Chris Oswald
commited on
Commit
•
76efad4
1
Parent(s):
c5eade3
refactored to enable sharding
Browse files
SPIDER.py
CHANGED
@@ -16,7 +16,7 @@
|
|
16 |
# Import packages
|
17 |
import csv
|
18 |
import os
|
19 |
-
from typing import Dict, List, Mapping, Optional, Sequence, Tuple, Union
|
20 |
|
21 |
import numpy as np
|
22 |
import pandas as pd
|
@@ -35,6 +35,13 @@ def import_csv_data(filepath: str) -> List[Dict[str, str]]:
|
|
35 |
results.append(line)
|
36 |
return results
|
37 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
38 |
def standardize_3D_image(
|
39 |
image: np.ndarray,
|
40 |
resize_shape: Tuple[int, int, int]
|
@@ -130,10 +137,21 @@ class CustomBuilderConfig(datasets.BuilderConfig):
|
|
130 |
description: Optional[str] = None,
|
131 |
scan_types: List[str] = DEFAULT_SCAN_TYPES,
|
132 |
resize_shape: Tuple[int, int, int] = DEFAULT_RESIZE,
|
|
|
133 |
):
|
134 |
super().__init__(name, version, data_dir, data_files, description)
|
135 |
-
self.scan_types = scan_types
|
136 |
self.resize_shape = resize_shape
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
137 |
|
138 |
|
139 |
class SPIDER(datasets.GeneratorBasedBuilder):
|
@@ -143,30 +161,44 @@ class SPIDER(datasets.GeneratorBasedBuilder):
|
|
143 |
DEFAULT_WRITER_BATCH_SIZE = 16 # PyArrow default is too large for image data
|
144 |
VERSION = datasets.Version("1.1.0")
|
145 |
BUILDER_CONFIG_CLASS = CustomBuilderConfig
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
157 |
|
158 |
def _info(self):
|
159 |
"""Specify datasets.DatasetInfo object containing information and typing
|
160 |
for the dataset."""
|
161 |
|
162 |
-
image_size = self.config.resize_shape
|
163 |
features = datasets.Features({
|
164 |
"patient_id": datasets.Value("string"),
|
165 |
"scan_type": datasets.Value("string"),
|
|
|
|
|
|
|
|
|
166 |
"image_path": datasets.Value("string"),
|
167 |
"mask_path": datasets.Value("string"),
|
168 |
-
"image_array": datasets.Array3D(shape=image_size, dtype='uint8'),
|
169 |
-
"mask_array": datasets.Array3D(shape=image_size, dtype='uint8'),
|
170 |
"metadata": {
|
171 |
"num_vertebrae": datasets.Value(dtype="string"), #TODO: more specific types
|
172 |
"num_discs": datasets.Value(dtype="string"),
|
@@ -228,87 +260,31 @@ class SPIDER(datasets.GeneratorBasedBuilder):
|
|
228 |
citation=_CITATION,
|
229 |
)
|
230 |
|
231 |
-
def _split_generators(
|
232 |
-
"""Download and extract data and define splits based on configuration."""
|
233 |
-
|
234 |
-
paths_dict = dl_manager.download_and_extract(_URLS)
|
235 |
-
return [
|
236 |
-
datasets.SplitGenerator(
|
237 |
-
name=datasets.Split.TRAIN,
|
238 |
-
gen_kwargs={
|
239 |
-
"paths_dict": paths_dict,
|
240 |
-
"split": "train",
|
241 |
-
"scan_types": self.scan_types,
|
242 |
-
"resize_shape": self.resize_shape,
|
243 |
-
},
|
244 |
-
),
|
245 |
-
datasets.SplitGenerator(
|
246 |
-
name=datasets.Split.VALIDATION,
|
247 |
-
gen_kwargs={
|
248 |
-
"paths_dict": paths_dict,
|
249 |
-
"split": "validate",
|
250 |
-
"scan_types": self.scan_types,
|
251 |
-
"resize_shape": self.resize_shape,
|
252 |
-
},
|
253 |
-
),
|
254 |
-
datasets.SplitGenerator(
|
255 |
-
name=datasets.Split.TEST,
|
256 |
-
gen_kwargs={
|
257 |
-
"paths_dict": paths_dict,
|
258 |
-
"split": "test",
|
259 |
-
"scan_types": self.scan_types,
|
260 |
-
"resize_shape": self.resize_shape,
|
261 |
-
},
|
262 |
-
),
|
263 |
-
]
|
264 |
-
|
265 |
-
def _generate_examples(
|
266 |
self,
|
267 |
-
|
268 |
-
split: str,
|
269 |
-
scan_types: List[str],
|
270 |
-
resize_shape: Tuple[int, int, int],
|
271 |
validate_share: float = 0.2,
|
272 |
test_share: float = 0.2,
|
273 |
random_seed: int = 9999,
|
274 |
-
)
|
275 |
"""
|
276 |
-
|
277 |
-
(key, example) tuples from the dataset. The `key` is for legacy reasons
|
278 |
-
(tfds) and is not important in itself, but must be unique for each example.
|
279 |
|
280 |
Args
|
281 |
-
|
282 |
-
split: specify training, validation, or testing set;
|
283 |
-
options = 'train', 'validate', OR 'test'
|
284 |
-
scan_types: list of sagittal scan types to use in examples;
|
285 |
-
options = ['t1', 't2', 't2_SPACE']
|
286 |
validate_share: float indicating share of data to use for validation;
|
287 |
must be in range (0.0, 1.0); note that training share is
|
288 |
calculated as (1 - validate_share - test_share)
|
289 |
test_share: float indicating share of data to use for testing;
|
290 |
must be in range (0.0, 1.0); note that training share is
|
291 |
calculated as (1 - validate_share - test_share)
|
292 |
-
|
293 |
-
Yields
|
294 |
-
Tuple (unique patient-scan ID, dict of
|
295 |
"""
|
296 |
# Set constants
|
297 |
train_share = (1.0 - validate_share - test_share)
|
298 |
np.random.seed(int(random_seed))
|
299 |
|
300 |
# Validate params
|
301 |
-
for item in scan_types:
|
302 |
-
if item not in ['t1', 't2', 't2_SPACE']:
|
303 |
-
raise ValueError(
|
304 |
-
'Scan type "{item}" not recognized as valid scan type.\
|
305 |
-
Verify scan type argument.'
|
306 |
-
)
|
307 |
-
if split not in ['train', 'validate', 'test']:
|
308 |
-
raise ValueError(
|
309 |
-
f'Split argument "{split}" is not recognized. \
|
310 |
-
Please enter one of ["train", "validate", "test"]'
|
311 |
-
)
|
312 |
if train_share <= 0.0:
|
313 |
raise ValueError(
|
314 |
f'Training share is calculated as (1 - validate_share - test_share) \
|
@@ -327,6 +303,33 @@ class SPIDER(datasets.GeneratorBasedBuilder):
|
|
327 |
{test_share}.'
|
328 |
)
|
329 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
330 |
# Generate train/validate/test partitions of patient IDs
|
331 |
partition = np.random.choice(
|
332 |
['train', 'dev', 'test'],
|
@@ -339,6 +342,19 @@ class SPIDER(datasets.GeneratorBasedBuilder):
|
|
339 |
test_ids = set(patient_ids[partition == 'test'])
|
340 |
assert len(train_ids.union(validate_ids, test_ids)) == N_PATIENTS
|
341 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
342 |
# Import patient/scanner data and radiological gradings data
|
343 |
overview_data = import_csv_data(paths_dict['overview'])
|
344 |
grades_data = import_csv_data(paths_dict['gradings'])
|
@@ -389,54 +405,74 @@ class SPIDER(datasets.GeneratorBasedBuilder):
|
|
389 |
if col not in ['Patient']
|
390 |
}
|
391 |
|
392 |
-
#
|
393 |
-
|
394 |
-
|
395 |
-
|
396 |
-
|
397 |
-
|
398 |
-
|
399 |
-
|
400 |
-
file for file in os.listdir(os.path.join(paths_dict['masks'], 'masks'))
|
401 |
-
if file.endswith('.mha')
|
402 |
-
]
|
403 |
-
assert len(mask_files) > 0, "No mask files found--check directory path."
|
404 |
-
|
405 |
-
# Filter image and mask data files based on scan types
|
406 |
-
image_files = [
|
407 |
-
file for file in image_files
|
408 |
-
if any(scan_type in file for scan_type in scan_types)
|
409 |
-
]
|
410 |
|
411 |
-
|
412 |
-
|
413 |
-
|
414 |
-
|
415 |
-
|
416 |
-
|
417 |
-
|
418 |
-
|
419 |
-
|
420 |
-
|
421 |
-
|
422 |
-
|
423 |
-
|
424 |
-
|
425 |
-
|
426 |
-
|
427 |
-
|
428 |
-
|
429 |
-
|
430 |
-
|
431 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
432 |
]
|
433 |
-
|
434 |
-
|
435 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
436 |
# Shuffle order of patient scans
|
437 |
# (note that only images need to be shuffled since masks and metadata
|
438 |
# will be linked to the selected image)
|
439 |
-
|
|
|
440 |
|
441 |
## Generate next example
|
442 |
# ----------------------
|
@@ -483,10 +519,10 @@ class SPIDER(datasets.GeneratorBasedBuilder):
|
|
483 |
return_dict = {
|
484 |
'patient_id':patient_id,
|
485 |
'scan_type':scan_type,
|
|
|
|
|
486 |
'image_path':image_path,
|
487 |
'mask_path':mask_path,
|
488 |
-
'image_array':image_array_standardized,
|
489 |
-
'mask_array':mask_array_standardized,
|
490 |
'metadata':image_overview,
|
491 |
'rad_gradings':patient_grades_dict,
|
492 |
}
|
|
|
16 |
# Import packages
|
17 |
import csv
|
18 |
import os
|
19 |
+
from typing import Dict, List, Mapping, Optional, Sequence, Set, Tuple, Union
|
20 |
|
21 |
import numpy as np
|
22 |
import pandas as pd
|
|
|
35 |
results.append(line)
|
36 |
return results
|
37 |
|
38 |
+
def subset_file_list(all_files: List[str], subset_ids: Set[int]):
|
39 |
+
"""Subset files pertaining to individuals in person_ids arg."""
|
40 |
+
return [
|
41 |
+
file for file in all_files
|
42 |
+
if any(str(id_val) == file.split('_')[0] for id_val in subset_ids)
|
43 |
+
]
|
44 |
+
|
45 |
def standardize_3D_image(
|
46 |
image: np.ndarray,
|
47 |
resize_shape: Tuple[int, int, int]
|
|
|
137 |
description: Optional[str] = None,
|
138 |
scan_types: List[str] = DEFAULT_SCAN_TYPES,
|
139 |
resize_shape: Tuple[int, int, int] = DEFAULT_RESIZE,
|
140 |
+
shuffle: bool = True,
|
141 |
):
|
142 |
super().__init__(name, version, data_dir, data_files, description)
|
143 |
+
self.scan_types = self.validate_scan_types(scan_types)
|
144 |
self.resize_shape = resize_shape
|
145 |
+
self.shuffle = shuffle
|
146 |
+
|
147 |
+
def validate_scan_types(self, scan_types):
|
148 |
+
for item in scan_types:
|
149 |
+
if item not in ['t1', 't2', 't2_SPACE']:
|
150 |
+
raise ValueError(
|
151 |
+
'Scan type "{item}" not recognized as valid scan type.\
|
152 |
+
Verify scan type argument.'
|
153 |
+
)
|
154 |
+
return scan_types
|
155 |
|
156 |
|
157 |
class SPIDER(datasets.GeneratorBasedBuilder):
|
|
|
161 |
DEFAULT_WRITER_BATCH_SIZE = 16 # PyArrow default is too large for image data
|
162 |
VERSION = datasets.Version("1.1.0")
|
163 |
BUILDER_CONFIG_CLASS = CustomBuilderConfig
|
164 |
+
BUILDER_CONFIGS = [
|
165 |
+
CustomBuilderConfig(
|
166 |
+
name="default",
|
167 |
+
description="Load the full dataset",
|
168 |
+
),
|
169 |
+
CustomBuilderConfig(
|
170 |
+
name="demo",
|
171 |
+
description="Generate 10 examples for demonstration",
|
172 |
+
)
|
173 |
+
]
|
174 |
+
DEFAULT_CONFIG_NAME = "default"
|
175 |
+
|
176 |
+
# def __init__(
|
177 |
+
# self,
|
178 |
+
# *args,
|
179 |
+
# scan_types: List[str] = DEFAULT_SCAN_TYPES,
|
180 |
+
# resize_shape: Tuple[int, int, int] = DEFAULT_RESIZE,
|
181 |
+
# shuffle: bool = True,
|
182 |
+
# **kwargs,
|
183 |
+
# ):
|
184 |
+
# super().__init__(*args, **kwargs)
|
185 |
+
# self.scan_types = self.config.scan_types
|
186 |
+
# self.resize_shape = self.config.resize_shape
|
187 |
+
# self.shuffle = self.config.shuffle
|
188 |
|
189 |
def _info(self):
|
190 |
"""Specify datasets.DatasetInfo object containing information and typing
|
191 |
for the dataset."""
|
192 |
|
|
|
193 |
features = datasets.Features({
|
194 |
"patient_id": datasets.Value("string"),
|
195 |
"scan_type": datasets.Value("string"),
|
196 |
+
"image": datasets.Image(),
|
197 |
+
"mask": datasets.Image(),
|
198 |
+
# "image": datasets.Array3D(shape=self.config.resize_shape, dtype='uint8'),
|
199 |
+
# "mask": datasets.Array3D(shape=self.config.resize_shape, dtype='uint8'),
|
200 |
"image_path": datasets.Value("string"),
|
201 |
"mask_path": datasets.Value("string"),
|
|
|
|
|
202 |
"metadata": {
|
203 |
"num_vertebrae": datasets.Value(dtype="string"), #TODO: more specific types
|
204 |
"num_discs": datasets.Value(dtype="string"),
|
|
|
260 |
citation=_CITATION,
|
261 |
)
|
262 |
|
263 |
+
def _split_generators(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
264 |
self,
|
265 |
+
dl_manager,
|
|
|
|
|
|
|
266 |
validate_share: float = 0.2,
|
267 |
test_share: float = 0.2,
|
268 |
random_seed: int = 9999,
|
269 |
+
):
|
270 |
"""
|
271 |
+
Download and extract data and define splits based on configuration.
|
|
|
|
|
272 |
|
273 |
Args
|
274 |
+
dl_manager: HuggingFace datasets download manager (automatically supplied)
|
|
|
|
|
|
|
|
|
275 |
validate_share: float indicating share of data to use for validation;
|
276 |
must be in range (0.0, 1.0); note that training share is
|
277 |
calculated as (1 - validate_share - test_share)
|
278 |
test_share: float indicating share of data to use for testing;
|
279 |
must be in range (0.0, 1.0); note that training share is
|
280 |
calculated as (1 - validate_share - test_share)
|
281 |
+
random_seed: seed for random draws of train/validate/test patient ids
|
|
|
|
|
282 |
"""
|
283 |
# Set constants
|
284 |
train_share = (1.0 - validate_share - test_share)
|
285 |
np.random.seed(int(random_seed))
|
286 |
|
287 |
# Validate params
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
288 |
if train_share <= 0.0:
|
289 |
raise ValueError(
|
290 |
f'Training share is calculated as (1 - validate_share - test_share) \
|
|
|
303 |
{test_share}.'
|
304 |
)
|
305 |
|
306 |
+
# Download images (returns dictionary to local cache)
|
307 |
+
paths_dict = dl_manager.download_and_extract(_URLS)
|
308 |
+
|
309 |
+
# Get list of image and mask data files
|
310 |
+
image_files = [
|
311 |
+
file for file in os.listdir(os.path.join(paths_dict['images'], 'images'))
|
312 |
+
if file.endswith('.mha')
|
313 |
+
]
|
314 |
+
assert len(image_files) > 0, "No image files found--check directory path."
|
315 |
+
|
316 |
+
mask_files = [
|
317 |
+
file for file in os.listdir(os.path.join(paths_dict['masks'], 'masks'))
|
318 |
+
if file.endswith('.mha')
|
319 |
+
]
|
320 |
+
assert len(mask_files) > 0, "No mask files found--check directory path."
|
321 |
+
|
322 |
+
# Filter image and mask data files based on scan types
|
323 |
+
image_files = [
|
324 |
+
file for file in image_files
|
325 |
+
if any(scan_type in file for scan_type in self.config.scan_types)
|
326 |
+
]
|
327 |
+
|
328 |
+
mask_files = [
|
329 |
+
file for file in mask_files
|
330 |
+
if any(scan_type in file for scan_type in self.config.scan_types)
|
331 |
+
]
|
332 |
+
|
333 |
# Generate train/validate/test partitions of patient IDs
|
334 |
partition = np.random.choice(
|
335 |
['train', 'dev', 'test'],
|
|
|
342 |
test_ids = set(patient_ids[partition == 'test'])
|
343 |
assert len(train_ids.union(validate_ids, test_ids)) == N_PATIENTS
|
344 |
|
345 |
+
# Subset train/validation/test partition images and mask files
|
346 |
+
train_image_files = subset_file_list(image_files, train_ids)
|
347 |
+
validate_image_files = subset_file_list(image_files, validate_ids)
|
348 |
+
test_image_files = subset_file_list(image_files, test_ids)
|
349 |
+
|
350 |
+
train_mask_files = subset_file_list(mask_files, train_ids)
|
351 |
+
validate_mask_files = subset_file_list(mask_files, validate_ids)
|
352 |
+
test_mask_files = subset_file_list(mask_files, test_ids)
|
353 |
+
|
354 |
+
assert len(train_image_files) == len(train_mask_files)
|
355 |
+
assert len(validate_image_files) == len(validate_mask_files)
|
356 |
+
assert len(test_image_files) == len(test_mask_files)
|
357 |
+
|
358 |
# Import patient/scanner data and radiological gradings data
|
359 |
overview_data = import_csv_data(paths_dict['overview'])
|
360 |
grades_data = import_csv_data(paths_dict['gradings'])
|
|
|
405 |
if col not in ['Patient']
|
406 |
}
|
407 |
|
408 |
+
# DEMO configuration: subset first 10 examples
|
409 |
+
if self.config.name == "demo":
|
410 |
+
train_image_files = train_image_files[:10]
|
411 |
+
train_mask_files = train_mask_files[:10]
|
412 |
+
validate_image_files = validate_image_files[:10]
|
413 |
+
validate_mask_files = validate_mask_files[:10]
|
414 |
+
test_image_files = test_image_files[:10]
|
415 |
+
test_mask_files = test_mask_files[:10]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
416 |
|
417 |
+
return [
|
418 |
+
datasets.SplitGenerator(
|
419 |
+
name=datasets.Split.TRAIN,
|
420 |
+
gen_kwargs={
|
421 |
+
"paths_dict": paths_dict,
|
422 |
+
"image_files": train_image_files,
|
423 |
+
"mask_files": train_mask_files,
|
424 |
+
"overview_dict": overview_dict,
|
425 |
+
"grades_dict": grades_dict,
|
426 |
+
"resize_shape": self.config.resize_shape,
|
427 |
+
"shuffle": self.config.shuffle,
|
428 |
+
},
|
429 |
+
),
|
430 |
+
datasets.SplitGenerator(
|
431 |
+
name=datasets.Split.VALIDATION,
|
432 |
+
gen_kwargs={
|
433 |
+
"paths_dict": paths_dict,
|
434 |
+
"image_files": validate_image_files,
|
435 |
+
"mask_files": validate_mask_files,
|
436 |
+
"overview_dict": overview_dict,
|
437 |
+
"grades_dict": grades_dict,
|
438 |
+
"resize_shape": self.config.resize_shape,
|
439 |
+
"shuffle": self.config.shuffle,
|
440 |
+
},
|
441 |
+
),
|
442 |
+
datasets.SplitGenerator(
|
443 |
+
name=datasets.Split.TEST,
|
444 |
+
gen_kwargs={
|
445 |
+
"paths_dict": paths_dict,
|
446 |
+
"image_files": test_image_files,
|
447 |
+
"mask_files": test_mask_files,
|
448 |
+
"overview_dict": overview_dict,
|
449 |
+
"grades_dict": grades_dict,
|
450 |
+
"resize_shape": self.config.resize_shape,
|
451 |
+
"shuffle": self.config.shuffle,
|
452 |
+
},
|
453 |
+
),
|
454 |
]
|
455 |
+
|
456 |
+
def _generate_examples(
|
457 |
+
self,
|
458 |
+
paths_dict: Dict[str, str],
|
459 |
+
image_files: List[str],
|
460 |
+
mask_files: List[str],
|
461 |
+
overview_dict: Dict,
|
462 |
+
grades_dict: Dict,
|
463 |
+
resize_shape: Tuple[int, int, int],
|
464 |
+
shuffle: bool,
|
465 |
+
) -> Tuple[str, Dict]:
|
466 |
+
"""
|
467 |
+
This method handles input defined in _split_generators to yield
|
468 |
+
(key, example) tuples from the dataset. The `key` is for legacy reasons
|
469 |
+
(tfds) and is not important in itself, but must be unique for each example.
|
470 |
+
"""
|
471 |
# Shuffle order of patient scans
|
472 |
# (note that only images need to be shuffled since masks and metadata
|
473 |
# will be linked to the selected image)
|
474 |
+
if shuffle:
|
475 |
+
np.random.shuffle(image_files)
|
476 |
|
477 |
## Generate next example
|
478 |
# ----------------------
|
|
|
519 |
return_dict = {
|
520 |
'patient_id':patient_id,
|
521 |
'scan_type':scan_type,
|
522 |
+
'image':image_array_standardized,
|
523 |
+
'mask':mask_array_standardized,
|
524 |
'image_path':image_path,
|
525 |
'mask_path':mask_path,
|
|
|
|
|
526 |
'metadata':image_overview,
|
527 |
'rad_gradings':patient_grades_dict,
|
528 |
}
|