-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
191 lines (160 loc) · 6.98 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
# Training script for the project
# Author: Simon Zhou, last modify Nov. 18, 2022
'''
Change log:
-Simon: file created, write some training code
-Simon: refine training script
'''
import argparse
import os
import sys
sys.path.append("../")
from tqdm import trange
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torchvision.models import vgg16_bn
import meta_config as config
from model import *
from our_utils import *
from dataset_loader import *
from loss import *
import wandb
parser = argparse.ArgumentParser(description='parameters for the training script')
parser.add_argument('--dataset', type=str, default="CT-MRI", help="which dataset to use, available option: CT-MRI, MRI-PET, MRI-SPECT")
parser.add_argument('--batch_size', type=int, default=4, help='batch size for training')
parser.add_argument('--epochs', type=int, default=100, help='number of epochs for training')
parser.add_argument('--lr', type=float, default=0.0001, help='learning rate for training')
parser.add_argument('--lr_decay', type=bool, default=False, help='decay learing rate?')
parser.add_argument('--accum_batch', type=int, default=1, help='number of batches for gradient accumulation')
parser.add_argument('--lambda1', type=float, default=0.5, help='weight for image gradient loss')
parser.add_argument('--lambda2', type=float, default=0.5, help='weight for perceptual loss')
#parser.add_argument('--checkpoint', type=str, default='./model', help='Path to checkpoint')
parser.add_argument('--cuda', action='store_true', help='whether to use cuda', default= True)
parser.add_argument('--seed', type=int, default=3407, help='random seed to use')
parser.add_argument('--base_loss', type=str, default='l1_charbonnier', help='which loss function to use for pixel-level (l2 or l1 charbonnier)')
opt = parser.parse_args()
######### whether to use cuda ####################
device = torch.device("cuda:0" if opt.cuda else "cpu")
#################################################
########## seeding ##############
seed_val = opt.seed
random_seed(seed_val, opt.cuda)
################################
############ making dirs########################
if not os.path.exists(config.res_dir):
os.mkdir(config.res_dir)
model_dir = os.path.join(config.res_dir, "pretrained_models")
if not os.path.exists(model_dir):
os.mkdir(model_dir)
if not os.path.exists(config.test_data_dir):
os.mkdir(config.test_data_dir)
################################################
####### loading dataset ####################################
target_dir = os.path.join(config.data_dir, opt.dataset)
ct, mri = get_common_file(target_dir)
train_ct, train_mri, test_ct, test_mri = load_data(ct, target_dir, config.test_num)
# torch.save(test_ct, os.path.join(c.test_data_dir, "ct_test.pt"))
# torch.save(test_mri, os.path.join(c.test_data_dir, "mri_test.pt"))
#print(train_ct.shape, train_mri.shape, test_ct.shape, test_mri.shape)
train_total = torch.cat((train_ct, train_mri), dim = 0).to(device)
# these loaders return index, not the actual image
train_loader, val_loader = get_loader(train_ct, train_mri, config.train_val_ratio, opt.batch_size)
print("train loader length: ", len(train_loader), " val loder length: ", len(val_loader))
# check the seed is working
# for batch_idx in train_loader:
# batch_idx = batch_idx.view(-1).long()
# print(batch_idx)
# print("validation index")
# for batch_idx in val_loader:
# batch_idx = batch_idx.view(-1).long()
# print(batch_idx)
# sys.exit()
############################################################
############ loading model #####################
model = fullModel().to(device)
optimizer = optim.Adam(model.parameters(), lr=opt.lr)
if opt.lr_decay:
stepLR = optim.lr_scheduler.StepLR(optimizer, step_size = 100, gamma=0.5)
###################################################
##### downloading pretrained vgg model ##################
vgg = vgg16_bn(pretrained = True)
########################################################
############## train model ##############
wandb.init(project="test-project", entity="csc2529", config=opt) # visualize in wandb
# wandb.config = {
# "learning_rate": opt.lr,
# "epochs": opt.epochs,
# "batch_size": opt.batch_size,
# "lambda1": c.lambda1,
# "lambda2": c.lambda2
# }
wandb.watch(model)
# gradient accumulation for small batch
NUM_ACCUMULATION_STEPS = opt.accum_batch
train_loss = []
val_loss = []
t = trange(opt.epochs, desc='Training progress...', leave=True)
lowest_val_loss = int(1e9)
for i in t:
print("new epoch {} starts!".format(i))
# clear gradient in model
model.zero_grad()
b_loss = 0
# train model
model.train()
for j, batch_idx in enumerate(train_loader):
# clear gradient in optimizer
optimizer.zero_grad()
batch_idx = batch_idx.view(-1).long()
img = train_total[batch_idx]
img_out = model(img)
# compute loss
loss,_,_,_ = loss_func2(vgg, img_out, img, opt.lambda1, opt.lambda2, config.block_idx, device)
# back propagate and update weights
#print("batch reg, grad, percep loss: ", reg_loss.item(), img_grad.item(), percep.item())
#loss = loss / NUM_ACCUMULATION_STEPS
loss.backward()
#if ((j + 1) % NUM_ACCUMULATION_STEPS == 0) or (j + 1 == len(train_loader)):
optimizer.step()
b_loss += loss.item()
#wandb.log({"loss": loss})
# store loss
ave_loss = b_loss / len(train_loader)
train_loss.append(ave_loss)
print("epoch {}, training loss is: {}".format(i, ave_loss))
# validation
val_loss = []
val_display_img = []
with torch.no_grad():
b_loss = 0
# eval model, unable update weights
model.eval()
for k, batch_idx in enumerate(val_loader):
batch_idx = batch_idx.view(-1).long()
val_img = train_total[batch_idx]
val_img_out = model(val_img)
# display first image to visualize, this can be changed
val_display_img.extend([val_img_out[i].squeeze(0).cpu().numpy() for i in range(1)])
loss, _,_,_= loss_func2(vgg, img_out, img, opt.lambda1, opt.lambda2, config.block_idx, device)
b_loss += loss.item()
ave_val_loss = b_loss / len(val_loader)
val_loss.append(ave_val_loss)
print("epoch {}, validation loss is: {}".format(i, ave_val_loss))
# define a metric we are interested in the minimum of
wandb.define_metric("train loss", summary="min")
# define a metric we are interested in the maximum of
wandb.define_metric("val loss", summary="min")
wandb.log({"train loss": ave_loss, "epoch": i})
wandb.log({"val loss": ave_val_loss, "epoch": i})
wandb.log({"val sample images": [wandb.Image(img) for img in val_display_img]})
# save model
if ave_val_loss < lowest_val_loss:
torch.save(model.state_dict(), model_dir+"/model_at_{}.pt".format(i))
lowest_val_loss = ave_val_loss
print("model is saved in epoch {}".format(i))
# lr decay update
if opt.lr_decay:
stepLR.step()
########################################