-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_frrn.py
220 lines (195 loc) · 9.54 KB
/
train_frrn.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
"""
Given yaml config file for model training
"""
from datetime import datetime
import os
from PIL import Image
import random
import torch
import numpy as np
from tqdm import tqdm
from torch.optim.lr_scheduler import CosineAnnealingLR
from torch.utils.tensorboard import SummaryWriter
import sys
import wandb
import yaml
from dataloader.a2d2_loader import get_dataloader
from models.frrn import FRRNet
from models.frrn_ensemble import Ensemble
from models.frrn_moe import MoE
from utils.metrics import Evaluator
from utils.saver import Saver
from utils.train_utils import cross_entropy2d, EarlyStopping, colorize
class Trainer():
"""Define model trainer"""
def __init__(self, _config):
self.params = params = _config
params["gpu_ids"] = [0]
self.is_cuda = len(params["gpu_ids"]) > 0
self.device = torch.device('cuda', params["gpu_ids"][0]) \
if self.is_cuda else torch.device('cpu')
self.saver = Saver(params)
log_dir = os.path.join(params["output"], params["start_time"] + "_" + params["checkname"] + "/")
self.writer = SummaryWriter(log_dir)
# Datasets
self.datasets = list(s for s in params["DATASET"]["dataset"].split(','))
self.train_loader, self.val_loader, self.label_names, self.label_colors \
= get_dataloader(params, self.datasets)
if 'moe' in params['MODEL']['expert'].lower():
expert_names = list(s for s in params['DATASET']['dataset'].split(','))
checkpoints = [os.path.join(*[params['DIR'],
params['MODEL']['arch'] + '_' + expert_name,
'model_best.pth'])
for expert_name in expert_names]
# fully connect layer feature depend on scale factor
l_feat = int((params['DATASET']['img_height']/16) * (params['DATASET']['img_width']/16))
self.model = MoE(3, params['DATASET']['num_class'],
l_feat,
checkpoints,
params['MODEL']['gate'],
params['MODEL']['with_conv'],
len(expert_names))
trainable_parameters = model._get_train_params()
else:
self.model = FRRNet(out_channels=params['DATASET']['num_class'])
trainable_parameters = model.parameters()
self.optimizer = torch.optim.Adam(
trainable_parameters, lr=params["TRAIN"]["lr"], betas=(0.9, 0.99),
weight_decay=params["TRAIN"]["weight_decay"])
self.scheduler = CosineAnnealingLR(
self.optimizer, params["TRAIN"]["num_epochs"], eta_min=1e-6)
self.loss = cross_entropy2d
self.evaluator = Evaluator(params['DATASET']['num_class'])
if self.is_cuda:
self.model = torch.nn.DataParallel(self.model, device_ids=params["gpu_ids"])
self.model = self.model.to(self.device)
pytorch_total_params = sum(p.numel() for p in self.model.parameters() if p.requires_grad)
print("Total trainable params", pytorch_total_params)
self.best_pred = 0.0
if os.path.isfile(params["TRAIN"]["resume"]):
checkpoint = torch.load(params["TRAIN"]["resume"])
params["TRAIN"]["start_epoch"] = checkpoint['epoch']
if self.is_cuda:
self.model.module.load_state_dict(checkpoint['state_dict'])
else:
self.model.load_state_dict(checkpoint['state_dict'])
self.optimizer.load_state_dict(checkpoint['optimizer'])
self.best_pred = checkpoint['best_pred']
print("=> loaded checkpoint '{}' (epoch {})".format(params["TRAIN"]["resume"],
checkpoint['epoch']))
def training(self, epoch):
"""train model iteration from one epoch"""
train_loss = 0.0
self.model.train()
tbar = tqdm(self.train_loader)
num_img_tr = len(self.train_loader)
for i, sample in enumerate(tbar):
image, label = sample
image = image.to(self.device)
label = label.to(self.device)
self.optimizer.zero_grad()
output = self.model(image)
loss = self.loss(output, label)
loss.backward()
self.optimizer.step()
# adjust learning rate each step
# self.scheduler.step()
train_loss += loss.item()
tbar.set_description('Epoch: {:<3d}, Train Loss: {:.3f}'.format(
epoch, train_loss / (i + 1)))
with torch.no_grad():
wandb.log({"train/total_loss_iter": float(loss.item())})
with torch.no_grad():
wandb.log({"train/total_loss_epoch": float(loss.item()), "epoch": epoch})
def validation(self, epoch):
"""validate model"""
# pylint: disable-msg=too-many-locals
test_loss = 0.0
self.model.eval()
self.evaluator.reset()
tbar = tqdm(self.val_loader, desc='\r')
num_img_val = len(self.val_loader)
for i, sample in enumerate(tbar):
image, label = sample
image = image.to(self.device)
label = label.to(self.device)
with torch.no_grad():
output = self.model(image)
loss = self.loss(output, label)
test_loss += loss.item()
tbar.set_description('Epoch: {:<3d}, Val Loss: {:.3f}'.format(
epoch, test_loss / (i + 1)))
pred = output.data.cpu().numpy()
label = label.cpu().numpy()
pred = np.argmax(pred, axis=1)
self.evaluator.add_batch(label, pred)
pred = output[0].data.max(0)[1].cpu().numpy()
ground_true = label[0]
wandb.log({"val_img": wandb.Image(image)})
wandb.log({"val_gt": wandb.Image(colorize(self.label_colors, ground_true))})
wandb.log({"val_pred": wandb.Image(colorize(self.label_colors, pred))})
# Fast test during the training
acc = self.evaluator.pixel_accuracy()
acc_class = self.evaluator.pixel_accuracy_class()
miou, log_miou_cls = self.evaluator.mean_intersection_over_union(self.label_names)
fwiou = self.evaluator.frequency_weighted_intersection_over_union()
self.writer.add_scalar("val/total_loss_epoch", test_loss, epoch)
self.writer.add_scalar("val/mIoU", miou, epoch)
self.writer.add_scalar("val/Acc", acc, epoch)
self.writer.add_scalar("val/Acc_class", acc_class, epoch)
self.writer.add_scalar("val/fwIoU", fwiou, epoch)
wandb.log({"val/total_loss_epoch": test_loss, "epoch": epoch})
wandb.log({"val/mIoU": miou, "epoch": epoch})
wandb.log({"val/Acc": acc, "epoch": epoch})
wandb.log({"val/Acc_class": acc_class, "epoch": epoch})
wandb.log({"val/fwIoU": fwiou, "epoch": epoch})
wandb.log({"val/loss": test_loss, "epoch": epoch})
new_pred = miou
if new_pred > self.best_pred:
is_best = True
self.best_pred = new_pred
self.saver.save_checkpoint({
'epoch': epoch + 1,
'state_dict': self.model.module.state_dict(),
'optimizer': self.optimizer.state_dict(),
'best_pred': self.best_pred,
}, is_best)
return test_loss/num_img_val
if __name__ == '__main__':
if len(sys.argv) != 2:
print('\nPlease pass the desired param file for training as an argument.\n'
'e.g: params/params_moe.py')
else:
print('STARTING EVALUATION WITH PARAM FILE: ', str(sys.argv[1]))
with open(str(sys.argv[1]), 'r') as stream:
try:
params = yaml.safe_load(stream)
torch.manual_seed(params["TRAIN"]["seed"])
torch.cuda.manual_seed(params["TRAIN"]["seed"])
np.random.seed(params["TRAIN"]["seed"])
random.seed(params["TRAIN"]["seed"])
# Setup device
params["gpu_ids"] = [0]
params["output"] = params["DIR"]
params["dataset"] = ""
params["checkname"] = params['MODEL']['arch']+'_'+params['MODEL']['expert']+ params['MODEL']['gate'] + str(params["DATASET"]["img_height"]) + "_" + str(params["DATASET"]["img_width"])
params["start_time"] = datetime.now().strftime("%Y%m%d_%H%M%S")
## initialize wandb
wandb.init(project="adv_attacks_on_moe", reinit=True)
wandb.config = params
wandb.run.name = params["start_time"] + "_" + params["checkname"]
wandb.run.save()
trainer = Trainer(params)
print('Starting Epoch: {}'.format(str(params["TRAIN"]["start_epoch"])))
print('Total Epoches: {}'.format(str(params["TRAIN"]["num_epochs"])))
early_stopping = EarlyStopping(params["TRAIN"]["patience"])
for epoch in range(params["TRAIN"]["start_epoch"], params["TRAIN"]["num_epochs"]):
trainer.training(epoch)
trainer.scheduler.step(epoch)
if epoch % params["TRAIN"]["eval_interval"] == (params["TRAIN"]["eval_interval"] - 1):
test_loss = trainer.validation(epoch)
early_stopping(test_loss)
if early_stopping.early_stop:
break
except yaml.YAMLError as exc:
print(exc)