-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_KD.py
198 lines (171 loc) · 8.48 KB
/
train_KD.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
import numpy as np
import os
import sys
import argparse
import matplotlib.pyplot as plt
import torch
from torch import nn
import torchvision.transforms as transforms
from losses import FocalLoss, mIoULoss
from model import UNet, Custom_Slim_UNet
from dataloader import segDataset
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
# T = 5 is best
T = 5
# alpha = 0.8 is best, 0.9 is too much, less than 0.8 decreases accuracy
alpha = 0.8
def pixel_wise_loss(student_output, teacher_output):
_,C,W,H = student_output.shape
# what would happen if we use softmax?
# Softmax didn't work
pred_T = torch.sigmoid(teacher_output/T)
pred_S = torch.sigmoid(student_output/T).log()
#criterion = torch.nn.KLDivLoss(reduction = 'batchmean')
#KLDloss = - criterion(pred_S, pred_T)
#TODO: map this to KLDL
#KDloss = - sum(p * log (p/q)) ---> refer notes page 15 - 16
#Pixelwise loss = sum(-p*logq)
#KLDiv = relative entropy
pixelwise_loss = (- pred_T * pred_S)
# Check to see if it is needed
return torch.sum(pixelwise_loss) / (W*H*10)
def loss_fn(student_output, teacher_output, gt , criterion):
'''student_output = student_output.round()
student_output[student_output<0] = 0
gt = torch.clamp(gt, min = 0, max = 1)
teacher_output = torch.clamp(teacher_output, min = 0, max = 1)'''
student_output = student_output.clamp(min = 0, max = 1)
teacher_output = teacher_output.clamp(min = 0, max = 1)
student_loss = criterion(student_output, gt)
kd_loss = pixel_wise_loss(student_output, teacher_output)
#not sure about using T, also check KLD
loss = (student_loss*(1-alpha) + (kd_loss)*(alpha)) # as per structured KD paper
return loss
def get_args():
parser = argparse.ArgumentParser()
parser.add_argument('--data', type=str, required=True, help='path to your dataset')
parser.add_argument('--test', type=str, help='path to your test dataset')
parser.add_argument('--meta', type=str, required=True, help='path to your metadata')
parser.add_argument('--name', type=str, default="unet", help='name to be appended to checkpoints')
parser.add_argument('--num_epochs', type=int, default=100, help='dnumber of epochs')
parser.add_argument('--batch', type=int, default=1, help='batch size')
parser.add_argument('--loss', type=str, default='focalloss', help='focalloss | iouloss | crossentropy')
return parser.parse_args()
def acc(label, predicted):
seg_acc = (y.cpu() == torch.argmax(pred_mask, axis=1).cpu()).sum() / torch.numel(y.cpu())
return seg_acc
if __name__ == '__main__':
args = get_args()
N_EPOCHS = args.num_epochs
BACH_SIZE = args.batch
color_shift = transforms.ColorJitter(.1,.1,.1,.1)
blurriness = transforms.GaussianBlur(3, sigma=(0.1, 2.0))
t = transforms.Compose([color_shift, blurriness])
dataset = segDataset(args.data, args.meta, training = True, transform= t)
n_classes = len(dataset.bin_classes)+1
print('Number of data : '+ str(len(dataset)))
if not args.test :
dataset = segDataset(args.data, args.meta, training = True, transform= t)
n_classes = len(dataset.bin_classes)+1
print('Number of data : '+ str(len(dataset)))
test_num = int(0.1 * len(dataset))
print(f'Test data : {test_num}')
print(f"Number of classes : {n_classes}")
train_dataset, test_dataset = torch.utils.data.random_split(dataset, [len(dataset)-test_num, test_num], generator=torch.Generator().manual_seed(101))
N_DATA, N_TEST = len(train_dataset), len(test_dataset)
train_dataloader = torch.utils.data.DataLoader(train_dataset, batch_size=BACH_SIZE, shuffle=True, num_workers=2)
test_dataloader = torch.utils.data.DataLoader(test_dataset, batch_size=BACH_SIZE, shuffle=False, num_workers=1)
else :
dataset = segDataset(args.data, args.meta, training = True, transform= t)
dataset2 = segDataset(args.test, args.meta, training = False, transform= t)
n_classes = len(dataset.bin_classes)+1
print('Number of train data : '+ str(len(dataset)))
test_num = len(dataset2)
print(f'Test data : {test_num}')
print(f"Number of classes : {n_classes}")
N_DATA, N_TEST = len(dataset), len(dataset2)
train_dataloader = torch.utils.data.DataLoader(dataset, batch_size=BACH_SIZE, shuffle=True, num_workers=2)
test_dataloader = torch.utils.data.DataLoader(dataset2, batch_size=BACH_SIZE, shuffle=False, num_workers=1)
if args.loss == 'focalloss':
criterion = FocalLoss(gamma=3/4).to(device)
elif args.loss == 'iouloss':
criterion = mIoULoss(n_classes=n_classes).to(device)
elif args.loss == 'crossentropy':
criterion = nn.CrossEntropyLoss().to(device)
else:
print('Loss function not found!')
model = UNet(n_channels=3, n_classes=n_classes, bilinear=True, channel_depth=8).to(device)
teacher = UNet(n_channels=3, n_classes=n_classes, bilinear=True).to(device)
teacher.load_state_dict(torch.load('C:\\projects\\LandUse\\Semantic-Segmentation-UNet-Federated-master\\saved_models\\BigCS_epoch_1_1.26202.pt'))
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.5)
min_loss = torch.tensor(float('inf'))
os.makedirs('./saved_models_small', exist_ok=True)
plot_losses = []
scheduler_counter = 0
for epoch in range(N_EPOCHS):
# training
model.train()
# teacher.eval()
loss_list = []
acc_list = []
for batch_i, (x, y) in enumerate(train_dataloader):
with torch.no_grad():
teach_mask = teacher(x.to(device))
pred_mask = model(x.to(device))
loss = loss_fn(pred_mask, teach_mask, y.to(device), criterion)
# loss = criterion(pred_mask.to(device),y.to(device))
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss_list.append(loss.cpu().detach().numpy())
acc_list.append(acc(y,pred_mask).numpy())
sys.stdout.write(
"\r[Epoch %d/%d] [Batch %d/%d] [Loss: %f (%f)]"
% (
epoch,
N_EPOCHS,
batch_i,
len(train_dataloader),
loss.cpu().detach().numpy(),
np.mean(loss_list),
)
)
scheduler_counter += 1
# testing
model.eval()
val_loss_list = []
val_acc_list = []
for batch_i, (x, y) in enumerate(test_dataloader):
with torch.no_grad():
pred_mask = model(x.to(device))
val_loss = criterion(pred_mask, y.to(device))
val_loss_list.append(val_loss.cpu().detach().numpy())
val_acc_list.append(acc(y,pred_mask).numpy())
print(' epoch {} - loss : {:.5f} - acc : {:.2f} - val loss : {:.5f} - val acc : {:.2f}'.format(epoch,
np.mean(loss_list),
np.mean(acc_list),
np.mean(val_loss_list),
np.mean(val_acc_list)))
plot_losses.append([epoch, np.mean(loss_list), np.mean(val_loss_list)])
compare_loss = np.mean(val_loss_list)
is_best = compare_loss < min_loss
if is_best == True:
scheduler_counter = 0
min_loss = min(compare_loss, min_loss)
torch.save(model.state_dict(), './saved_models/{}_epoch_{}_{:.5f}.pt'.format(args.name,epoch,np.mean(val_loss_list)))
if scheduler_counter > 5:
lr_scheduler.step()
print(f"lowering learning rate to {optimizer.param_groups[0]['lr']}")
scheduler_counter = 0
# plot loss
plot_losses = np.array(plot_losses)
plt.figure(figsize=(12,8))
plt.plot(plot_losses[:,0], plot_losses[:,1], color='b', linewidth=4)
plt.plot(plot_losses[:,0], plot_losses[:,2], color='r', linewidth=4)
plt.title(args.loss, fontsize=20)
plt.xlabel('epoch',fontsize=20)
plt.ylabel('loss',fontsize=20)
plt.grid()
plt.legend(['training', 'validation']) # using a named size
plt.savefig('loss_plots.png')