-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
210 lines (199 loc) · 9.47 KB
/
train.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
import torch.optim as optim
import torch.utils.data
import torch
from tqdm import tqdm, trange
import os
import numpy as np
from dataset.glas import get_glas_dataset
from dataset.MoNuBrain import get_monu_dataset
from segment_anything import sam_model_registry
from segment_anything.utils.transforms import ResizeLongestSide
import torch.nn.functional as F
from utils import *
def preprocess_logits(logits, size=256, th=0.01, is_resize=True):
logits = F.sigmoid(logits)
logits = (logits - logits.min()) / (logits.max() - logits.min())
if is_resize:
logits = F.interpolate(logits, (size, size), mode='bilinear', align_corners=True)
logits[logits>th] = 1
logits[logits<=th] = 0
return logits
def gen_step(optimizer, out_mask, logits, score, criterion):
loss = 0
dice_loss, _ = Dice_loss(logits, out_mask)
bce_loss = criterion(out_mask, logits)
loss = loss + 0.00*(1 - score) + dice_loss + bce_loss
loss.backward()
optimizer.step()
return loss.item(), dice_loss.item(), bce_loss.item()
def train_single_image(imgs, gts, original_sz, img_sz, sam, optimizer, args, file, ix):
loss_list = []
image = imgs.to(sam.device)
batched_input = get_input_dict(image, original_sz, img_sz)
pbar = trange(int(args['epoches']))
criterion = torch.nn.BCELoss()
gts_large = sam.postprocess_masks(gts.unsqueeze(dim=0),
input_size=img_sz[0],
original_size=original_sz.squeeze().long().tolist())
gts_large[gts_large>0.5] = 1
gts_large[gts_large<=0.5] = 0
with torch.no_grad():
zero_pred, p_i, n_i = get_similarity_maps(sam.eval(), imgs, gts.unsqueeze(dim=0).cuda(), th=0.5, pos=args['pos'], neg=args['neg'])
points_pred = zero_sam_with_points(sam, batched_input, p_i, n_i)
masks_pred = zero_sam_with_masks(sam, zero_pred, batched_input)
zero_pred = F.interpolate(zero_pred.unsqueeze(dim=0).unsqueeze(dim=0),
(256, 256),
mode='bilinear', align_corners=True).cuda()
zero_pred[zero_pred>0.5]=1
zero_pred[zero_pred<=0.5]=0
zero_pred_large = sam.postprocess_masks(zero_pred,
input_size=img_sz[0],
original_size=original_sz.squeeze().long().tolist())
zero_pred_large[zero_pred_large>0.5] = 1
zero_pred_large[zero_pred_large<=0.5] = 0
points_pred_large = sam.postprocess_masks(points_pred.unsqueeze(dim=0).unsqueeze(dim=0),
input_size=img_sz[0],
original_size=original_sz.squeeze().long().tolist())
points_pred_large[points_pred_large>0.5] = 1
points_pred_large[points_pred_large<=0.5] = 0
masks_pred_large = sam.postprocess_masks(masks_pred.unsqueeze(dim=0).unsqueeze(dim=0),
input_size=img_sz[0],
original_size=original_sz.squeeze().long().tolist())
masks_pred_large[masks_pred_large>0.5] = 1
masks_pred_large[masks_pred_large<=0.5] = 0
best = 0.0
res2 = 0.0
best_mask = None
best_i = None
for i in pbar:
optimizer.zero_grad()
out, score = get_sam_model_output(sam.train(), batched_input, zero_pred, it=args['it'])
loss, _, _ = gen_step(optimizer, out, zero_pred.detach(), score, criterion)
with torch.no_grad():
pred = out
pred[pred>0.5] = 1
pred[pred<=0.5] = 0
pred = sam.postprocess_masks(pred,
input_size=img_sz[0],
original_size=original_sz.squeeze().long().tolist())
dice_gt, ji_gt = get_dice_ji(pred.cpu().numpy(), gts_large.cpu().numpy())
_, ji = get_dice_ji(pred.cpu().numpy(), zero_pred_large.cpu().numpy())
if ji >= best:
best = ji
res2 = (dice_gt, ji_gt)
best_mask = pred.clone()
best_i = i
loss_list.append(loss)
pbar.set_description(
'(train | {}) epoch {epoch} ::'
' res {res:.4f}'.format(
'Medical',
epoch=i,
res=res2[1]
))
if best_mask is None:
best_mask = pred
best_i = args['epoches']
res1 = get_dice_ji(zero_pred_large.cpu().numpy(), gts_large.cpu().numpy())
file.write(str(ix) + ',' +
str(best_i) + ',' +
str(best)[:6] + ',' +
str(res1[1])[:6] + ',' +
str(res2[1])[:6] + '\n')
file.flush()
return res1, res2
def zero_sam(args=None, sam_args=None):
if torch.cuda.is_available():
device = torch.device("cuda")
else:
device = torch.device("cpu")
transform = ResizeLongestSide(1024)
if args['task'] == 'monu':
trainset, testset = get_monu_dataset(args, sam_trans=transform)
elif args['task'] == 'glas':
trainset, testset = get_glas_dataset(args, sam_trans=transform)
ds = torch.utils.data.DataLoader(testset, batch_size=1, shuffle=False, num_workers=int(args['nW_eval']), drop_last=False)
pbar = tqdm(ds)
file = open(os.path.join('vis', args['task'], args['vit'], 'results.csv'), 'w')
file.write('file,best_ix,iou_loss,iou_owl,iou_gt\n')
file.flush()
dice1_list, dice2_list, iou1_list, iou2_list = [], [], [], []
for ix, (imgs, gts, original_sz, img_sz) in enumerate(pbar):
sam = sam_model_registry[sam_args['model_type']](checkpoint=sam_args['sam_checkpoint'])
sam.to(device=device)
opt_model = sam.image_encoder #mask_decoder, prompt_encoder, image_encoder
for param in sam.parameters():
param.requires_grad = False
for param in opt_model.parameters():
param.requires_grad = True
optimizer = optim.Adam(opt_model.parameters(),
lr=float(args['learning_rate']),
weight_decay=float(args['WD']))
res1, res2 = train_single_image(imgs, gts, original_sz, img_sz, sam.train(), optimizer, args, file, ix)
dice1_list.append(res1[0])
dice2_list.append(res2[0])
iou1_list.append(res1[1])
iou2_list.append(res2[1])
pbar.set_description(
'(train | {}) epoch {epoch} ::'
'iou1: {iou1:.4f} :: iou2 {iou2:.4f}'.format(
'Medical',
epoch=ix,
iou1=np.mean(iou1_list),
iou2=np.mean(iou2_list)
))
return np.mean(dice1_list), np.mean(iou1_list), np.mean(dice2_list), np.mean(iou2_list)
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser(description='Description of your program')
parser.add_argument('-lr', '--learning_rate', default=1e-5, help='learning_rate', required=False)
parser.add_argument('-bs', '--Batch_size', default=1, help='batch_size', required=False)
parser.add_argument('-epoches', '--epoches', default=2, help='number of epoches', required=False)
parser.add_argument('-nW', '--nW', default=0, help='evaluation iteration', required=False)
parser.add_argument('-nW_eval', '--nW_eval', default=0, help='evaluation iteration', required=False)
parser.add_argument('-WD', '--WD', default=0, help='evaluation iteration', required=False)
parser.add_argument('-task', '--task', default='glas', help='evaluation iteration', required=False)
parser.add_argument('-rotate', '--rotate', default=22, help='image size', required=False)
parser.add_argument('-scale1', '--scale1', default=0.75, help='image size', required=False)
parser.add_argument('-scale2', '--scale2', default=1.25, help='image size', required=False)
parser.add_argument('-Idim', '--Idim', default=512, help='image size', required=False)
parser.add_argument('-it', '--it', default=1, help='image size', required=False)
parser.add_argument('-vit', '--vit', default='vit_b', help='image size', required=False)
parser.add_argument('-pos', '--pos', default=1, help='image size', required=False)
parser.add_argument('-neg', '--neg', default=1, help='image size', required=False)
args = vars(parser.parse_args())
os.makedirs('vis', exist_ok=True)
os.makedirs(os.path.join('vis', args['task'], args['vit']), exist_ok=True)
sam_args = {
'sam_checkpoint': "cp/sam_" + args['vit'] + ".pth",
'model_type': args['vit'],
'generator_args': {
'points_per_side': 8,
'pred_iou_thresh': 0.95,
'stability_score_thresh': 0.7,
'crop_n_layers': 0,
'crop_n_points_downscale_factor': 2,
'min_mask_region_area': 0,
'point_grids': None,
'box_nms_thresh': 0.7,
},
'gpu_id': 0,
}
name = args['task'] + '_' + args['vit'] + '_' +str(args['epoches'])
f = open('results_' + name + '.csv', 'w')
f.write('dataset,vit,pos,neg,dice1,iou1,dice2,iou2\n')
f.flush()
for points in [(1,1), (2,2), (3,3), (4,4), (5,5), (1,0), (2,0), (3,0), (4,0), (5,0)]:
args['pos'] = points[0]
args['neg'] = points[1]
dice1,iou1,dice2,iou2 = zero_sam(args=args, sam_args=sam_args)
f.write(args['task'] + ',' +
args['vit'] + ',' +
str(args['pos']) + ',' +
str(args['neg']) + ',' +
str(dice1) + ',' +
str(iou1) + ',' +
str(dice2) + ',' +
str(iou2) + '\n'
)
f.flush()