-
Notifications
You must be signed in to change notification settings - Fork 0
/
baseline_model.py
236 lines (211 loc) · 10.7 KB
/
baseline_model.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from tqdm import tqdm
import cv2
from PIL import Image
import json
import numpy as np
import pandas as pd
from torch.utils.tensorboard import SummaryWriter
import os, sys
sys.path.append(os.path.dirname(os.path.abspath(__file__)))
import os.path as osp
from copy import deepcopy
import importlib
from metrics.all_metric import SegMetricAll
from losses.softce_dice_loss import SoftCrossEntropy_DiceLoss
from utils.net_utils import load_network
from utils.meter_utils import AverageMeter
from utils.color_utils import generate_random_colormap
from segmentation_models_pytorch.losses import DiceLoss, SoftCrossEntropyLoss
class BaselineModel:
"""
Baseline Model:
input augmented image, output prob map for each class
standard training strategy, from scratch
loss conducted on prob/logit maps, celoss/iouloss/lovaszloss/diceloss etc.
"""
def __init__(self, opt) -> None:
self.opt = opt
self.opt_dataset = opt['datasets']
self.device = opt['device']
def prepare_training(self):
self.opt_train = self.opt['train']
self.max_perf = 0.0
os.makedirs(self.opt['log_dir'], exist_ok=True)
log_path = osp.join(self.opt['log_dir'], self.opt['exp_name'])
self.writer = SummaryWriter(log_path)
# prepare dataloader
self.train_loader, self.val_loader = self.get_trainval_dataloaders(self.opt_dataset)
# prepare network for training
opt_model_arch = deepcopy(self.opt_train['model_arch'])
arch_type = opt_model_arch.pop('type')
load_path = opt_model_arch.pop('load_path') if 'load_path' in opt_model_arch else None
self.network = self.get_network(arch_type, load_path, **opt_model_arch).to(self.device)
if self.opt['multi_gpu']:
self.network = nn.DataParallel(self.network)
self.network.train()
# prepare optimizer and corresponding net params
opt_optim = deepcopy(self.opt_train['optimizer'])
optim_type = opt_optim.pop('type')
optim_params = []
for k, v in self.network.named_parameters():
if v.requires_grad:
optim_params.append(v)
else:
print(f'Params {k} will not be optimized.')
self.optimizer = self.get_optimizer(optim_type, optim_params, **opt_optim)
# prepare lr scheduler
opt_scheduler = deepcopy(self.opt_train['scheduler'])
scheduler_type = opt_scheduler.pop('type')
self.scheduler = self.get_scheduler(scheduler_type, self.optimizer, **opt_scheduler)
# prepare criterion
self.criterion = self.get_criterion(self.opt_train['criterion'])
# prepare metric for evaluation
opt_metric = deepcopy(self.opt_train['metric'])
metric_type = opt_metric.pop('type')
self.metric = self.get_metric(metric_type, num_classes=self.opt_train['model_arch']['classes'])
def get_trainval_dataloaders(self, opt_dataset):
opt_train = deepcopy(opt_dataset['train_dataset'])
opt_val = deepcopy(opt_dataset['val_dataset'])
train_type = opt_train.pop('type')
val_type = opt_val.pop('type')
opt_train['phase'] = 'train'
opt_val['phase'] = 'valid'
train_loader = getattr(importlib.import_module('data'), train_type)(opt_train)
val_loader = getattr(importlib.import_module('data'), val_type)(opt_val)
return train_loader, val_loader
def get_network(self, arch_type, load_path, **kwargs):
# get torch original lr_scheduler based on their names
network = getattr(importlib.import_module('archs'), arch_type)(**kwargs)
if load_path is not None:
network = load_network(network, load_path)
print(f"[MODEL] Locally load pretrained network from {load_path}")
else:
print(f"[MODEL] Locally network train from scratch")
return network
def get_optimizer(self, optim_type, params, lr, **kwargs):
# get torch original optimizers based on their names
# e.g. Adam, AdamW, SGD ...
optimizer = getattr(importlib.import_module('torch.optim'), optim_type)(params, lr, **kwargs)
return optimizer
def get_scheduler(self, scheduler_type, optimizer, **kwargs):
# get torch original lr_scheduler based on their names
# e.g. CosineAnnealingWarmRestarts, MultiStepLR, ExponentialLR ...
sch_cls = getattr(importlib.import_module('torch.optim.lr_scheduler'), scheduler_type)
scheduler = sch_cls(optimizer, **kwargs)
return scheduler
def get_criterion(self, opt_criterion):
# TODO: add iou-based criterions
crit_type = opt_criterion.pop('type')
if crit_type.lower() == 'celoss':
loss_func = nn.CrossEntropyLoss(**opt_criterion)
elif crit_type.lower() == 'softceloss':
loss_func = SoftCrossEntropyLoss(**opt_criterion)
elif crit_type.lower() == 'diceloss':
loss_func = DiceLoss(**opt_criterion)
elif crit_type.lower() == 'softce_diceloss':
loss_func = SoftCrossEntropy_DiceLoss(**opt_criterion)
else:
raise NotImplementedError(f'loss func type {crit_type} is currently not supported')
return loss_func
def get_metric(self, metric_type, **kwargs):
metric = SegMetricAll(metric_type=metric_type, **kwargs)
return metric
# train one epoch
def train_epoch(self, epoch_id):
self.network.train()
batch_size = self.opt_dataset['train_dataset']['batch_size']
loss_tot_avm = AverageMeter()
pbar = tqdm(enumerate(self.train_loader), total=len(self.train_loader))
iters_per_epoch = len(self.train_loader)
# for iter_id, batch in enumerate(self.train_loader):
for iter_idx, batch in pbar:
# load data mini-batch
img, label = batch['img'], batch['label']
img, label = img.to(self.device), label.to(self.device)
# start optimize
self.optimizer.zero_grad()
probs = self.network(img)
loss = self.criterion(probs, label)
loss.backward()
self.optimizer.step()
loss_tot_avm.update(loss.detach().item(), batch_size)
# print(f"Train_Loss: {loss_score.avg}, Epoch: {epoch_id} iter {iter_id}, LR: {self.optimizer.param_groups[0]['lr']}")
pbar.set_postfix(Loss=loss_tot_avm.avg, Epoch=epoch_id, LR=self.optimizer.param_groups[0]['lr'])
self.writer.add_scalar('loss/loss', loss_tot_avm.avg, epoch_id * iters_per_epoch + iter_idx)
self.writer.add_scalar('learning_rate', self.optimizer.param_groups[0]['lr'], epoch_id * iters_per_epoch + iter_idx)
self.scheduler.step()
# eval after one epoch
def eval_epoch(self, epoch_id):
self.network.eval()
self.metric.reset()
print(f"[MODEL] Begin evaluation ...")
with torch.no_grad():
pbar = tqdm(enumerate(self.val_loader), total=len(self.val_loader))
for iter_id, batch in pbar:
img, label = batch['img'], batch['label']
img, label = img.to(self.device), label.to(self.device)
probs = self.network(img)
pred = torch.max(probs, dim=1)[1]
self.metric.update(label, pred)
pbar.set_postfix(Idx=iter_id, NumSamples=self.metric.num_sample())
# finished all eval images, summarize
cur_perf_all = self.metric.calc()
cur_perf = cur_perf_all[self.opt_train['metric']['type']]
print("\n\t >>> [MODEL] Evaluate Summary:")
print(f" Epoch {epoch_id}, total {len(self.val_loader)} eval images, "
f" Metric Type: {self.opt_train['metric']}"
f" Eval Score: {cur_perf}")
self.writer.add_scalar(f'eval/{self.opt_train["metric"]}', cur_perf, epoch_id)
if cur_perf > self.max_perf:
print(f"[MODEL] New best performance Epoch {epoch_id} Metric {cur_perf}, save and update")
self.save_model(epoch_id, self.opt_train['metric']['type'], cur_perf, copy_best=True)
self.max_perf = cur_perf
def save_model(self, epoch_id, val_metric_type, val_metric, copy_best=True):
save_dir = osp.join(self.opt['save_dir'], self.opt['exp_name'], 'ckpt')
os.makedirs(save_dir, exist_ok=True)
save_path = osp.join(save_dir, f'epoch{epoch_id:05}_{val_metric_type}{val_metric:.4f}.pth.tar')
torch.save(self.network.state_dict(), save_path)
if copy_best:
best_path = osp.join(save_dir, 'best.pth.tar')
torch.save(self.network.state_dict(), best_path)
return save_path
def inference(self):
iscolor = self.opt['infer']['output_color']
if iscolor:
SEG_COLORMAP = generate_random_colormap(self.opt['model_arch']['classes'])
opt_test = deepcopy(self.opt['datasets']['test_dataset'])
test_type = opt_test.pop('type')
self.test_loader = getattr(importlib.import_module('data'), test_type)(opt_test)
# prepare network for inference
opt_model_arch = deepcopy(self.opt['model_arch'])
arch_type = opt_model_arch.pop('type')
load_path = opt_model_arch.pop('load_path') if 'load_path' in opt_model_arch else None
self.network = self.get_network(arch_type, load_path, **opt_model_arch).to(self.device)
self.network.eval()
vis_dir = osp.join(self.opt['result_dir'], self.opt['exp_name'], 'visualization')
os.makedirs(vis_dir, exist_ok=True)
print(f"[MODEL] Begin inference ...")
# result_df = pd.DataFrame()
with torch.no_grad():
pbar = tqdm(enumerate(self.test_loader), total=len(self.test_loader))
for iter_id, batch in pbar:
img = batch['img'].to(self.device)
img_path = batch['img_path'][0]
ori_size_wh = batch['ori_size_wh']
ori_size_wh = (ori_size_wh[0].item(), ori_size_wh[1].item())
imname = osp.basename(img_path)
probs = F.softmax(self.network(img), dim=1)
pred_clsmap_ten = torch.argmax(probs, dim=1, keepdim=False)[0] # class map (h, w)
pred_clsmap = pred_clsmap_ten.detach().cpu().numpy().astype(np.uint8)
pred_clsmap = np.array(Image.fromarray(pred_clsmap).resize(ori_size_wh, resample=0)) # 0 for nearest
if iscolor:
pred_mask = SEG_COLORMAP[pred_clsmap]
else:
pred_mask = pred_clsmap
mask_img = Image.fromarray(pred_mask)
output_path = osp.join(vis_dir, imname.split('.')[0] + '.png')
mask_img.save(output_path)
pbar.set_postfix(Idx=iter_id)