Datasets:
Update terramesh.py
Browse files- terramesh.py +84 -9
terramesh.py
CHANGED
@@ -19,18 +19,20 @@
|
|
19 |
import os
|
20 |
import io
|
21 |
import re
|
|
|
|
|
22 |
import zarr
|
23 |
import fsspec
|
24 |
import itertools
|
25 |
import braceexpand
|
26 |
import numpy as np
|
|
|
27 |
import webdataset as wds
|
28 |
from collections.abc import Callable, Iterable
|
29 |
from sympy.printing.pytorch import torch
|
30 |
from torch.utils.data._utils.collate import default_collate
|
31 |
from webdataset.handlers import warn_and_continue
|
32 |
|
33 |
-
|
34 |
# Definition of all shard files in TerraMesh
|
35 |
split_files = {
|
36 |
"ssl4eos12": {
|
@@ -56,7 +58,6 @@ def build_terramesh_dataset(
|
|
56 |
batch_size: int = 8,
|
57 |
*args, **kwargs,
|
58 |
):
|
59 |
-
|
60 |
if len(modalities) == 1:
|
61 |
# Build standard WebDataset for single modality
|
62 |
dataset = build_wds_dataset(
|
@@ -172,12 +173,11 @@ def build_multimodal_dataset(
|
|
172 |
lst.remove(value)
|
173 |
return lst
|
174 |
|
175 |
-
|
176 |
majortom_mod = f"[{','.join(filter_list(modalities, 'S1GRD'))}]"
|
177 |
ssl4eos12_mod = f"[{','.join(filter_list(modalities, 'S1RTC'))}]"
|
178 |
|
179 |
# Joins majortom and ssl4eos12 shard files with "::"
|
180 |
-
urls = (os.path.join(path, split, majortom_mod,
|
181 |
+ "::" + os.path.join(path, split, ssl4eos12_mod, split_files["ssl4eos12"][split][0]))
|
182 |
|
183 |
dataset = build_datapipeline(urls, transform, batch_size, *args, **kwargs)
|
@@ -193,12 +193,12 @@ def build_datapipeline(urls, transform, batch_size, *args, **kwargs):
|
|
193 |
wds.decode(zarr_decoder), # Decode from bytes to PIL images, numpy arrays, etc.
|
194 |
wds.map(drop_time_dim), # Remove time dimension from tensors
|
195 |
wds.map(remove_extensions), # Remove "file extensions" from dictionary keys
|
196 |
-
(
|
197 |
wds.map(transform)
|
198 |
if transform is not None
|
199 |
else wds.map(identity)
|
200 |
),
|
201 |
-
(
|
202 |
wds.batched(batch_size, collation_fn=default_collate, partial=False)
|
203 |
if batch_size is not None
|
204 |
else wds.map(identity)
|
@@ -240,8 +240,8 @@ def remove_extensions(sample):
|
|
240 |
|
241 |
|
242 |
def multi_tarfile_samples(
|
243 |
-
|
244 |
-
|
245 |
):
|
246 |
"""
|
247 |
This function is adapted from https://github.com/apple/ml-4m/blob/main/fourm/data/unified_datasets.py.
|
@@ -289,7 +289,7 @@ def multi_tarfile_samples(
|
|
289 |
merged_dict["__url__"] = src["url"]
|
290 |
|
291 |
for modality_name, modality_dict in zip(
|
292 |
-
|
293 |
):
|
294 |
_key = modality_dict.pop("__key__")
|
295 |
_url = modality_dict.pop("__url__")
|
@@ -315,3 +315,78 @@ def multi_tarfile_samples(
|
|
315 |
continue
|
316 |
else:
|
317 |
break
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
import os
|
20 |
import io
|
21 |
import re
|
22 |
+
|
23 |
+
import numpy
|
24 |
import zarr
|
25 |
import fsspec
|
26 |
import itertools
|
27 |
import braceexpand
|
28 |
import numpy as np
|
29 |
+
import albumentations
|
30 |
import webdataset as wds
|
31 |
from collections.abc import Callable, Iterable
|
32 |
from sympy.printing.pytorch import torch
|
33 |
from torch.utils.data._utils.collate import default_collate
|
34 |
from webdataset.handlers import warn_and_continue
|
35 |
|
|
|
36 |
# Definition of all shard files in TerraMesh
|
37 |
split_files = {
|
38 |
"ssl4eos12": {
|
|
|
58 |
batch_size: int = 8,
|
59 |
*args, **kwargs,
|
60 |
):
|
|
|
61 |
if len(modalities) == 1:
|
62 |
# Build standard WebDataset for single modality
|
63 |
dataset = build_wds_dataset(
|
|
|
173 |
lst.remove(value)
|
174 |
return lst
|
175 |
|
|
|
176 |
majortom_mod = f"[{','.join(filter_list(modalities, 'S1GRD'))}]"
|
177 |
ssl4eos12_mod = f"[{','.join(filter_list(modalities, 'S1RTC'))}]"
|
178 |
|
179 |
# Joins majortom and ssl4eos12 shard files with "::"
|
180 |
+
urls = (os.path.join(path, split, majortom_mod, split_files["majortom"][split][0])
|
181 |
+ "::" + os.path.join(path, split, ssl4eos12_mod, split_files["ssl4eos12"][split][0]))
|
182 |
|
183 |
dataset = build_datapipeline(urls, transform, batch_size, *args, **kwargs)
|
|
|
193 |
wds.decode(zarr_decoder), # Decode from bytes to PIL images, numpy arrays, etc.
|
194 |
wds.map(drop_time_dim), # Remove time dimension from tensors
|
195 |
wds.map(remove_extensions), # Remove "file extensions" from dictionary keys
|
196 |
+
( # Apply transformation
|
197 |
wds.map(transform)
|
198 |
if transform is not None
|
199 |
else wds.map(identity)
|
200 |
),
|
201 |
+
( # Batching
|
202 |
wds.batched(batch_size, collation_fn=default_collate, partial=False)
|
203 |
if batch_size is not None
|
204 |
else wds.map(identity)
|
|
|
240 |
|
241 |
|
242 |
def multi_tarfile_samples(
|
243 |
+
src_iter: Iterable[dict],
|
244 |
+
handler: Callable[[Exception], bool] = warn_and_continue,
|
245 |
):
|
246 |
"""
|
247 |
This function is adapted from https://github.com/apple/ml-4m/blob/main/fourm/data/unified_datasets.py.
|
|
|
289 |
merged_dict["__url__"] = src["url"]
|
290 |
|
291 |
for modality_name, modality_dict in zip(
|
292 |
+
modality_names, multi_tar_files
|
293 |
):
|
294 |
_key = modality_dict.pop("__key__")
|
295 |
_url = modality_dict.pop("__url__")
|
|
|
315 |
continue
|
316 |
else:
|
317 |
break
|
318 |
+
|
319 |
+
|
320 |
+
class Transpose(albumentations.ImageOnlyTransform):
|
321 |
+
"""
|
322 |
+
Rearrange is a generic image transformation that reshapes an input tensor using a custom einops pattern.
|
323 |
+
|
324 |
+
This transform allows flexible reordering of tensor dimensions based on the provided pattern and arguments.
|
325 |
+
"""
|
326 |
+
|
327 |
+
def __init__(self, axis: list):
|
328 |
+
"""
|
329 |
+
Initialize the Transpose transform.
|
330 |
+
|
331 |
+
Args:
|
332 |
+
axis (list): Axis for numpy.transpose.
|
333 |
+
"""
|
334 |
+
super().__init__(p=1)
|
335 |
+
self.axis = axis
|
336 |
+
|
337 |
+
def apply(self, img, **params):
|
338 |
+
return numpy.transpose(img, self.axis)
|
339 |
+
|
340 |
+
def get_transform_init_args_names(self):
|
341 |
+
return "transpose"
|
342 |
+
|
343 |
+
|
344 |
+
def default_non_image_transform(array):
|
345 |
+
if hasattr(array, 'dtype') and (array.dtype == float or array.dtype == int):
|
346 |
+
return torch.from_numpy(array)
|
347 |
+
else:
|
348 |
+
return array
|
349 |
+
|
350 |
+
|
351 |
+
class MultimodalTransforms:
|
352 |
+
"""
|
353 |
+
MultimodalTransforms applies albumentations transforms to multiple image modalities.
|
354 |
+
|
355 |
+
This class supports both shared transformations across modalities and separate transformations for each modality.
|
356 |
+
It also handles non-image modalities by applying a specified non-image transform.
|
357 |
+
|
358 |
+
This code is adapted from https://github.com/IBM/terratorch/blob/main/terratorch/datasets/transforms.py.
|
359 |
+
"""
|
360 |
+
|
361 |
+
def __init__(
|
362 |
+
self,
|
363 |
+
transforms: dict | albumentations.Compose,
|
364 |
+
shared: bool = True,
|
365 |
+
non_image_modalities: list[str] | None = None,
|
366 |
+
non_image_transforms: object | None = None,
|
367 |
+
):
|
368 |
+
"""
|
369 |
+
Initialize the MultimodalTransforms.
|
370 |
+
|
371 |
+
Args:
|
372 |
+
transforms (dict or A.Compose): The transformation(s) to apply to the data.
|
373 |
+
non_image_modalities (list[str] | None): List of keys corresponding to non-image modalities.
|
374 |
+
non_image_transforms (object | None): A transform to apply to non-image modalities.
|
375 |
+
If None, a default transform is used.
|
376 |
+
"""
|
377 |
+
self.transforms = transforms
|
378 |
+
self.non_image_modalities = non_image_modalities or []
|
379 |
+
self.non_image_transforms = non_image_transforms or default_non_image_transform
|
380 |
+
|
381 |
+
def __call__(self, data: dict):
|
382 |
+
# albumentations requires a key 'image' and treats all other keys as additional targets
|
383 |
+
image_modality = [k for k in data.keys() if k not in self.non_image_modalities][0]
|
384 |
+
data['image'] = data.pop(image_modality)
|
385 |
+
data = self.transforms(**data)
|
386 |
+
data[image_modality] = data.pop('image')
|
387 |
+
|
388 |
+
# Process sequence data which is ignored by albumentations as 'global_label'
|
389 |
+
for modality in self.non_image_modalities:
|
390 |
+
data[modality] = self.non_image_transforms(data[modality])
|
391 |
+
|
392 |
+
return data
|