-
Notifications
You must be signed in to change notification settings - Fork 2
/
main.py
172 lines (156 loc) · 7.2 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
# -*- coding: utf-8 -*-
import torch, time, os, shutil
import models, utils
import numpy as np
import pandas as pd
#from tensorboard_logger import Logger
from torch import nn, optim
from torch.utils.data import DataLoader
from dataset import ECGDataset
from config import config
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torch.manual_seed(41)
torch.cuda.manual_seed(41)
# save current weight and update new best weight
def save_ckpt(state, is_best, model_save_dir):
current_w = os.path.join(model_save_dir, config.current_w)
best_w = os.path.join(model_save_dir, config.best_w)
torch.save(state, current_w)
if is_best: shutil.copyfile(current_w, best_w)
def train_epoch(model, optimizer, criterion, train_dataloader, show_interval=10):
model.train()
f1_meter, loss_meter, it_count = 0, 0, 0
for inputs, target in train_dataloader:
inputs = inputs.to(device)
target = target.to(device)
# zero the parameter gradients
optimizer.zero_grad()
# forward
output = model(inputs)
loss = criterion(output, target)
loss.backward()
optimizer.step()
loss_meter += loss.item()
it_count += 1
f1 = utils.calc_f1(target, torch.sigmoid(output))
f1_meter += f1
if it_count != 0 and it_count % show_interval == 0:
print("%d,loss:%.3e f1:%.3f" % (it_count, loss.item(), f1))
return loss_meter / it_count, f1_meter / it_count
def val_epoch(model, criterion, val_dataloader, threshold=0.5):
model.eval()
f1_meter,acc_meter,recall_meter,precision_meter,loss_meter, it_count = 0, 0, 0,0,0,0
with torch.no_grad():
for inputs, target in val_dataloader:
inputs = inputs.to(device)
target = target.to(device)
output = model(inputs)
loss = criterion(output, target)
loss_meter += loss.item()
it_count += 1
output = torch.sigmoid(output)
f1 = utils.calc_f1(target, output, threshold)
acc = utils.cal_accuracy_score(target,output)
recall = utils.cal_recall_score(target,output)
precision = utils.cal_percision_score(target,output)
f1_meter += f1
acc_meter +=acc
recall_meter +=recall
precision_meter += precision
return loss_meter / it_count, f1_meter / it_count , acc_meter / it_count , recall_meter / it_count, precision_meter / it_count
def train(args):
# get model
model = getattr(models, config.model_name)()
if args.ckpt and not args.resume:
state = torch.load(args.ckpt, map_location='cpu')
model.load_state_dict(state['state_dict'])
print('train with pretrained weight val_f1', state['f1'])
model = model.to(device)
print(model)
# data
train_dataset = ECGDataset(data_path=config.train_data, train=True)
train_dataloader = DataLoader(train_dataset, batch_size=config.batch_size, shuffle=True, num_workers=6)
val_dataset = ECGDataset(data_path=config.train_data, train=False)
val_dataloader = DataLoader(val_dataset, batch_size=config.batch_size, num_workers=4)
print("train_datasize", len(train_dataset), "val_datasize", len(val_dataset))
# optimizer and loss
optimizer = optim.Adam(model.parameters(), lr=config.lr)
w = torch.tensor(train_dataset.wc, dtype=torch.float).to(device)
criterion = utils.FocalLoss()
# model save dir
model_save_dir = '%s/%s_%s' % (config.ckpt, config.model_name, time.strftime("%Y%m%d%H%M"))
utils.mkdirs(model_save_dir)
if args.ex: model_save_dir += args.ex
best_f1 = -1
lr = config.lr
start_epoch = 1
stage = 1
# train from last save point
if args.resume:
if os.path.exists(args.ckpt): # weight path
model_save_dir = args.ckpt
current_w = torch.load(os.path.join(args.ckpt, config.current_w))
best_w = torch.load(os.path.join(model_save_dir, config.best_w))
best_f1 = best_w['loss']
start_epoch = current_w['epoch'] + 1
lr = current_w['lr']
stage = current_w['stage']
model.load_state_dict(current_w['state_dict'])
if start_epoch - 1 in config.stage_epoch:
stage += 1
lr /= config.lr_decay
utils.adjust_learning_rate(optimizer, lr)
model.load_state_dict(best_w['state_dict'])
print("=> loaded checkpoint (epoch {})".format(start_epoch - 1))
#logger = Logger(logdir=model_save_dir, flush_secs=2)
# =========>start training<=========
for epoch in range(start_epoch, config.max_epoch + 1):
since = time.time()
train_loss, train_f1 = train_epoch(model, optimizer, criterion, train_dataloader, show_interval=100)
val_loss, val_f1 = val_epoch(model, criterion, val_dataloader)
print('#epoch:%02d stage:%d train_loss:%.3e train_f1:%.3f val_loss:%0.3e val_f1:%.3f time:%s\n'
% (epoch, stage, train_loss, train_f1, val_loss, val_f1, utils.print_time_cost(since)))
# logger.log_value('train_loss', train_loss, step=epoch)
# logger.log_value('train_f1', train_f1, step=epoch)
# logger.log_value('val_loss', val_loss, step=epoch)
# logger.log_value('val_f1', val_f1, step=epoch)
state = {"state_dict": model.state_dict(), "epoch": epoch, "loss": val_loss, 'f1': val_f1, 'lr': lr,
'stage': stage}
save_ckpt(state, best_f1 < val_f1, model_save_dir)
best_f1 = max(best_f1, val_f1)
print(best_f1)
if epoch in config.stage_epoch:
stage += 1
lr /= config.lr_decay
best_w = os.path.join(model_save_dir, config.best_w)
model.load_state_dict(torch.load(best_w)['state_dict'])
print("*" * 10, "step into stage%02d lr %.3ef" % (stage, lr))
utils.adjust_learning_rate(optimizer, lr)
# test
def val(args):
list_threhold = [0.5]
model = models.resnet34()
if args.ckpt: model.load_state_dict(torch.load(args.ckpt, map_location='cpu')['state_dict'])
model = model.to(device)
criterion = nn.BCEWithLogitsLoss()
val_dataset = ECGDataset(data_path=config.train_data, train=False)
val_dataloader = DataLoader(val_dataset, batch_size=config.batch_size, num_workers=4)
for threshold in list_threhold:
val_loss, val_f1,val_acc,val_recall,val_precision = val_epoch(model, criterion, val_dataloader, threshold)
print('threshold %.2f val_loss:%0.6e val_f1:%.8f\n , val_acc: %.8f\n val_recall: %.8f\n val_precision : %.8f\n' % (threshold, val_loss, val_f1,val_acc,val_recall,val_precision))
if __name__ == '__main__':
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("command", metavar="<command>", help="train or infer")
parser.add_argument("--ckpt", type=str, help="the path of model weight file")
parser.add_argument("--ex", type=str, help="experience name")
parser.add_argument("--resume", action='store_true', default=False)
args = parser.parse_args()
if (args.command == "train"):
train(args)
if (args.command == "test"):
test(args)
if (args.command == "val"):
val(args)