-
Notifications
You must be signed in to change notification settings - Fork 9
/
train.py
257 lines (216 loc) · 10.3 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
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
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
import os
import pandas as pd
import numpy as np
from tqdm import tqdm
import albumentations as A
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.transforms as tfms
from torch.utils.data import DataLoader, Dataset
from sklearn.model_selection import StratifiedKFold, train_test_split
from seed import seed_everything
from model import model_define
from data import AnimalKeypoint
from eval import calc_dists, dist_acc, accuracy
from inference import get_final_preds, get_max_preds
from config import SingleModelConfig
from loss import JointsRMSELoss, OffsetMSELoss, OffsetL1Loss, HeatmapMSELoss, HeatmapOHKMMSELoss
def calc_coord_loss(pred, gt):
batch_size = gt.size(0)
valid_mask = gt[:, :, -1].view(batch_size, -1, 1)
gt = gt[:, :, :2]
return torch.mean(torch.sum(torch.abs(pred-gt) * valid_mask, dim=-1))
def train(cfg, meta_info_dir, train_img_path, train_tfms=None, valid_tfms=None):
# for reporduction
seed = cfg.seed
torch.cuda.empty_cache()
seed_everything(2021)
# device type
device = 'cuda' if torch.cuda.is_available() else 'cpu'
# define model
if cfg.target_type=='offset':
pass
elif cfg.target_type=='gaussian':
yaml_name = "./config/heatmap_train.yaml"
yaml_path = os.path.join(cfg.main_dir, yaml_name)
model = model_define(yaml_path, cfg.init_training)
model = model.to(device)
# define criterions
if cfg.target_type == "offset":
main_criterion = OffsetMSELoss(True)
elif cfg.target_type == "gaussian":
if cfg.loss_type == "MSE":
main_criterion = HeatmapMSELoss(True)
elif cfg.loss_type == "OHKMMSE":
main_criterion = HeatmapOHKMMSELoss(True)
rmse_criterion = JointsRMSELoss()
# define optimizer and scheduler
optimizer = torch.optim.Adam(model.parameters(), lr=cfg.lr)
total_df = pd.read_csv(meta_info_dir)
train_df, valid_df = train_test_split(total_df.iloc[:, :], test_size=cfg.test_ratio, random_state=seed)
train_ds = AnimalKeypoint(cfg, train_img_path, train_df, train_tfms, mode='train')
valid_ds = AnimalKeypoint(cfg, train_img_path, valid_df, valid_tfms, mode='valid')
train_dl = DataLoader(train_ds, batch_size=cfg.batch_size, shuffle=True)
valid_dl = DataLoader(valid_ds, batch_size=cfg.batch_size, shuffle=False)
print("Train Transformation:\n", train_tfms, "\n")
print("Valid Transformation:\n", valid_tfms, "\n")
best_loss = float('INF')
for epoch in range(cfg.epochs):
################
# Train #
################
with tqdm(train_dl, total=train_dl.__len__(), unit="batch") as train_bar:
train_acc_list = []
train_rmse_list = []
train_heatmap_list = []
train_coord_list = []
train_offset_list = []
train_total_list = []
for sample in train_bar:
train_bar.set_description(f"Train Epoch {epoch+1}")
optimizer.zero_grad()
images, targ_coords = sample['image'].to(device), sample['keypoints'].to(device)
target, target_weight = sample['target'].to(device), sample['target_weight'].to(device)
model.train()
with torch.set_grad_enabled(True):
preds = model(images)
if cfg.target_type == "offset":
loss_hm, loss_os = main_criterion(preds, target, target_weight)
loss = loss_hm + loss_os
elif cfg.target_type == "gaussian":
loss = main_criterion(preds, target, target_weight)
if cfg.target_type=="offset":
heatmap_height = preds.shape[2]
heatmap_width = preds.shape[3]
pred_coords = get_final_preds(cfg, preds.detach().cpu().numpy())
elif cfg.target_type=='gaussian':
heatmap_height = preds.shape[2]
heatmap_width = preds.shape[3]
pred_coords, _ = get_max_preds(preds.detach().cpu().numpy())
pred_coords[:, :, 0] = pred_coords[:, :, 0] / (heatmap_width - 1.0) * (4 * heatmap_width - 1.0)
pred_coords[:, :, 1] = pred_coords[:, :, 1] / (heatmap_height - 1.0) * (4 * heatmap_height - 1.0)
pred_coords = torch.tensor(pred_coords).float().to(device)
coord_loss = calc_coord_loss(pred_coords, targ_coords)
rmse_loss = rmse_criterion(pred_coords, targ_coords)
_, avg_acc, cnt, pred = accuracy(preds.detach().cpu().numpy()[:, ::3, :, :],
target.detach().cpu().numpy()[:, ::3, :, :])
loss.backward()
optimizer.step()
if cfg.target_type == "offset":
train_heatmap_list.append(loss_hm.item())
train_offset_list.append(loss_os.item())
train_rmse_list.append(rmse_loss.item())
train_total_list.append(loss.item())
train_coord_list.append(coord_loss.item())
train_acc_list.append(avg_acc)
train_acc = np.mean(train_acc_list)
train_rmse = np.mean(train_rmse_list)
train_coord = np.mean(train_coord_list)
train_total = np.mean(train_total_list)
if cfg.target_type == "offset":
train_offset = np.mean(train_offset_list)
train_heatmap = np.mean(train_heatmap_list)
train_bar.set_postfix(heatmap_loss = train_heatmap,
coord_loss = train_coord,
offset_loss = train_offset,
rmse_loss = train_rmse,
total_loss = train_total,
train_acc = train_acc)
else:
train_bar.set_postfix(coord_loss = train_coord,
rmse_loss = train_rmse,
total_loss = train_total,
train_acc = train_acc)
################
# Valid #
################
with tqdm(valid_dl, total=valid_dl.__len__(), unit="batch") as valid_bar:
valid_acc_list = []
valid_rmse_list = []
valid_heatmap_list = []
valid_coord_list = []
valid_offset_list = []
valid_total_list = []
for sample in valid_bar:
valid_bar.set_description(f"Valid Epoch {epoch+1}")
images, targ_coords = sample['image'].to(device), sample['keypoints'].to(device)
target, target_weight = sample['target'].to(device), sample['target_weight'].to(device)
model.eval()
with torch.no_grad():
preds = model(images)
if cfg.target_type == "offset":
loss_hm, loss_os = main_criterion(preds, target, target_weight)
loss = loss_hm + loss_os
elif cfg.target_type == "gaussian":
loss = main_criterion(preds, target, target_weight)
pred_coords = get_final_preds(cfg, preds.detach().cpu().numpy())
pred_coords = torch.tensor(pred_coords).float().to(device)
coord_loss = calc_coord_loss(pred_coords, targ_coords)
rmse_loss = rmse_criterion(pred_coords, targ_coords)
_, avg_acc, cnt, pred = accuracy(preds.detach().cpu().numpy()[:, ::3, :, :],
target.detach().cpu().numpy()[:, ::3, :, :])
if cfg.target_type == "offset":
valid_heatmap_list.append(loss_hm.item())
valid_offset_list.append(loss_os.item())
valid_rmse_list.append(rmse_loss.item())
valid_total_list.append(loss.item())
valid_coord_list.append(coord_loss.item())
valid_acc_list.append(avg_acc)
valid_acc = np.mean(valid_acc_list)
valid_rmse = np.mean(valid_rmse_list)
valid_coord = np.mean(valid_coord_list)
valid_total = np.mean(valid_total_list)
if cfg.target_type == "offset":
valid_offset = np.mean(valid_offset_list)
valid_heatmap = np.mean(valid_heatmap_list)
valid_bar.set_postfix(heatmap_loss = valid_heatmap,
coord_loss = valid_coord,
offset_loss = valid_offset,
rmse_loss = valid_rmse,
total_loss = valid_total,
valid_acc = valid_acc)
else:
valid_bar.set_postfix(coord_loss = valid_coord,
rmse_loss = valid_rmse,
total_loss = valid_total,
valid_acc = valid_acc)
if best_loss > valid_total:
best_model = model
save_dir = cfg.save_folder
save_name = f'best_model_{valid_total}.pth'
torch.save(model.state_dict(), os.path.join(save_dir, save_name))
print(f"Valid Loss: {valid_total:.8f}\nBest Model saved.")
best_loss = valid_total
return best_model
def main():
meta_info_dir = '../data/annotations_1.csv'
train_img_path = '../images/images_1'
train_tfms = A.Compose([
A.OneOf([
A.ChannelShuffle(p=1.0),
A.HueSaturationValue(p=1.0),
A.RGBShift(p=1.0),
], p=0.5),
A.RandomBrightnessContrast(p=0.6),
A.RandomContrast(p=0.6),
A.RandomGamma(p=0.6),
A.CLAHE(p=0.5),
A.Normalize(p=1.0),
])
valid_tfms = A.Normalize(p=1.0)
cfg = SingleModelConfig(
epochs=30,
input_size=[640, 480],
learning_rate=5e-4,
sigma=3.0,
batch_size=4,
shift = True,
init_training=True,
loss_type = "MSE",
target_type = "gaussian",
save_folder='./weight'
)
best_model = train(cfg, meta_info_dir, train_img_path, train_tfms=train_tfms, valid_tfms=valid_tfms)
if __name__=="__main__":
main()