-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_MKTransformer.py
336 lines (281 loc) · 10.1 KB
/
train_MKTransformer.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
import argparse
import math
import random
import os
import numpy as np
from sklearn.metrics import roc_auc_score
import torch
from torch import nn, autograd, optim
from torch.nn import functional as F
from torch.utils import data
import torch.distributed as dist
from torchvision import transforms, utils
from tqdm import tqdm
import scipy.stats as stats
try:
import wandb
except ImportError:
wandb = None
from model import PyramidEncoder
import torchvision.models as models
from dataset import CXR14Dataset
import pdb
class Classifier(nn.Module):
def __init__(self, channel, out_dim, enc_state_dict=None, gray=False):
super().__init__()
self.encoder = PyramidEncoder(channel, gray=gray)
self.fc = nn.Linear(2048, out_dim)
# init encoder
if enc_state_dict is not None:
self.encoder.load_state_dict(enc_state_dict)
print("loaded")
# init fc
X = stats.truncnorm(-2, 2, scale=0.1)
values = torch.as_tensor(X.rvs(self.fc.weight.numel()), dtype=self.fc.weight.dtype)
values = values.view(self.fc.weight.size())
with torch.no_grad():
self.fc.weight.copy_(values)
def forward(self, data):
# with torch.no_grad():
# code, _ = self.encoder(data, run_tex=False, multi_str=False)
_, code = self.encoder(data, run_str=False, multi_tex=False)
# code = code.flatten(start_dim=1)
return self.fc(code)
def data_sampler(dataset, shuffle, distributed):
if distributed:
return data.distributed.DistributedSampler(dataset, shuffle=shuffle)
if shuffle:
return data.RandomSampler(dataset)
else:
return data.SequentialSampler(dataset)
def requires_grad(model, flag=True):
for p in model.parameters():
p.requires_grad = flag
def sample_data(loader):
while True:
for batch in loader:
yield batch
def bce_loss(pred, label):
loss = F.binary_cross_entropy_with_logits(pred, label, None,
pos_weight=None,
reduction='mean')
return loss
def adjust_learning_rate(optimizer, init_lr, iter, lr_steps):
nexp = 0
for step in lr_steps:
if iter < step:
break
else:
nexp += 1
lr = init_lr * math.pow(10, -nexp)
for param_group in optimizer.param_groups:
param_group['lr'] = lr
def compute_avg_auc(gt, pred, verbose=False):
AUROCs = []
if pred.ndim == 2:
n_classes = pred.shape[1]
elif pred.ndim == 1:
n_classes = 1
else:
raise ValueError("Prediction shape wrong")
for i in range(n_classes):
try:
auc = roc_auc_score(gt[:, i], pred[:, i])
except (IndexError, ValueError) as error:
if isinstance(error, IndexError):
auc = roc_auc_score(gt_np, pred_np)
elif isinstance(error, ValueError):
auc = 0
else:
raise Exception("Unexpected Error")
AUROCs.append(auc)
if verbose:
print(AUROCs)
AUROC_avg = np.array(AUROCs).mean()
return AUROC_avg
def train(
args,
tr_loader,
ts_loader,
classifier,
c_optim,
device,
):
tr_loader = sample_data(tr_loader)
pbar = range(args.iter)
pbar = tqdm(pbar, initial=args.start_iter, dynamic_ncols=True, smoothing=0.01)
# define the global variables
avg_auc = 0
loss_dict = {}
# create output folders
if args.proj_name != "":
ckpt_dir = f"outputs/ckpt-{args.proj_name}"
else:
ckpt_dir = "outputs/ckpt"
if not os.path.exists(ckpt_dir):
os.makedirs(ckpt_dir)
if args.distributed:
c_module = classifier.module
else:
c_module = classifier
for idx in pbar:
i = idx + args.start_iter
if i > args.iter:
print("Well!")
break
adjust_learning_rate(c_optim, args.lr, i, args.lr_steps)
real_img, label = next(tr_loader)
real_img = real_img.to(device)
label = label.to(device)
##### classifier optim #####
classifier.train()
requires_grad(classifier, True)
cls_logits = classifier(real_img)
# Minimize the loss of attr classification
cls_loss = bce_loss(cls_logits, label)
# attr classification loss
loss_dict["cls"] = cls_loss
# update discriminator
c_optim.zero_grad()
cls_loss.backward()
c_optim.step()
# loss_reduced = reduce_loss_dict(loss_dict)
cls_loss_val = loss_dict["cls"].mean().item()
if wandb and args.wandb and i % 10 == 0:
wandb.log(
{
"cls_loss": cls_loss_val,
"avg_auc": avg_auc,
},
step=i,
)
if i % 200 == 0:
with torch.no_grad():
scores_list = []
labels_list = []
for img, label in ts_loader:
img = img.to(device)
label = label.to(device)
classifier.eval()
cls_logits = classifier(img)
cls_scores = torch.sigmoid(cls_logits)
scores_list.append(cls_scores.cpu().data.numpy())
labels_list.append(label.cpu().data.numpy())
scores_arr = np.concatenate(scores_list, axis=0)
labels_arr = np.concatenate(labels_list, axis=0)
if i % 1000 == 0:
avg_auc = compute_avg_auc(labels_arr, scores_arr,verbose=True)
else:
avg_auc = compute_avg_auc(labels_arr, scores_arr)
# show information
pbar.set_description((f"c: {cls_loss_val:.4f}; avg_auc: {avg_auc:.4f}; lr: {c_optim.param_groups[0]['lr']:.4f}"))
if i % 5000 == 0:
torch.save(
{
"c": c_module.state_dict(),
"c_optim": c_optim.state_dict(),
"args": args,
"avg_auc": avg_auc,
"iter": i,
},
f"{ckpt_dir}/{str(i).zfill(6)}.pt",
)
if __name__ == "__main__":
device = "cuda"
torch.backends.cudnn.benchmark = True
parser = argparse.ArgumentParser()
parser.add_argument("--path", type=str, nargs="+")
parser.add_argument("--trlist", type=str)
parser.add_argument("--tslist", type=str)
parser.add_argument("--iter", type=int, default=800000)
parser.add_argument("--batch", type=int, default=16)
parser.add_argument("--size", type=int, default=256)
parser.add_argument("--r1", type=float, default=10)
parser.add_argument("--mlp_r1", type=float, default=1)
parser.add_argument("--ref_crop", type=int, default=4)
parser.add_argument("--n_crop", type=int, default=8)
parser.add_argument("--d_reg_every", type=int, default=16)
parser.add_argument("--ckpt", type=str, default=None)
parser.add_argument("--enc_ckpt", type=str, default=None)
parser.add_argument("--lr", type=float, default=0.002)
parser.add_argument("--lr_steps", type=int, nargs="+")
parser.add_argument("--channel", type=int, default=32)
parser.add_argument("--channel_multiplier", type=int, default=1)
parser.add_argument("--resume", action="store_true")
parser.add_argument("--wandb", action="store_true")
parser.add_argument("--local_rank", type=int, default=0)
parser.add_argument("--proj_name", type=str, default="SDoM_pneumoconiosis")
args = parser.parse_args()
n_gpu = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
args.distributed = n_gpu > 1
args.latent = 512
args.n_mlp = 8
args.start_iter = 0
try:
enc_ckpt = torch.load(args.enc_ckpt, map_location=lambda storage, loc: storage)
classifier = Classifier(args.channel, out_dim=14, enc_state_dict=enc_ckpt['e'], gray=True).to(device)
print("Encoder dictionary loaded")
except:
print("Rand Init")
classifier = Classifier(args.channel, out_dim=14, enc_state_dict=None, gray=True).to(device)
classifier = classifier.to(device)
print("Total model size: ", sum(p.numel() for p in classifier.parameters()))
c_optim = optim.Adam(
list(classifier.parameters()),
lr=args.lr,
betas=(0.9, 0.999)
)
if args.ckpt is not None:
print("load model:", args.ckpt)
ckpt = torch.load(args.ckpt, map_location=lambda storage, loc: storage)
if args.resume:
args.start_iter = ckpt["iter"]
classifier.load_state_dict(ckpt["c"])
c_optim.load_state_dict(ckpt["c_optim"])
else:
classifier.load_state_dict(ckpt["c"])
path = args.path[0]
# train dataset and dataloader
tr_transform = transforms.Compose(
[
transforms.RandomResizedCrop(256),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize((0.5,), (0.5,), inplace=True),
]
)
tr_dataset = CXR14Dataset(path, args.trlist, tr_transform, gray=True)
tr_loader = data.DataLoader(
tr_dataset,
batch_size=args.batch,
sampler=data_sampler(tr_dataset, shuffle=True, distributed=args.distributed),
num_workers=8,
drop_last=True,
)
# test dataset and dataloader
ts_transform = transforms.Compose(
[
transforms.Resize(256),
transforms.CenterCrop(256),
transforms.ToTensor(),
transforms.Normalize((0.5,), (0.5,), inplace=True),
]
)
ts_dataset = CXR14Dataset(path, args.tslist, ts_transform, gray=True)
ts_loader = data.DataLoader(
ts_dataset,
batch_size=args.batch,
sampler=data_sampler(ts_dataset, shuffle=False, distributed=args.distributed),
num_workers=8,
drop_last=False,
)
if wandb is not None and args.wandb:
wandb.init(project=f"proj_{args.proj_name}")
train(
args,
tr_loader,
ts_loader,
classifier,
c_optim,
device,
)