-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval.py
532 lines (490 loc) · 21.8 KB
/
eval.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
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
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
"""
Run predictions over a list of images
- Visualize predictions
- Calculate the metrics if ground truth is available
"""
import argparse
import csv
import json
import os
from collections import defaultdict
import cv2
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from tqdm import tqdm
from config import ConfigParams
from data.dataset import BloodVesselDatasetTest
from data.rle import rle_encode
from data.transforms import get_test_transform
from metrics.fast_surface_dice.fast_surface_dice import compute_surface_dice_score
from metrics.metrics import DiceScore, Metric
from model import init_model
from tta import predict_crops_tta_max, predict_no_tta
from utils import get_device
from visualization.image import convert_to_image
def do_parsing():
parser = argparse.ArgumentParser(
formatter_class=argparse.ArgumentDefaultsHelpFormatter,
description="Eval the model",
)
parser.add_argument(
"--config_path", required=True, type=str, help="Configuration filepath"
)
parser.add_argument(
"--batch_size",
required=False,
type=int,
help="Evaluation batch size, otherwise training one is used",
)
parser.add_argument(
"--inference_input_size",
required=False,
type=int,
nargs="*",
help="Evaluation input size width and height, if only one value is passed it will be used both for "
"width and height. If no value is passed the training one is used",
)
parser.add_argument(
"--threshold",
required=False,
type=float,
help="Probability threshold to set 1 as prediction",
)
parser.add_argument(
"--model_path", required=True, type=str, help="Model checkpoint filepath"
)
parser.add_argument(
"--rescale",
action="store_true",
help="Rescale to the original image size (otherwise keep the model input size) "
"for prediction visualization as images",
)
parser.add_argument(
"--tta_mode",
choices=["4+fullmax", "5+fullmax", "4max", "5max", "None", "none"],
required=False,
default=None,
help="Test time augmentation mode, don't pass it to don't use TTA",
)
parser.add_argument(
"--input_paths",
required=True,
type=str,
nargs="*",
help="List of input directories, each one should correspond to a kidney and "
"must contain images dir and optionally labels dir",
)
parser.add_argument(
"--output_dir",
required=True,
type=str,
help="Output directory with predictions and metrics",
)
parser.add_argument(
"--save_3d_predictions",
action="store_true",
help="Save 3d prediction for each dataset subset kidney",
)
parser.add_argument(
"--save_3d_labels",
action="store_true",
help="Save 3d label for each dataset subset kidney",
)
args = parser.parse_args()
return args
def main():
args = do_parsing()
print(args)
config = ConfigParams(args.config_path)
batch_size = (
args.batch_size if args.batch_size is not None else config.train_batch_size
)
threshold = (
args.threshold if args.threshold is not None else config.inference_threshold
)
assert threshold is not None
tta_mode = args.tta_mode if args.tta_mode is not None else config.tta_mode
if tta_mode in ["", "None", "none"]:
tta_mode = None
device = get_device()
model, preprocess_function, inverse_preprocess_function = init_model(config)
model.load_state_dict(torch.load(args.model_path))
model.eval()
model.to(device)
inference_input_size = (
args.inference_input_size
if args.inference_input_size
else [config.model_train_input_size]
)
if len(inference_input_size) == 2:
inference_input_width = inference_input_size[0]
inference_input_height = inference_input_size[1]
else:
inference_input_width = inference_input_size[0]
inference_input_height = inference_input_size[0]
data_transform_test = get_test_transform(
input_size_height=inference_input_height, input_size_width=inference_input_width
)
labels_exists = [
os.path.exists(os.path.join(input_path, "labels"))
for input_path in args.input_paths
]
labels_exists_check = np.all(labels_exists)
if labels_exists_check is False:
print(f"Labels are not available for at least one of {args.input_paths}")
return
print(
f"Preparing dataset for inference size w x h: {inference_input_width} x {inference_input_height}"
)
test_dataset = BloodVesselDatasetTest(
input_size_width=inference_input_width,
input_size_height=inference_input_height,
selected_dirs=args.input_paths,
transform=data_transform_test,
preprocess_function=preprocess_function,
dataset_with_gt=labels_exists_check,
in_channels=config.model_input_channels,
tta_mode=tta_mode,
)
test_dataloader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)
test_batches = len(test_dataset) // config.train_batch_size + int(
len(test_dataset) % batch_size > 0
)
kidney_subset_names = [
os.path.basename(input_path) for input_path in args.input_paths
]
for kidney_subset_name in kidney_subset_names:
os.makedirs(
os.path.join(args.output_dir, f"images_{kidney_subset_name}"),
exist_ok=False,
)
os.makedirs(os.path.join(args.output_dir, "3d"), exist_ok=False)
dice_score_class: Metric = DiceScore(to_monitor=False)
dice_2d_scores = defaultdict(list)
single_2d_slices_scores = list()
# Keep track of all the predictions for 3D evaluation
predictions_for_3d = defaultdict(list)
labels_for_3d = defaultdict(list)
predictions_df_dict = dict()
labels_df_dict = dict()
# Iterate on test batches
with torch.no_grad():
for batch_id, data in tqdm(
enumerate(test_dataloader), desc="batch", total=test_batches
):
# get the input images and labels
images = data["image"]
labels_model_size_nchw = (
data["label_model_size"] if "label_model_size" in data else None
)
labels_full_size_nhw = (
data["label_full_size"] if "label_full_size" in data else None
)
images_paths = data["file"]
images_shapes = data["shape"]
# Move to GPU
images = images.to(device)
if labels_full_size_nhw is not None:
labels_full_size_nhw = labels_full_size_nhw.to(device)
# forward pass to get outputs
predictions_raw = nn.Sigmoid()(model(images))
# Prepare results for TTA
if tta_mode:
tl_predictions_raw = nn.Sigmoid()(model(data["top_left"].to(device)))
tr_predictions_raw = nn.Sigmoid()(model(data["top_right"].to(device)))
bl_predictions_raw = nn.Sigmoid()(model(data["bottom_left"].to(device)))
br_predictions_raw = nn.Sigmoid()(
model(data["bottom_right"].to(device))
)
center_predictions_raw = nn.Sigmoid()(model(data["center"].to(device)))
for i in range(images.shape[0]):
image_path = images_paths[i]
dataset_kidney_name = image_path.split(os.sep)[-3]
# Shape format is a list of 3 elements (H-W-C).
# Each element is a list of batch_size length with the values
original_image_height, original_image_width = (
images_shapes[0][i].item(),
images_shapes[1][i].item(),
)
original_shape_wh = (original_image_width, original_image_height)
resize_to_wh = original_shape_wh if args.rescale else None
# Restore original image values range
# Select the first slice and then readd the dimension
original_image = inverse_preprocess_function(
x_norm=torch.unsqueeze(images[i][0], dim=0)
)
full_image = convert_to_image(original_image, resize_to_wh)
label_img_model_size = (
convert_to_image(labels_model_size_nchw[i], resize_to_wh)
if labels_model_size_nchw is not None
else None
)
# Calculate prediction and labels upscaled for 3D metrics and 3D results export (only 0-1 values)
# the fast_surface_dice implementation requires two dataframes as input: prediction and label
if tta_mode:
prediction_raw, prediction_thresholded = predict_crops_tta_max(
full_image_prediction_raw=predictions_raw[i],
tl_prediction_raw=tl_predictions_raw[i],
tr_prediction_raw=tr_predictions_raw[i],
bl_prediction_raw=bl_predictions_raw[i],
br_prediction_raw=br_predictions_raw[i],
center_prediction_raw=center_predictions_raw[i],
threshold=threshold,
tta_mode=tta_mode,
)
else:
prediction_raw, prediction_thresholded = predict_no_tta(
prediction_raw=predictions_raw[i], threshold=threshold
)
# print(f"Prediction thresholded shape: {prediction_thresholded.shape}")
# print(f"Original shape w x h: {original_image_width} x {original_image_height}")
# Force resize to model input for visualization, it could be bigger with TTA
# TODO: Find a better solution for this resize
prediction_raw_img = convert_to_image(
prediction_raw,
(inference_input_width, inference_input_height)
if args.rescale is False
else resize_to_wh,
)
prediction_thresholded_img = convert_to_image(
prediction_thresholded,
(inference_input_width, inference_input_height)
if args.rescale is False
else resize_to_wh,
)
# Manipulate prediction with torch, 4D input is necessary for upsampling,
# upsampling is made at the original image size
prediction_thresholded_4d = torch.unsqueeze(
prediction_thresholded, dim=0
) # 1 x 1 x height x width shape
prediction_upscaled_th = nn.UpsamplingNearest2d(
size=[original_image_height, original_image_width]
)(prediction_thresholded_4d)
# Back to 3 dimensions CHW
prediction_upscaled_th = torch.squeeze(prediction_upscaled_th)
# Numpy shape HW for rle encode
prediction_upscaled_npy = (
prediction_upscaled_th.cpu().data.detach().numpy(force=True)
)
prediction_rle = rle_encode(np.squeeze(prediction_upscaled_npy))
prediction_row_df = pd.DataFrame.from_dict(
{
"id": [
f"{dataset_kidney_name}_{os.path.basename(image_path[:-4])}"
],
"rle": [prediction_rle],
"width": [original_image_width], # Not mandatory
"height": [original_image_height], # Not mandatory
"image_id": [dataset_kidney_name], # Not mandatory
"slice_id": [
os.path.basename(image_path[:-4]).split("_")[-1]
], # Not mandatory
}
)
if dataset_kidney_name not in predictions_df_dict:
predictions_df_dict[dataset_kidney_name] = pd.DataFrame(
columns=["id", "rle", "width", "height", "image_id", "slice_id"]
)
predictions_df_dict[dataset_kidney_name] = pd.concat(
[predictions_df_dict[dataset_kidney_name], prediction_row_df],
ignore_index=True,
)
predictions_for_3d[dataset_kidney_name].append(
prediction_upscaled_npy.astype(bool)
)
if labels_model_size_nchw is None:
all_in_one = cv2.hconcat(
[full_image, prediction_raw_img, prediction_thresholded_img]
)
cv2.imwrite(
os.path.join(
args.output_dir,
f"images_{dataset_kidney_name}",
os.path.basename(image_path),
),
all_in_one,
)
else:
# HW label at full resolution
label_full_size_npy = labels_full_size_nhw[i].cpu().data.numpy()
# WH for rle encode
label_rle = rle_encode(np.squeeze(label_full_size_npy))
label_row_df = pd.DataFrame.from_dict(
{
"id": [
f"{dataset_kidney_name}_{os.path.basename(image_path[:-4])}"
],
"rle": [label_rle],
"width": [original_image_width],
"height": [original_image_height],
"image_id": [dataset_kidney_name],
"slice_id": [
os.path.basename(image_path[:-4]).split("_")[-1]
],
}
)
if dataset_kidney_name not in labels_df_dict:
labels_df_dict[dataset_kidney_name] = pd.DataFrame(
columns=[
"id",
"rle",
"width",
"height",
"image_id",
"slice_id",
]
)
labels_df_dict[dataset_kidney_name] = pd.concat(
[labels_df_dict[dataset_kidney_name], label_row_df],
ignore_index=True,
)
# Save label full for dice score (HW)
labels_for_3d[dataset_kidney_name].append(
label_full_size_npy.astype(bool)
)
# TODO: Extract function(s)
diff_on_image = np.copy(full_image)
diff_on_image = cv2.cvtColor(diff_on_image, cv2.COLOR_GRAY2BGR)
diff_on_black = np.zeros_like(label_img_model_size, np.uint8)
diff_on_black = cv2.cvtColor(diff_on_black, cv2.COLOR_GRAY2BGR)
# Green
correct = (
label_img_model_size & prediction_thresholded_img
) # np.uint8 of 0 or 1
correct_bool = np.squeeze(correct.astype(bool))
diff_on_image[correct_bool] = [0, 255, 0]
diff_on_black[correct_bool] = [0, 255, 0]
# Red
false_positive = (
label_img_model_size == 0
) & prediction_thresholded_img
false_positive_bool = np.squeeze(false_positive.astype(bool))
diff_on_image[false_positive_bool] = [0, 0, 255]
diff_on_black[false_positive_bool] = [0, 0, 255]
# Blue
false_negative = label_img_model_size & (
prediction_thresholded_img == 0
)
false_negative_bool = np.squeeze(false_negative.astype(bool))
diff_on_image[false_negative_bool] = [255, 0, 0]
diff_on_black[false_negative_bool] = [255, 0, 0]
image_bgr = cv2.cvtColor(full_image, cv2.COLOR_GRAY2BGR)
label_bgr = cv2.cvtColor(label_img_model_size, cv2.COLOR_GRAY2BGR)
predictions_raw_img_bgr = cv2.cvtColor(
prediction_raw_img, cv2.COLOR_GRAY2BGR
)
predictions_thresholded_img_bgr = cv2.cvtColor(
prediction_thresholded_img, cv2.COLOR_GRAY2BGR
)
row_one = cv2.hconcat([image_bgr, label_bgr, diff_on_image])
row_two = cv2.hconcat(
[
predictions_raw_img_bgr,
predictions_thresholded_img_bgr,
diff_on_black,
]
)
all_in_one = cv2.vconcat([row_one, row_two])
dice_score = dice_score_class.evaluate(
prediction_upscaled_th, labels_full_size_nhw[i]
)
print(f"{image_path} {dice_score_class.name} {dice_score:.2f}")
dice_2d_scores[dataset_kidney_name].append(dice_score)
single_2d_slices_scores.append(
{
"image_path": image_path,
dice_score_class.name: f"{dice_score:.2f}",
}
)
cv2.imwrite(
os.path.join(
args.output_dir,
f"images_{dataset_kidney_name}",
os.path.basename(image_path)[:-4]
+ f"_{dice_score_class.name}_{dice_score:.2f}.png",
),
all_in_one,
)
if labels_exists_check:
metrics = defaultdict(dict)
# 2D metrics: average 2D dice score (on each 2D slice)
for dataset_kidney_name, scores in dice_2d_scores.items():
avg_2d_dice_score = np.mean(scores)
print(f"{dataset_kidney_name} avg 2D dice score: {avg_2d_dice_score:.2f}")
metrics[dataset_kidney_name][
"avg_2D_dice_score"
] = f"{avg_2d_dice_score:.3f}"
print(f"{dataset_kidney_name}: calculating surface dice score...")
surface_dice_score = compute_surface_dice_score(
predictions_df_dict[dataset_kidney_name],
labels_df_dict[dataset_kidney_name],
device,
)
print(f"Surface dice score: {surface_dice_score}")
metrics[dataset_kidney_name]["surface_dice_score"] = f"{surface_dice_score:.3f}"
# Save also the dataframes
predictions_df_dict[dataset_kidney_name].to_csv(
os.path.join(
args.output_dir, f"slices_predictions_rle_{dataset_kidney_name}.csv"
),
index=False,
sep=";",
)
labels_df_dict[dataset_kidney_name].to_csv(
os.path.join(
args.output_dir, f"slices_labels_rle_{dataset_kidney_name}.csv"
),
index=False,
sep=";",
)
with open(os.path.join(args.output_dir, "metrics.json"), "w") as out_fp:
json.dump(metrics, out_fp, indent=4)
print(f"Metrics written to {os.path.join(args.output_dir, 'metrics.json')}")
with open(
os.path.join(args.output_dir, "single_2d_slices_scores.csv"), "w"
) as out_fp:
fieldnames = ["image_path", dice_score_class.name]
writer = csv.DictWriter(out_fp, fieldnames=fieldnames)
writer.writeheader()
for single_2d_slice_scores_row in single_2d_slices_scores:
writer.writerow(single_2d_slice_scores_row)
print(
f"2D slices scores written to {os.path.join(args.output_dir, 'single_2d_slices_scores.csv')}"
)
# 3D metrics: surface dice and save 3D numpy array for later analysis
for dataset_kidney_name in predictions_for_3d.keys():
print(f"{dataset_kidney_name}: saving 3d numpy prediction and label...")
if args.save_3d_predictions:
kidney_predictions_for_3d = predictions_for_3d[dataset_kidney_name]
kidney_prediction_3d_zyx = np.stack(
kidney_predictions_for_3d, axis=0
) # CHW (slices x height x width)
kidney_prediction_3d_xyz = np.transpose(kidney_prediction_3d_zyx, (2, 1, 0))
np.save(
os.path.join(
args.output_dir, "3d", f"{dataset_kidney_name}_prediction_xyz.npy"
),
kidney_prediction_3d_xyz,
)
if args.save_3d_labels:
if labels_exists_check:
kidney_labels_for_3d = labels_for_3d[dataset_kidney_name]
kidney_label_3d_zyx = np.stack(
kidney_labels_for_3d, axis=0
) # CHW (slices x height x width)
kidney_label_3d_xyz = np.transpose(kidney_label_3d_zyx, (2, 1, 0))
np.save(
os.path.join(
args.output_dir, "3d", f"{dataset_kidney_name}_label_xyz.npy"
),
kidney_label_3d_xyz,
)
else:
print(
"Labels not available for all the input directories, saving 3D labels will be skipped"
)
if __name__ == "__main__":
main()