-
Notifications
You must be signed in to change notification settings - Fork 5
/
train_lt_mxp.py
381 lines (317 loc) · 18.8 KB
/
train_lt_mxp.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
import sys, json, os, argparse, time
from operator import itemgetter
import shutil
import os.path as osp
from datetime import datetime
import operator
from tqdm import trange
import numpy as np
import torch
import torch.nn.functional as F
from models.get_model import get_arch
from utils.get_loaders import get_train_val_cls_loaders, modify_dataset, modify_loader, get_combo_loader
from utils.evaluation import evaluate_multi_cls
from utils.model_saving_loading import save_model, str2bool, load_model
from utils.reproducibility import set_seeds
from torch.optim.lr_scheduler import ReduceLROnPlateau
from typing import Tuple
# argument parsing
parser = argparse.ArgumentParser()
# as seen here: https://stackoverflow.com/a/15460288/3208255
# parser.add_argument('--layers', nargs='+', type=int, help='unet configuration (depth/filters)')
# annoyingly, this does not get on well with guild.ai, so we need to reverse to this one:
parser.add_argument('--csv_train', type=str, default='data/train_eyepacs.csv', help='path to training data csv')
parser.add_argument('--data_path', type=str, default='data/eyepacs_all_ims/', help='path data')
parser.add_argument('--sampling', type=str, default='instance', help='sampling mode (instance, class, sqrt, prog)')
parser.add_argument('--model_name', type=str, default='bit_resnext50_1', help='architecture')
parser.add_argument('--n_classes', type=int, default=5, help='binary disease detection (1) or multi-class (5)')
parser.add_argument('--loss_fn', type=str, default='ce', help='loss function (ce)')
parser.add_argument('--do_mixup', type=float, default=0.0, help='mixup coeff (so far only for multi-class')
parser.add_argument('--batch_size', type=int, default=8, help='batch size')
parser.add_argument('--optimizer', type=str, default='sgd', help='optimizer choice')
parser.add_argument('--patience', type=int, default=2, help='patience before lr reduction')
parser.add_argument('--lr', type=float, default=0.01, help='learning rate')
parser.add_argument('--min_lr', type=float, default=-1, help='learning rate (defaults to stopping just after 3rd decay)')
parser.add_argument('--wd', type=float, default=0, help='weight decay')
parser.add_argument('--n_epochs', type=int, default=7, help='nr epochs') #
parser.add_argument('--metric', type=str, default='auc', help='which metric to use for monitoring progress (loss/dice)')
parser.add_argument('--im_size', help='delimited list input, could be 500, or 600,400', type=str, default='512,512')
parser.add_argument('--pretrained_weights', type=str, default=None, help='start from eyepacs-pretrained weights (path to)')
parser.add_argument('--do_not_save', type=str2bool, nargs='?', const=True, default=False, help='avoid saving anything')
parser.add_argument('--save_path', type=str, default='date_time', help='path to save model (defaults to date/time')
parser.add_argument('--num_workers', type=int, default=8, help='number of parallel (multiprocessing) workers to launch '
'for data loading tasks (handled by pytorch) [default: %(default)s]')
parser.add_argument('--n_checkpoints', type=int, default=1, help='nr of best checkpoints to keep (defaults to 3)')
def sgd_optimizer(model, lr, momentum, weight_decay):
params = []
for key, value in model.named_parameters():
if not value.requires_grad:
continue
apply_weight_decay = weight_decay
apply_lr = lr
if 'bias' in key or 'bn' in key:
apply_weight_decay = 0
# print('set weight decay=0 for {}'.format(key))
if 'bias' in key:
apply_lr = 2 * lr # Just a Caffe-style common practice. Made no difference.
params += [{'params': [value], 'lr': apply_lr, 'weight_decay': apply_weight_decay}]
optimizer = torch.optim.SGD(params, lr, momentum=momentum)
return optimizer
def compare_op(metric):
'''
This should return an operator that given a, b returns True if a is better than b
Also, let us return which is an appropriately terrible initial value for such metric
'''
if metric == 'auc':
return operator.gt, 0
elif metric == 'mcc':
return operator.gt, 0
elif metric == 'kappa':
return operator.gt, 0
elif metric == 'f1':
return operator.gt, 0
elif metric == 'bacc':
return operator.gt, 0
elif metric == 'loss':
return operator.lt, np.inf
else:
raise NotImplementedError
def get_lr(optimizer):
for param_group in optimizer.param_groups:
return param_group['lr']
def partial_mixup(input: torch.Tensor,
gamma: float,
indices: torch.Tensor
) -> torch.Tensor:
if input.size(0) != indices.size(0):
raise RuntimeError("Size mismatch!")
perm_input = input[indices]
return input.mul(gamma).add(perm_input, alpha=1 - gamma)
def mixup(input: torch.Tensor,
target: torch.Tensor,
gamma: float,
) -> Tuple[torch.Tensor, torch.Tensor]:
indices = torch.randperm(input.size(0), device=input.device, dtype=torch.long)
return partial_mixup(input, gamma, indices), partial_mixup(target, gamma, indices)
def cross_entropy_loss(input: torch.Tensor,
target: torch.Tensor
) -> torch.Tensor:
return -(input.log_softmax(dim=-1) * target).sum(dim=-1).mean()
def run_one_epoch(loader, model, criterion, do_mixup=0., optimizer=None, assess=False):
device='cuda' if next(model.parameters()).is_cuda else 'cpu'
train = optimizer is not None # if we are in training mode there will be an optimizer and train=True here
n_classes = model.n_classes
if train: model.train()
else: model.eval()
if assess:
probs_all, preds_all, labels_all = [], [], []
with trange(len(loader)) as t:
n_elems, running_loss = 0, 0
for i_batch, batch in enumerate(loader):
if train:
lam = np.random.beta(a=do_mixup, b=1)
inputs, labels = batch[0][0], batch[0][1]
balanced_inputs, balanced_labels = batch[1][0], batch[1][1]
inputs, labels = inputs.to(device), labels.squeeze().to(device)
balanced_inputs, balanced_labels = balanced_inputs.to(device), balanced_labels.squeeze().to(device)
inputs = (1 - lam) * inputs + lam * balanced_inputs
mixed_labels = (1 - lam) * F.one_hot(labels, n_classes) + lam * F.one_hot(balanced_labels, n_classes)
del balanced_inputs
del balanced_labels
else:
inputs, labels = batch[0].to(device), batch[1].squeeze().to(device)
logits = model(inputs)
loss = cross_entropy_loss(logits, mixed_labels) if train else criterion(logits, labels)
if train: # only in training mode
loss.backward()
optimizer.step()
optimizer.zero_grad()
if assess:
probs = logits.softmax(dim=1)
preds = np.argmax(probs.detach().cpu().numpy(), axis=1)
probs_all.extend(probs.detach().cpu().numpy())
preds_all.extend(preds)
labels_all.extend(labels.cpu().numpy())
# Compute running loss
running_loss += loss.detach().item() * inputs.size(0)
n_elems += inputs.size(0)
run_loss = running_loss / n_elems
if train: t.set_postfix(loss_lr="{:.4f}/{:.6f}".format(float(run_loss), get_lr(optimizer)))
else: t.set_postfix(vl_loss="{:.4f}".format(float(run_loss)))
t.update()
if assess: return np.stack(preds_all), np.stack(probs_all), np.stack(labels_all), run_loss
return None, None, None, None
def train_model(model, sampling, optimizer, train_criterion, val_criterion, do_mixup, train_loader, val_loader,
scheduler, metric, n_epochs, exp_path, n_checkpoints):
best_loss, best_auc, best_bacc, best_k, best_mcc, best_f1, best_epoch, best_models = 10, 0, 0, 0, 0, 0, 0, []
is_better, best_monitoring_metric = compare_op(metric)
greater_is_better = best_monitoring_metric == 0
all_tr_aucs, all_vl_aucs, all_tr_mccs, all_vl_mccs = [], [], [], []
all_tr_ks, all_vl_ks, all_tr_baccs, all_vl_baccs, all_tr_losses, all_vl_losses = [], [], [], [], [], []
if model.n_classes == 5: class_names = ['DR0', 'DR1', 'DR2', 'DR3', 'DR4']
else: class_names = ['C{}'.format(i) for i in range(model.n_classes)]
print_conf, text_file_train, text_file_val = False, None, None
for epoch in range(n_epochs):
print('\nEpoch {:d}/{:d}'.format(epoch+1, n_epochs))
# Modify sampling
combo_loader = get_combo_loader(train_loader, base_sampling=sampling)
# train one epoch
_, _, _, _ = run_one_epoch(combo_loader, model, train_criterion, do_mixup, optimizer, assess=True)
with torch.no_grad():
tr_preds, tr_probs, tr_labels, tr_loss = run_one_epoch(train_loader, model, val_criterion, assess=True)
vl_preds, vl_probs, vl_labels, vl_loss = run_one_epoch(val_loader, model, val_criterion, assess=True)
if exp_path is not None:
print_conf = True
text_file_train = osp.join(exp_path,'performance_epoch_{}.txt'.format(str(epoch+1).zfill(2)))
text_file_val = osp.join(exp_path, 'performance_epoch_{}.txt'.format(str(epoch+1).zfill(2)))
tr_auc, tr_k, tr_mcc, tr_f1, tr_bacc, tr_auc_all, tr_f1_all = evaluate_multi_cls(tr_labels, tr_preds, tr_probs, print_conf=print_conf,
class_names=class_names, text_file=text_file_train, loss=tr_loss)
vl_auc, vl_k, vl_mcc, vl_f1, vl_bacc, vl_auc_all, vl_f1_all = evaluate_multi_cls(vl_labels, vl_preds, vl_probs, print_conf=print_conf,
class_names=class_names, text_file=text_file_val, loss=vl_loss, lr=get_lr(optimizer))
print('Train||Val Loss: {:.4f}||{:.4f} - K: {:.2f}||{:.2f} - mAUC: {:.2f}||{:.2f} - MCC: {:.2f}||{:.2f} - BACC: {:.2f}||{:.2f}'.format(
tr_loss, vl_loss, 100 * tr_k, 100 * vl_k, 100 * tr_auc, 100 * vl_auc, 100 * tr_mcc, 100 * vl_mcc, 100 * tr_bacc, 100 * vl_bacc))
all_tr_aucs.append(tr_auc_all)
all_vl_aucs.append(vl_auc_all)
all_tr_mccs.append(tr_mcc)
all_vl_mccs.append(vl_mcc)
all_tr_baccs.append(tr_bacc)
all_vl_baccs.append(vl_bacc)
all_tr_ks.append(tr_k)
all_vl_ks.append(vl_k)
all_tr_losses.append(tr_loss)
all_vl_losses.append(vl_loss)
# check if performance was better than anyone before and checkpoint if so
if metric == 'loss': tr_monitoring_metric, vl_monitoring_metric = tr_loss, vl_loss
elif metric == 'kappa': tr_monitoring_metric, vl_monitoring_metric = tr_k, vl_k
elif metric == 'mcc': tr_monitoring_metric, vl_monitoring_metric = tr_mcc, vl_mcc
elif metric == 'f1': tr_monitoring_metric, vl_monitoring_metric = tr_f1, vl_f1
elif metric == 'auc': tr_monitoring_metric, vl_monitoring_metric = tr_auc, vl_auc
elif metric == 'bacc': tr_monitoring_metric, vl_monitoring_metric = tr_bacc, vl_bacc
if tr_monitoring_metric > vl_monitoring_metric: # only if we do not underfit
scheduler.step(vl_monitoring_metric)
if is_better(vl_monitoring_metric, best_monitoring_metric):
print('-------- Best {} attained. {:.2f} --> {:.2f} --------'.format(metric, 100*best_monitoring_metric, 100*vl_monitoring_metric))
best_loss, best_k, best_mcc, best_f1, best_auc, best_bacc, best_epoch = vl_loss, vl_k, vl_mcc, vl_f1, vl_auc, vl_bacc, epoch+1
best_monitoring_metric = vl_monitoring_metric
else:
print('-------- Best {} so far {:.2f} at epoch {:d} --------'.format(metric, 100 * best_monitoring_metric,
best_epoch))
# SAVE n best - keep deleting worse ones
if exp_path is not None:
s_name = 'epoch_{}_K_{:.2f}_mAUC_{:.2f}_MCC_{:.2f}'.format(str(epoch + 1).zfill(2), 100 * vl_k,
100 * vl_auc, 100 * vl_mcc)
best_models.append([osp.join(exp_path, s_name), vl_monitoring_metric])
if epoch < n_checkpoints: # first n_checkpoints epochs save always
print('-------- Checkpointing to {}/ --------'.format(s_name))
save_model(osp.join(exp_path, s_name), model, optimizer, weights=True)
else:
worst_model = sorted(best_models, key=itemgetter(1), reverse=greater_is_better)[-1][
0] # False for Loss, True for K
if s_name != worst_model: # this model was better than one of the best n_checkpoints models, remove that one
print('-------- Checkpointing to {}/ --------'.format(s_name))
save_model(osp.join(exp_path, s_name), model, optimizer, weights=True)
# print('before deleting', os.listdir(osp.join(exp_path, s_name)))
print('----------- Deleting {}/ -----------'.format(worst_model.split('/')[-1]))
shutil.rmtree(worst_model)
best_models = sorted(best_models, key=itemgetter(1), reverse=greater_is_better)[:n_checkpoints]
if np.isclose(get_lr(optimizer), scheduler.min_lrs[0]):
print('Early stopping')
del model
torch.cuda.empty_cache()
return best_auc, best_bacc, best_mcc, best_k, all_tr_aucs, all_vl_aucs, all_tr_mccs, all_vl_mccs, \
all_tr_ks, all_vl_ks, all_tr_losses, all_vl_losses, best_epoch
del model
torch.cuda.empty_cache()
return best_auc, best_bacc, best_mcc, best_k, all_tr_aucs, all_vl_aucs, all_tr_mccs, all_vl_mccs, \
all_tr_ks, all_vl_ks, all_tr_losses, all_vl_losses, best_epoch
if __name__ == '__main__':
args = parser.parse_args()
use_cuda = torch.cuda.is_available()
device = torch.device("cuda:0" if use_cuda else "cpu")
# reproducibility
seed_value = 0
set_seeds(seed_value, use_cuda)
data_path = args.data_path
n_classes = args.n_classes
# gather parser parameters
sampling = args.sampling
model_name = args.model_name
optimizer_choice = args.optimizer
lr, min_lr, patience, bs = args.lr, args.min_lr, args.patience, args.batch_size
if min_lr == -1: min_lr = lr * 1e-3
n_epochs, metric = args.n_epochs, args.metric
im_size = tuple([int(item) for item in args.im_size.split(',')])
if isinstance(im_size, tuple) and len(im_size)==1:
tg_size = (im_size[0], im_size[0])
elif isinstance(im_size, tuple) and len(im_size)==2:
tg_size = (im_size[0], im_size[1])
else:
sys.exit('im_size should be a number or a tuple of two numbers')
do_not_save = str2bool(args.do_not_save)
if do_not_save is False:
save_path = args.save_path
if save_path == 'date_time':
save_path = datetime.now().strftime("%Y-%m-%d-%H:%M:%S")
experiment_path=osp.join('experiments', save_path)
args.experiment_path = experiment_path
os.makedirs(experiment_path, exist_ok=True)
n_checkpoints = args.n_checkpoints
config_file_path = osp.join(experiment_path,'config.cfg')
with open(config_file_path, 'w') as f:
json.dump(vars(args), f, indent=2)
else: experiment_path, n_checkpoints=None, 0
csv_train = args.csv_train
csv_val = csv_train.replace('train', 'val')
print('* Instantiating a {} model'.format(model_name))
model, mean, std = get_arch(model_name, n_classes=n_classes)
print('* Creating Dataloaders, batch size = {}, workers = {}'.format(bs, args.num_workers))
train_loader, val_loader = get_train_val_cls_loaders(csv_path_train=csv_train, csv_path_val=csv_val,
data_path=data_path, batch_size=bs,
tg_size=tg_size, mean=mean, std=std,
num_workers=args.num_workers)
model = model.to(device)
print("Total params: {0:,}".format(sum(p.numel() for p in model.parameters() if p.requires_grad)))
if optimizer_choice == 'adam':
optimizer = torch.optim.Adam(model.parameters(), lr=lr)
elif optimizer_choice == 'sgd':
optimizer = torch.optim.SGD(model.parameters(), lr=lr, weight_decay=args.wd)
else:
sys.exit('please choose a valid optimizer')
if args.pretrained_weights is True:
weights_path = osp.join('data/pretrained_weights/', model_name)
try:
model, stats, optimizer_state_dict = load_model(model, args.resume_path, device=device, with_opt=True)
optimizer.load_state_dict(optimizer_state_dict)
except:
sys.exit('Pretrained weights not compatible for this model')
for i, param_group in enumerate(optimizer.param_groups):
param_group['lr'] = lr
scheduler = ReduceLROnPlateau(optimizer, mode='max', factor=0.1, patience=patience, min_lr=min_lr, verbose=False)
train_criterion, val_criterion = torch.nn.CrossEntropyLoss(), torch.nn.CrossEntropyLoss()
do_mixup = args.do_mixup
print('* Instantiating loss function', str(train_criterion))
print('* Starting to train\n','-' * 10)
start = time.time()
b_mauc, b_bacc, b_mcc, b_k, tr_aucs, vl_aucs, \
tr_mccs, vl_mccs, tr_ks, vl_ks, tr_ls, vl_ls, b_epoch = train_model(model, sampling, optimizer, train_criterion, val_criterion,
do_mixup, train_loader, val_loader, scheduler,
metric, n_epochs, experiment_path, n_checkpoints)
end = time.time()
hours, rem = divmod(end - start, 3600)
minutes, seconds = divmod(rem, 60)
print("b_mauc: %f" % b_mauc)
print("b_mcc: %f" % b_mcc)
print("b_k: %f" % b_k)
print("b_epoch: %d" % b_epoch)
if do_not_save is False:
with open(osp.join(experiment_path, 'val_metrics.txt'), 'w') as f:
print(
'Best K = {:.2f}\nBest mAUC = {:.2f}\nBest MCC = {:.2f}\nBest BACC = {:.2f}\nBest epoch = {}\n'.format(
100 * b_k, 100 * b_mauc, 100 * b_mcc, 100 * b_bacc, b_epoch), file=f)
for j in range(len(vl_aucs)):
print(
'Epoch = {} -> K={:.2f}/{:.2f}, mAUC={:.2f}/{:.2f}, MCC={:.2f}/{:.2f}, Loss={:.4f}/{:.4f},'.format(
j + 1, 100 * tr_ks[j], 100 * vl_ks[j],
100 * np.mean(tr_aucs[j]), 100 * np.mean(vl_aucs[j]),
100 * tr_mccs[j], 100 * vl_mccs[j], tr_ls[j], vl_ls[j]), file=f)
print('\nTraining time: {:0>2}h {:0>2}min {:05.2f}secs'.format(int(hours), int(minutes), seconds), file=f)