-
Notifications
You must be signed in to change notification settings - Fork 2
/
finaledgegcncopy.py
693 lines (617 loc) · 31.8 KB
/
finaledgegcncopy.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
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
import glob
import torch
from multiprocessing import Pool
import math
from torch.nn import Sequential, Linear, ReLU
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from torch_geometric.nn import DataParallel as GeoParallel
from torch_geometric.data import Data, Batch
from torchvision import datasets, transforms
import torchvision
import torch.utils.data as data
from torch.autograd import Variable
from skimage.color import rgb2hed
from skimage.util import img_as_float
from skimage.segmentation import slic, mark_boundaries
from skimage.exposure import rescale_intensity
from skimage import io
import cv2
import sys
import numpy as np
import random
from skimage import segmentation
import matplotlib.pyplot as plt
import torch.nn.init
from PIL import ImageFile, Image
import os
from collections import defaultdict
from utilities import *
from model import *
from args_parser import *
#amp_handle = amp.init(enabled=True)
global_segments = 0
if args.half_precision: from apex import amp
ImageFile.LOAD_TRUNCATED_IMAGES = True
##CUDA_LAUNCH_BLOCKING=1 #if cuda fails to backpropagate use this toggle to debug
OUTPUT_SIZE = args.output_size
use_cuda = torch.cuda.is_available()
SAVE_PATH = args.train_path
SD_SAVE_PATH = args.checkpoint_path
DATA_PATH = args.input_path
class random_dataloader(torch.utils.data.Dataset):
"""
Proposed Dataset with Random Matched Pairs
This function is a dataset designed for small batch initialization.
When initializing with small batches (< 8), less than ideal GCN convergence may occur.
This dataset loads image patches from two subfolder: ./positives ./negatives
-and combines positive and negative samples by adjustable ratio self.positive_negative_ratio
-in one mini-batch.
Keyword arguments:
path --parent path which contains ./positives ./negatives subfolders
"""
def __init__(self, path):
super(random_dataloader, self).__init__()
self.path = path
self.positive_images = [names for names in os.listdir(os.path.join(path, 'positives'))]
self.negative_images = [names for names in os.listdir(os.path.join(path, 'negatives'))]
self.positive_image_files = []
self.negative_image_files = []
self.toTensor = transforms.ToTensor()
self.positive_negative_ratio = 3 #3 positives and 1 negative in one mini-batch
self.ratio_counter = 0
self.negative_index = np.random.randint(low = 0, high = self.positive_negative_ratio + 1)
for img in self.positive_images:
if img[-4:] is not None and (img[-4:] == '.png' or img[-4:] == '.jpg'):
img_file = os.path.join(os.path.join(path, 'positives'), "%s" % img)
self.positive_image_files.append({
"img": img_file
})
for img in self.negative_images:
if img[-4:] is not None and (img[-4:] == '.png' or img[-4:] == '.jpg'):
img_file = os.path.join(os.path.join(path, 'negatives'), "%s" % img)
self.negative_image_files.append({
"img": img_file
})
def __len__(self):
return (len(self.positive_image_files) + len(self.negative_image_files))
def __getitem__(self, index):
if self.ratio_counter == self.negative_index:
index = index % len(self.negative_images)
data_file = self.negative_image_files[index]
else:
index = index % len(self.positive_images)
data_file = self.positive_image_files[index]
self.ratio_counter += 1
if self.ratio_counter > self.positive_negative_ratio:
self.negative_index = np.random.randint(low = 0, high = self.positive_negative_ratio + 1)
self.ratio_counter = 0
image = cv2.imread(data_file["img"])
image = cv2.resize(image, (OUTPUT_SIZE, OUTPUT_SIZE))
angle = np.random.randint(4)
image = rotate(image, angle)
image = cv2.flip(image, np.random.randint(2) - 1)
if args.color_channel_separation:
ihc_hed = rgb2hed(image)
h = rescale_intensity(ihc_hed[:, :, 0], out_range=(0, 1))
d = rescale_intensity(ihc_hed[:, :, 2], out_range=(0, 1))
image = np.dstack((np.zeros_like(h), d, h))
#image = zdh.transpose(2, 0, 1).astype('float32')/255
name = data_file["img"]
path, file = os.path.split(name)
split_filename = file.split("_")
gt_percent = float(split_filename[0])
moe = float(split_filename[1])
image = self.toTensor(image)
return (image, name, gt_percent, moe)
class normal_dataloader(torch.utils.data.Dataset):
"""
Typical Dataset
This function loads images from path and returns:
image: image (resized, color channel separated(optional))
name: full image path(contains image name)
gt_percent: Ground-Truth percent, an image-level weak annotation
moe: Margin of Error, an image-level weak annotation
Keyword arguments:
path --path which contains *.png or *.jpg image patches
"""
def __init__(self, path):
super(normal_dataloader, self).__init__()
self.path = path
self.images = [names for names in os.listdir(path)]
self.image_files = []
self.toTensor = transforms.ToTensor()
for img in self.images:
if img[-4:] is not None and (img[-4:] == '.png' or img[-4:] == '.jpg'):
img_file = os.path.join(path, "%s" % img)
self.image_files.append({
"img": img_file,
"label": "1"
})
def __len__(self):
return len(self.image_files)
def __getitem__(self, index):
index = index % len(self.image_files)
data_file = self.image_files[index]
image = cv2.imread(data_file["img"])
image = cv2.resize(image, (OUTPUT_SIZE, OUTPUT_SIZE))
angle = np.random.randint(4)
image = rotate(image, angle)
image = cv2.flip(image, np.random.randint(2) - 1)
if args.color_channel_separation:
ihc_hed = rgb2hed(image)
h = rescale_intensity(ihc_hed[:, :, 0], out_range=(0, 1))
d = rescale_intensity(ihc_hed[:, :, 2], out_range=(0, 1))
image = np.dstack((np.zeros_like(h), d, h))
#image = zdh.transpose(2, 0, 1).astype('float32')/255
name = data_file["img"]
path, file = os.path.split(name)
split_filename = file.split("_")
gt_percent = float(split_filename[0])
moe = float(split_filename[1])
image = self.toTensor(image)
return (image, name, gt_percent, moe)
class inference_dataset(torch.utils.data.Dataset):
"""
dataset loader used when inferencing
"""
def __init__(self, path):
super(inference_dataset, self).__init__()
self.path = path
self.images = [names for names in os.listdir(path)]
self.image_files = []
self.toTensor = transforms.ToTensor()
for img in self.images:
if img[-4:] is not None and (img[-4:] == '.jpg' or img[-4:] == '.png'):
img_file = os.path.join(path, "%s" % img)
#print(img)
#print(img_file)
self.image_files.append({
"img": img_file,
"label": "1"
})
def __len__(self):
return len(self.image_files)
def __getitem__(self, index):
index = index % len(self.image_files)
data_file = self.image_files[index]
image = cv2.imread(data_file["img"])
image = cv2.resize(image, (OUTPUT_SIZE, OUTPUT_SIZE))
name = data_file["img"]
image = self.toTensor(image)
return (image, name, 0,0)
def load_dataset(batch_size):
"""
Typical Pytorch Dataloader
This function loades image from Dataset
Keyword arguments:
path --please refer to class random_dataloader and class normal_dataloader
"""
data_path = DATA_PATH
train_dataset = normal_dataloader(path = data_path)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size= batch_size, # <8*1024 res @ P40 cards, <25*512 res @ 1 P40 card
num_workers= 8, #depends on RAM and CPU
shuffle=True
)
return train_loader
def inference_loader(batch_size = 1):
"""
Typical Pytorch Dataloader
This function loades image from Dataset
Keyword arguments:
path --please refer to class random_dataloader and class normal_dataloader
"""
data_path = DATA_PATH
train_dataset = inference_dataset(path = data_path)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size= batch_size, # <8*1024 res @ P40 cards, <25*512 res @ 1 P40 card
num_workers= 8, #depends on RAM and CPU
shuffle=False
)
return train_loader
def multithread_slic(multi_input):
"""
Multi-Thread SLIC superpixel
Since SLIC algorithm is implemented on CPU, we use Python Pool to accelerate
-SLIC algorithm by taking full advantage of CPU's multi-threading capability.
This function also calculate mutable edges' weight as one of GCN's input data.
Keyword arguments:
multi_input --tuple lists containing:
(FCN output, Max Channel Response, Mutable Edge Weight Ratio)
"""
(multi_output, max_channel_response, weight_ratio, i) = multi_input
if i > 0:
multi_output_slic = slic(multi_output, n_segments = i, compactness = args.compactness\
, sigma = 0, multichannel = True)
else:
multi_output_slic = slic(multi_output, n_segments = global_segments, compactness = args.compactness\
, sigma = 0, multichannel = True)
num_segments = len(np.unique(multi_output_slic))
multi_adj = adjacency3(multi_output_slic,num_segments)
## true_euclidean_distance = []
## f_norm = []
## mses = []
chisq = []
classes_raw = np.zeros((num_segments, args.nChannel),dtype="float32")
for y in range(OUTPUT_SIZE):
for x in range(OUTPUT_SIZE):
curr_index = multi_output_slic[x][y]
max_channel = max_channel_response[x][y]
classes_raw[curr_index][max_channel] += 1.0
for x in range(len(classes_raw)):
max_in_row = np.amax(classes_raw[x])
classes_raw[x] = classes_raw[x] / max_in_row
for (p1, p2) in multi_adj:
p1_class = np.asarray(classes_raw[p1])
p2_class = np.asarray(classes_raw[p2])
#true_euclidean_distance.append( euclidean_dist(p1center, p2center) )
#f_norm.append( fnorm(p1_class,p2_class) )
chisq.append( np.absolute(chisq_dist(p1_class, p2_class)) )
#mses.append( mse(p1_class, p2_class) )
chisq_max_value = np.amax(chisq)
chisq = chisq / chisq_max_value
#complementary_weight = np.ones_like(chisq) - chisq
#chisq = chisq + weight_ratio * complementary_weight
edge_weight = torch.from_numpy(chisq)
return multi_output_slic, multi_adj, edge_weight, num_segments
if __name__ == '__main__':
# train
model = FCN(3, args.nChannel)
modelgcn = GCN(args.nChannel, 1)
gcn_batch_iter = 1 #how many iteration on one GCN batch
batch_counter = 0
global_segments = args.num_superpixels #mutable superpixel quantity
change_dataloader = args.switch_iter #switch GCN into small batch training after # of iterations
model_loss = 99999 #variable keeping track of FCN loss during warmup phase
in_GCN = False #True when FCN exiting warmup phase and feeding output into GCN
inference_mode = False
#slic_multiscale_descending = False #True when mutable superpixel quantity is decreasing per iteration
slic_adjust_ratio = args.adjust_iter #each iteration, global_segments *= (1+/- slic_adjust_ratio)
weight_ratio = args.weight_ratio #edge weight complementing (to 1) ratio, decreasing with training
warmup_threshold = args.warmup_threshold# 0.5 #when FCN warmup loss reaches #, terminate warmup and start training GCN
half_precision = args.half_precision
if args.checkpoint > 0:
#if given a checkpoint to resume training
model.load_state_dict(torch.load(os.path.join(SD_SAVE_PATH, "FCN" + str(args.checkpoint) + ".pt")))
modelgcn.load_state_dict(torch.load(os.path.join(SD_SAVE_PATH, "GCN" + str(args.checkpoint) + ".pt")))
batch_counter = int(args.checkpoint)
change_dataloader = batch_counter - 1
in_GCN = True
if args.inference_path is not None:
dataset_loader = inference_loader()
inference_mode = True
args.maxIter = 1
if args.checkpoint < 1:
print("please define a inference checkpoint using --checkpoint")
exit(-1)
else:
dataset_loader = load_dataset(args.batch_size)
if not inference_mode and args.cpu_threads > args.batch_size:
args.cpu_threads = args.batch_size
if use_cuda:
model = model.to('cuda')
modelgcn = modelgcn.to('cuda')
if half_precision:
model, optimizer = amp.initialize(model, optimizer)
modelgcn, optimizergcn = amp.initialize(modelgcn, optimizergcn)
else:
print("model using CPU, please check CUDA settings")
if use_cuda:
if torch.cuda.device_count() > 1:
print(str(torch.cuda.device_count()) + " GPUs visible")
model = nn.DataParallel(model)
modelgcn = GeoParallel(modelgcn)
if inference_mode:
model.train()
else:
model.train()
modelgcn = modelgcn.float() #necessary for edge_weight initialized training
loss_fn = torch.nn.CrossEntropyLoss()
if args.optimizer == 'SGD':
optimizer = optim.SGD(model.parameters(), lr=args.fcn_lr, momentum=0)
optimizergcn = optim.SGD(modelgcn.parameters(), lr=args.gcn_lr, momentum=0)
elif args.optimizer == 'Adam':
optimizer = optim.Adam(model.parameters(), lr=args.fcn_lr)
optimizergcn = optim.Adam(modelgcn.parameters(), lr=args.gcn_lr)
else:
print("please reselect optimizer, curr value:", args.optimizer)
exit(-1)
if args.checkpoint > 0:
#if given a checkpoint to resume training
optimizer.load_state_dict(torch.load(os.path.join(SD_SAVE_PATH, "FCNopt" + str(args.checkpoint) + ".pt")))
optimizergcn.load_state_dict(torch.load(os.path.join(SD_SAVE_PATH, "GCNopt" + str(args.checkpoint) + ".pt")))
print("training successfully resumed!")
#an RGB colors array used to visualize channel response
label_colours = np.array([[0,0,0], [255,255,255],[0,0,255], [0,255,0],
[255,0,0],[128,0,0],[0,128,0],[0,0,128],
[255,255,0], [255,128,0], [128,255,0],
[0,255,255],[255,0,255],[255,255,255],
[128,128,128],[255,0,128],[0,128,255],
[128,0,255],[0,255,128],[100,200,200],
[200,100,100],[200,255,0],[100,255,0],
[200,0,255],[30,99,212],[40,222,100],
[100,200,25],[30,199,20],[0,211,200],
[3,44,122],[23,44,100],[90,22,0],[233,111,222],
[122,122,150],[0,233,149],[3,111,23]])
for epoch in range(args.maxIter):
"""switch large batch into small batch"""
# if batch_counter < change_dataloader:
# dataset_loader = load_dataset(2)
# else:
# gcn_batch_iter = 1
# dataset_loader = load_dataset(2)
#
for batch_idx, (data, name, gt_percent, moe) in \
enumerate(dataset_loader):
if not inference_mode:
print("iteration: " + str(batch_counter) + " epoch: " + str(epoch))
else:
print("---------------------------------------------------")
print("inferencing ", str(os.path.basename(name[0])))
if args.visualize:
"""visualize using opencv"""
originalimg = cv2.imread(name[0])
originalimg = cv2.resize(originalimg, (OUTPUT_SIZE,OUTPUT_SIZE))
cv2.imshow("original", originalimg)
cv2.waitKey(1)
batch_counter += 1 #records in-epoch progress
if use_cuda:
data = data.to('cuda')
optimizer.zero_grad()
output = model(data)
nLabels = -1
if not inference_mode and ((not in_GCN and batch_counter % 50 == 0) or (in_GCN and batch_counter % 10 == 0)):
"""FCN output visualization, either save to SAVE_PATH or display using opencv"""
#model.eval()
#output = model(data)
ignore, target = torch.max( output, 1 )
im_target = target.data.cpu().numpy() #label map original
num_in_minibatch = 0
for i in im_target:
im_target = i.flatten()
nLabels = len(np.unique(im_target))
label_num = rank(im_target)
label_rank = [i[0] for i in label_num]
im_target_rgb = np.array([label_colours [label_rank.index(c)] for c in im_target])
im_target_rgb = im_target_rgb.reshape( OUTPUT_SIZE, OUTPUT_SIZE, 3 ).astype( np.uint8 )
curr_filename = name[num_in_minibatch][-16:-4]
if args.visualize:
cv2.imshow("pre", im_target_rgb)
cv2.waitKey(1)
else:
cv2.imwrite(os.path.join(SAVE_PATH, 'PRE' + str(batch_counter) \
+ 'N' + str(num_in_minibatch) + "_" \
+ str(curr_filename) + '.png'), \
cv2.cvtColor(im_target_rgb, cv2.COLOR_RGB2BGR))
num_in_minibatch += 1
torch.save(model.module.state_dict(), os.path.join(SD_SAVE_PATH,"FCN" + str(batch_counter) + ".pt"))
if not inference_mode and (model_loss > warmup_threshold and not in_GCN): #stable 0.3 2000epoch
"""warning up FCN, loss is the cross entropy between pixel-wise
-max channel responses and FCN model's output"""
ignore, target = torch.max( output, 1 )
loss = loss_fn(output, target)
if half_precision:
with amp.scale_loss(loss, optimizer) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
optimizer.step()
if model_loss > loss.data:
model_loss = loss.data
print (epoch, '/', args.maxIter, ':', nLabels, loss.data)
change_dataloader += 1
continue
else:
"""when FCN model_loss is below warmup_threshold, enter GCN traning"""
#optimizer = torch.optim.Adam(model.parameters(), lr=0.0005)
#change the learning rate of FCN to a less conservative number
#optimizer = optim.Adam(model.parameters(), lr=0.00005)
#dataset_loader = load_dataset(2)
in_GCN = True
if model_loss < warmup_threshold or in_GCN:
"""Proposed method bridging output of FCN and input to GCN"""
"""mutable superpixel count, changes every training iterations"""
"""count changes by a fixed pattern"""
# if slic_multiscale_descending:
# global_segments = int(global_segments * (1 - slic_adjust_ratio))
# if global_segments < 2000:
# slic_adjust_ratio = slic_adjust_ratio * 0.7
# slic_multiscale_descending = False
# else:
# global_segments = int(global_segments * (1 + slic_adjust_ratio))
# if global_segments > 8000:
# global_segments = global_segments + 17
# slic_multiscale_descending = True
"""count is randomized between [2000, 8000)"""
global_segments = int(random.random() * 6000.0 + 2000.0)
if not inference_mode: print("slic segments count:" + str(global_segments))
gcn_batch_list = [] #a batch which later feed into GCN
segments_list = [] #SLIC segmentations of each image in current GCN batch
batch_node_num = [] #number of nodes(superpixel tiles) that each image have in\
#current GCN batch
print("computing multithread slic")
"""prepares FCN output for using in <def multithread_slic()>,
computes SLIC segmentations, adjacency edges, & edge weight using
multi-threads"""
multi_input = []
for multi_one_graph in output:
multi_one_graph = multi_one_graph.permute( 1 , 2 , 0 )
_, max_channel_response = torch.max(multi_one_graph, 2)
multi_one_graph = multi_one_graph.cpu().detach().numpy().astype(np.float64)
max_channel_response = max_channel_response.cpu().detach().numpy()
if not inference_mode:
multi_input.append((multi_one_graph, max_channel_response, weight_ratio, -1))
else:
for i in (2000,3000,4000,5000,6000,7000,8000):
multi_input.append((multi_one_graph, max_channel_response, weight_ratio, i))
with Pool(args.cpu_threads) as p:
multi_slic_adj_list = p.map(multithread_slic, multi_input)
print("multithread slic finished")
if weight_ratio > 0.2:
weight_ratio = weight_ratio * 0.99 #reduce edge weight's complementary
for one_graph, (segments, adj, edge_weight, num_segments) in zip(output, multi_slic_adj_list):
"""Bridging FCN's output into GCN's input, initialize GCN batch"""
segments_list.append(segments)
one_graph = one_graph.permute( 1 , 2 , 0 )
#one_graph.shape: [img_size, img_size, channel_size]
original_one_graph = torch.flatten(one_graph, start_dim = 0, end_dim = 1)
#original_one_graph.shape: [channel_size, img_size*img_size]
one_graph = None
batch_node_num.append(num_segments)
slic_pixels = [[] for _ in range (num_segments)]
"""slic_pixels stores x,y(flatten) corrdinates according to superpixel's index"""
for y in range (OUTPUT_SIZE):
for x in range (OUTPUT_SIZE):
curr_label = segments[x,y]
slic_pixels[curr_label].append(x * OUTPUT_SIZE + y)
#each slic seg's x y axis
classes = None
"""
For each superpixel's tile, select the PyTorch Variable(FCN) inside.
These Variables(FCN) combines into new Nodes Variable(GCN) while
-carrying gradients.
"""
for n in slic_pixels:
index_tensor = torch.LongTensor(n)
if use_cuda:
index_tensor = index_tensor.to('cuda')
one_class = torch.unsqueeze(torch.sum( \
torch.index_select(original_one_graph, dim = 0, index = index_tensor), \
0), dim = 0)
if classes is None:
classes = one_class
else:
classes = torch.cat((classes, one_class), 0)
one_class = None
index_tensor = None
original_one_graph = None
temp_ind = 0
adj = np.asarray(adj)
adj = torch.from_numpy(adj)
adj = adj.t().contiguous()
adj = Variable(adj).type(torch.LongTensor)
"""datagcn: GCN-ready wrapped data for one image
gcn_batch_list: GCN-ready minibatch containings same images from
-previous FCN's minibatch."""
datagcn = Data(x = classes, edge_index = adj, \
edge_weight = edge_weight.type(torch.FloatTensor))
gcn_batch_list.append(datagcn)
#print(gcn_batch_list)
classes = None #releases cached GPU memory immediately
adj = None
datagcn = None
if torch.cuda.device_count() == 1:
gcn_batch_list = Batch.from_data_list(gcn_batch_list)
if use_cuda:
gcn_batch_list = gcn_batch_list.to('cuda')
"""GCN training iterations"""
if inference_mode:
modelgcn.train()
else:
modelgcn.train()
print(gt_percent.data.cpu().numpy())
print(moe.data.cpu().numpy())
for epochgcn in range(0, gcn_batch_iter):
optimizergcn.zero_grad()
outgcn = modelgcn(gcn_batch_list)
"""visualize GCN output"""
if not inference_mode and epochgcn == gcn_batch_iter - 1 and batch_counter % 10 == 0:
start_index = 0
counter = 0
#modelgcn.eval()
#outgcn = modelgcn(gcn_batch_list)
for curr_batch_idx in range(len(name)):
outgcn_slice = torch.narrow(input = outgcn, dim = 0, \
start = start_index, \
length = batch_node_num[curr_batch_idx])
start_index += batch_node_num[curr_batch_idx]
outputgcn_np = outgcn_slice.detach().data.cpu().numpy()
segments_copy = segments_list[curr_batch_idx].copy()
segments_copy = segments_copy.astype(np.float64)
for segInd in range(len(outputgcn_np)):
segments_copy[segments_copy == segInd] = outputgcn_np[segInd]
gcn_target_rgb = np.array([[255*(c + 1) / 2, 255*(c + 1) / 2, 255*(c + 1) / 2] \
for c in segments_copy])
gcn_target_rgb = np.moveaxis(gcn_target_rgb, 1, 2)
gcn_target_rgb = gcn_target_rgb.reshape( (OUTPUT_SIZE,OUTPUT_SIZE,3) ).astype( np.uint8 )
if args.visualize:
cv2.imshow("gcn", gcn_target_rgb)
cv2.waitKey(1)
else:
cv2.imwrite(os.path.join(SAVE_PATH, 'GCN' + str(batch_counter) + "N" + str(counter)\
+ "_" + str(name[curr_batch_idx][-16:-4]) + '.png'), \
cv2.cvtColor(gcn_target_rgb, cv2.COLOR_RGB2BGR))
counter += 1
outgcn_slice = None
if not inference_mode:
"""
loss_top --cost, positive <gt_percent-moe> % of nodes responded correctly
loss_bottom --cost, negative <1-gt_percent-moe> % of nodes responded correctly
positive refers to desired region(cancer), negative refers to other regions(background)
"""
loss_top, loss_bottom = one_label_loss(gt_percent = gt_percent.data.cpu().numpy(), \
predict = outgcn, \
moe = moe.data.cpu().numpy(), \
batch_node_num = batch_node_num)
if loss_top is None:
total_gcn_loss = loss_bottom
elif loss_bottom is None:
total_gcn_loss = loss_top
else:
total_gcn_loss = loss_top + loss_bottom
if half_precision:
with amp.scale_loss(total_gcn_loss, optimizergcn) as scaled_loss2:
scaled_loss2.backward(retain_graph=True)
else:
total_gcn_loss.backward(retain_graph=True)
#backward calculating GCN gradients according to combined loss
#print(total_gcn_loss)
print("GCN+FCN loss: " + str(total_gcn_loss.data.cpu().numpy()))
if not inference_mode:
#backpropagate through GCN's layers
optimizergcn.step()
#backpropagate accumulated gradients through FCN's layers
optimizer.step()
#saving models & optimizers state_dict for later training and inference
#cpu = torch.device("cpu")
#model = model.to(cpu)
#modelgcn = modelgcn.to(cpu)
torch.save(model.module.state_dict(), os.path.join(SD_SAVE_PATH,"FCN" + str(batch_counter) + ".pt"))
torch.save(modelgcn.module.state_dict(), os.path.join(SD_SAVE_PATH,"GCN" + str(batch_counter) + ".pt"))
torch.save(optimizer.state_dict(), os.path.join(SD_SAVE_PATH,"FCNopt" + str(batch_counter) + ".pt"))
torch.save(optimizergcn.state_dict(), os.path.join(SD_SAVE_PATH,"GCNopt" + str(batch_counter) + ".pt"))
#model = model.to('cuda')
#modelgcn = modelgcn.to('cuda')
if inference_mode:
start_index = 0
counter = 0
final_map = np.zeros((args.output_size, args.output_size))
multi_input = []
print("fusing")
for input_index in range(len(segments_list)):
outgcn_numpy = torch.narrow(input = outgcn, dim = 0, start = start_index, length = batch_node_num[input_index]).detach().data.cpu().numpy()
segments_copy = segments_list[input_index].astype(np.float64)
multi_input.append((outgcn_numpy, segments_copy))
start_index += batch_node_num[input_index]
with Pool(args.cpu_threads) as p:
multi_graph = p.map(fuse_results, multi_input)
for graph in multi_graph:
final_map += graph
final_map = final_map / float(len(multi_graph))
final_map += 1.0
final_map = final_map / 2.0
final_map[final_map < args.fuse_thresh] = 0
gcn_target_rgb = np.array([[255 * c , 255* c , 255* c] for c in final_map])
gcn_target_rgb = np.moveaxis(gcn_target_rgb, 1, 2)
gcn_target_rgb = gcn_target_rgb.reshape( (args.output_size,args.output_size,3) ).astype( np.uint8 )
if args.visualize:
cv2.imshow("gcn", gcn_target_rgb)
cv2.waitKey(1)
else:
basename = os.path.basename(name[0])
cv2.imwrite(os.path.join(args.inference_path, basename), cv2.cvtColor(gcn_target_rgb, cv2.COLOR_RGB2BGR))
print("inference for", str(basename), "saved to", str(args.inference_path))