-
Notifications
You must be signed in to change notification settings - Fork 3
/
main.py
192 lines (158 loc) · 9.98 KB
/
main.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
import os
import shutil
import time
import argparse
import cv2
import numpy as np
import pandas as pd
import torch
import torch.nn.functional as F
import torchvision
import tqdm
from sklearn.metrics import accuracy_score, roc_auc_score, classification_report
from sklearn.utils import class_weight
from datasets import *
from utils import *
from losses import *
def main(args):
# Set model/output directory name
MODEL_NAME = args.dataset
MODEL_NAME += f'_{args.model_name}'
MODEL_NAME += f'_rand' if args.rand_init else ''
MODEL_NAME += f'_bal-mixup-{args.mixup_alpha}' if args.bal_mixup else ''
MODEL_NAME += f'_mixup-{args.mixup_alpha}' if args.mixup else ''
MODEL_NAME += f'_decoupling-{args.decoupling_method}' if args.decoupling_method != '' else ''
MODEL_NAME += f'_rw-{args.rw_method}' if args.rw_method != '' else ''
MODEL_NAME += f'_{args.loss}'
MODEL_NAME += '-drw' if args.drw else ''
MODEL_NAME += f'_cb-beta-{args.cb_beta}' if args.rw_method == 'cb' else ''
MODEL_NAME += f'_fl-gamma-{args.fl_gamma}' if args.loss == 'focal' else ''
MODEL_NAME += f'_lr-{args.lr}'
MODEL_NAME += f'_bs-{args.batch_size}'
# Create output directory for model (and delete if already exists)
if not os.path.isdir(args.out_dir):
os.mkdir(args.out_dir)
model_dir = os.path.join(args.out_dir, MODEL_NAME)
if os.path.isdir(model_dir):
shutil.rmtree(model_dir)
os.mkdir(model_dir)
# Set all seeds for reproducibility
set_seed(args.seed)
# Create datasets + loaders
if args.dataset == 'nih-lt':
dataset = NIH_CXR_Dataset
N_CLASSES = 20
else:
dataset = MIMIC_CXR_Dataset
N_CLASSES = 19
train_dataset = dataset(data_dir=args.data_dir, label_dir=args.label_dir, split='train')
val_dataset = dataset(data_dir=args.data_dir, label_dir=args.label_dir, split='balanced-val')
bal_test_dataset = dataset(data_dir=args.data_dir, label_dir=args.label_dir, split='balanced-test')
test_dataset = dataset(data_dir=args.data_dir, label_dir=args.label_dir, split='test')
if args.bal_mixup:
cls_weights = [len(train_dataset) / cls_count for cls_count in train_dataset.cls_num_list]
instance_weights = [cls_weights[label] for label in train_dataset.labels]
sampler = torch.utils.data.WeightedRandomSampler(torch.Tensor(instance_weights), len(train_dataset))
bal_train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=False, num_workers=8, pin_memory=True, worker_init_fn=worker_init_fn, sampler=sampler)
imbal_train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=8, pin_memory=True, worker_init_fn=worker_init_fn)
train_loader = ComboLoader([imbal_train_loader, bal_train_loader])
elif args.decoupling_method == 'cRT':
cls_weights = [len(train_dataset) / cls_count for cls_count in train_dataset.cls_num_list]
instance_weights = [cls_weights[label] for label in train_dataset.labels]
sampler = torch.utils.data.WeightedRandomSampler(torch.Tensor(instance_weights), len(train_dataset))
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=False, num_workers=8, pin_memory=True, worker_init_fn=worker_init_fn, sampler=sampler)
else:
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=8, pin_memory=True, worker_init_fn=worker_init_fn)
val_loader = torch.utils.data.DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=2, pin_memory=True, worker_init_fn=val_worker_init_fn)
# Create csv documenting training history
history = pd.DataFrame(columns=['epoch', 'phase', 'loss', 'balanced_acc', 'mcc', 'auroc'])
history.to_csv(os.path.join(model_dir, 'history.csv'), index=False)
# Set device
device = torch.device('cuda:0')
# Instantiate model
model = torchvision.models.resnet50(pretrained=(not args.rand_init))
model.fc = torch.nn.Linear(model.fc.in_features, N_CLASSES)
if args.decoupling_method == 'tau_norm':
msg = model.load_state_dict(torch.load(args.decoupling_weights, map_location='cpu')['weights'])
print(f'Loaded weights from {args.decoupling_weights} with message: {msg}')
model.fc.bias.data = torch.zeros_like(model.fc.bias.data)
fc_weights = model.fc.weight.data.clone()
weight_norms = torch.norm(fc_weights, 2, 1)
model.fc.weight.data = torch.stack([fc_weights[i] / torch.pow(weight_norms[i], -4) for i in range(N_CLASSES)], dim=0)
elif args.decoupling_method == 'cRT':
msg = model.load_state_dict(torch.load(args.decoupling_weights, map_location='cpu')['weights'])
print(f'Loaded weights from {args.decoupling_weights} with message: {msg}')
model.fc = torch.nn.Linear(model.fc.in_features, N_CLASSES) # re-initialize classifier head
model = model.to(device)
# Set loss and weighting method
if args.rw_method == 'sklearn':
weights = class_weight.compute_class_weight(class_weight='balanced', classes=np.unique(train_dataset.labels), y=np.array(train_dataset.labels))
weights = torch.Tensor(weights).to(device)
elif args.rw_method == 'cb':
weights = get_CB_weights(samples_per_cls=train_dataset.cls_num_list, beta=args.cb_beta)
weights = torch.Tensor(weights).to(device)
else:
weights = None
if weights is None:
print('No class reweighting')
else:
print(f'Class weights with rw_method {args.rw_method}:')
for i, c in enumerate(train_dataset.CLASSES):
print(f'\t{c}: {weights[i]}')
loss_fxn = get_loss(args, None if args.drw else weights, train_dataset)
# Set optimizer
if args.decoupling_method != '':
optimizer = torch.optim.Adam(model.fc.parameters(), lr=args.lr)
else:
optimizer = torch.optim.Adam(model.parameters(), lr=args.lr)
# Train with early stopping
if args.decoupling_method != 'tau_norm':
epoch = 1
early_stopping_dict = {'best_acc': 0., 'epochs_no_improve': 0}
best_model_wts = None
while epoch <= args.max_epochs and early_stopping_dict['epochs_no_improve'] <= args.patience:
if args.bal_mixup:
history = bal_mixup_train(model=model, device=device, loss_fxn=loss_fxn, optimizer=optimizer, data_loader=train_loader, history=history, epoch=epoch, model_dir=model_dir, classes=train_dataset.CLASSES, mixup_alpha=args.mixup_alpha)
else:
history = train(model=model, device=device, loss_fxn=loss_fxn, optimizer=optimizer, data_loader=train_loader, history=history, epoch=epoch, model_dir=model_dir, classes=train_dataset.CLASSES, mixup=args.mixup, mixup_alpha=args.mixup_alpha)
history, early_stopping_dict, best_model_wts = validate(model=model, device=device, loss_fxn=loss_fxn, optimizer=optimizer, data_loader=val_loader, history=history, epoch=epoch, model_dir=model_dir, early_stopping_dict=early_stopping_dict, best_model_wts=best_model_wts, classes=val_dataset.CLASSES)
if args.drw and epoch == 10:
for g in optimizer.param_groups:
g['lr'] *= 0.1 # anneal LR
loss_fxn = get_loss(args, weights, train_dataset) # get class-weighted loss
early_stopping_dict['epochs_no_improve'] = 0 # reset patience
epoch += 1
else:
best_model_wts = model.state_dict()
# Evaluate on balanced test set
evaluate(model=model, device=device, loss_fxn=loss_fxn, dataset=bal_test_dataset, split='balanced-test', batch_size=args.batch_size, history=history, model_dir=model_dir, weights=best_model_wts)
# Evaluate on imbalanced test set
evaluate(model=model, device=device, loss_fxn=loss_fxn, dataset=test_dataset, split='test', batch_size=args.batch_size, history=history, model_dir=model_dir, weights=best_model_wts)
if __name__ == '__main__':
# Command-line arguments
parser = argparse.ArgumentParser()
parser.add_argument('--data_dir', default='/ssd1/greg/NIH_CXR/images', type=str)
parser.add_argument('--label_dir', default='labels/', type=str)
parser.add_argument('--out_dir', default='results/', type=str, help="path to directory where results and model weights will be saved")
parser.add_argument('--dataset', required=True, type=str, choices=['nih-lt', 'mimic-cxr-lt'])
parser.add_argument('--loss', default='ce', type=str, choices=['ce', 'focal', 'ldam'])
parser.add_argument('--drw', action='store_true', default=False)
parser.add_argument('--rw_method', default='', choices=['', 'sklearn', 'cb'])
parser.add_argument('--cb_beta', default=0.9999, type=float)
parser.add_argument('--fl_gamma', default=2., type=float)
parser.add_argument('--bal_mixup', action='store_true', default=False)
parser.add_argument('--mixup', action='store_true', default=False)
parser.add_argument('--mixup_alpha', default=0.2, type=float)
parser.add_argument('--decoupling_method', default='', choices=['', 'cRT', 'tau_norm'], type=str)
parser.add_argument('--decoupling_weights', type=str)
parser.add_argument('--model_name', default='resnet50', type=str, help="CNN backbone to use")
parser.add_argument('--max_epochs', default=60, type=int, help="maximum number of epochs to train")
parser.add_argument('--batch_size', default=256, type=int, help="batch size for training, validation, and testing (will be lowered if TTA used)")
parser.add_argument('--lr', default=1e-4, type=float)
parser.add_argument('--patience', default=15, type=int, help="early stopping 'patience' during training")
parser.add_argument('--rand_init', action='store_true', default=False)
parser.add_argument('--n_TTA', default=0, type=int, help="number of augmented copies to use during test-time augmentation (TTA), default 0")
parser.add_argument('--seed', default=0, type=int, help="set random seed")
args = parser.parse_args()
print(args)
main(args)