tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt")
|
| 443 |
+
.unsqueeze(0).to(self.device)
|
| 444 |
+
)
|
| 445 |
+
# Generate
|
| 446 |
+
stime = time.time()
|
| 447 |
+
|
| 448 |
+
with torch.inference_mode():
|
| 449 |
+
output_ids = self.generate(
|
| 450 |
+
input_ids,
|
| 451 |
+
images=image_tensor,
|
| 452 |
+
do_sample=True if temperature > 0 else False,
|
| 453 |
+
temperature=temperature,
|
| 454 |
+
top_p=top_p,
|
| 455 |
+
num_beams=num_beams,
|
| 456 |
+
pad_token_id=tokenizer.pad_token_id,
|
| 457 |
+
max_new_tokens=max_new_tokens,
|
| 458 |
+
use_cache=True,
|
| 459 |
+
# stopping_criteria=[stopping_criteria],
|
| 460 |
+
)
|
| 461 |
+
|
| 462 |
+
# print('inference over')
|
| 463 |
+
generation_time = time.time() - stime
|
| 464 |
+
outputs = tokenizer.batch_decode(
|
| 465 |
+
output_ids, skip_special_tokens=True
|
| 466 |
+
)[0]
|
| 467 |
+
|
| 468 |
+
outputs = outputs.strip()
|
| 469 |
+
|
| 470 |
+
return outputs, generation_time
|
| 471 |
+
|
| 472 |
|
| 473 |
AutoConfig.register("tinyllava", TinyLlavaConfig)
|
| 474 |
AutoModelForCausalLM.register(TinyLlavaConfig, TinyLlavaForConditionalGeneration)
|