-
Notifications
You must be signed in to change notification settings - Fork 4
/
train.py
197 lines (166 loc) · 7.92 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
import sys
import pickle
import numpy as np
import pandas as pd
from loss import MultiLabelNBLoss
from naive_bayes import NaiveBayes
from dataset import VOCDataset, collate_wrapper
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision import transforms
import torchvision.models as models
from eval import get_AP
import PIL
torch.multiprocessing.set_sharing_strategy('file_system')
directory = 'VOC2012'
use_cuda = 1
batch_size = 48
device = torch.device("cuda" if use_cuda else "cpu")
torch.manual_seed(0)
def print_nb_matrix(dataset, mat):
cols = ["x={}".format(key) for key in dataset.labels_dict.keys()]
rows = ["P({}|x)".format(key) for key in dataset.labels_dict.keys()]
mat = pd.DataFrame(mat, columns=cols, index=rows).round(5).T
print(mat)
def train(model, device, train_loader, optimizer, epoch, loss_function):
model.train()
losses = []
for idx, batch in enumerate(train_loader):
data = batch.image.to(device)
target = batch.labels.to(device)
optimizer.zero_grad()
output = model(data)
loss = loss_function(output, target)
loss.backward()
optimizer.step()
losses.append(loss.item())
print('Epoch: {}, Samples: {}/{}, Loss: {}'.format(epoch, idx*batch_size,
len(train_loader)*batch_size,
loss.item()))
train_loss = torch.mean(torch.tensor(losses))
print('\nEpoch: {}'.format(epoch))
print('Training set: Average loss: {:.4f}'.format(train_loss))
return train_loss
def validate(model, device, val_loader, loss_function):
model.eval()
val_loss = 0
with torch.no_grad():
for idx, batch in enumerate(val_loader):
data = batch.image.to(device)
target = batch.labels.to(device)
output = model(data)
batch_loss = loss_function(output, target)
val_loss += batch_loss.item()
pred = torch.sigmoid(output)
if idx == 0:
predictions = pred
targets = target
else:
predictions = torch.cat((predictions, pred))
targets = torch.cat((targets, target))
# divide by the number of batches of batch size
# get the average validation over all bins
val_loss /= len(val_loader)
print('Validation set: Average loss: {:.4f}'.format(val_loss))
print(' AP: {:.4f}'.format(
get_AP(predictions.reshape(-1, 20), targets.reshape(-1, 20))))
return val_loss, predictions, targets
def main(mode, num_epochs, num_workers, lr, sc, model_name=None):
tr = transforms.Compose([transforms.RandomResizedCrop(300),
transforms.ToTensor(),
transforms.Normalize([0.4589, 0.4355, 0.4032],[0.2239, 0.2186, 0.2206])])
augs = transforms.Compose([transforms.RandomResizedCrop(300),
transforms.RandomRotation(20),
transforms.ToTensor(),
transforms.Normalize([0.4589, 0.4355, 0.4032],[0.2239, 0.2186, 0.2206])])
# Get the NB matrix from the dataset,
# counting multiple instances of labels.
nb_dataset = VOCDataset(directory, 'train', transforms=tr, multi_instance=True)
nb = NaiveBayes(nb_dataset, 1)
mat = nb.get_nb_matrix()
print_nb_matrix(nb_dataset, mat)
mat = torch.Tensor(mat).to(device)
# Define the training dataset, removing
# multiple instances for the training problem.
train_set = VOCDataset(directory, 'train', transforms=augs, multi_instance=False)
train_loader = DataLoader(train_set, batch_size=batch_size, collate_fn=collate_wrapper, shuffle=True, num_workers=num_workers)
val_set = VOCDataset(directory, 'val', transforms=tr)
val_loader = DataLoader(val_set, batch_size=batch_size, collate_fn=collate_wrapper, shuffle=True, num_workers=num_workers)
model = models.resnet34(pretrained=True)
model.fc = nn.Linear(512, 20)
if model_name == None:
train_losses = []
val_losses = []
curr_epoch = 0
else:
model.load_state_dict(torch.load(model_name + '.pt'))
print('Loading history')
train_losses = np.load('train_history_{}_{}.npy'.format(mode, model_name)).tolist()
val_losses = np.load('val_history_{}_{}.npy'.format(mode, model_name)).tolist()
curr_epoch = int(model_name.split('_')[-2])
model.to(device)
print('Starting optimizer with LR={}'.format(lr))
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=0.9)
# ====================================== #
# Use either: #
# loss_function = nn.BCEWithLogitsLoss() #
# loss_function = MultiLabelNBLoss(mat) #
# ====================================== #
if mode == 'BCE':
classwise_frequencies = np.array(list(train_set.classes_count.values()))
minimum_frequency = np.min(classwise_frequencies)
loss_weights = minimum_frequency / classwise_frequencies
loss_weights = torch.Tensor(loss_weights).to(device)
loss_function = nn.BCEWithLogitsLoss(weight=loss_weights)
elif mode == 'NB':
loss_function = MultiLabelNBLoss(mat, scaling_c=sc)
try:
for epoch in range(1, num_epochs + 1):
train_loss = train(model, device, train_loader, optimizer, curr_epoch+1, loss_function)
val_loss, predictions, targets = validate(model, device, val_loader, loss_function)
print("Saving raw predictions for epoch {}...".format(curr_epoch+1))
with open("pred_{}_{}.pkl".format(mode, curr_epoch+1), 'wb') as f:
pred_targets = torch.cat((predictions.unsqueeze(0), targets.unsqueeze(0)))
pickle.dump(pred_targets, f)
if (len(val_losses) > 0) and (val_loss < min(val_losses)):
torch.save(model.state_dict(), "lr{}_sc{}_model_{}_{}_{:.4f}.pt".format(lr, sc, mode, curr_epoch+1, val_loss))
print("Saving model (epoch {}) with lowest validation loss: {}"
.format(epoch, val_loss))
train_losses.append(train_loss)
val_losses.append(val_loss)
torch.save(model.state_dict(), 'temp_model.pt')
curr_epoch += 1
model_save_name = "stop_lr{}_sc{}_model_{}_{}_{:.4f}.pt".format(lr, sc, mode, curr_epoch, val_losses[-1])
torch.save(model.state_dict(), model_save_name)
except KeyboardInterrupt:
model.load_state_dict(torch.load('temp_model.pt'))
model_save_name = "pause_lr{}_sc{}_model_{}_{}_{:.4f}.pt".format(lr, sc, mode, curr_epoch, val_losses[-1])
torch.save(model.state_dict(), model_save_name)
print("Saving model (epoch {}) with current validation loss: {}".format(curr_epoch, val_losses[-1]))
train_history = np.array(train_losses)
val_history = np.array(val_losses)
print('Saving history')
np.save("train_history_{}_{}".format(mode, model_save_name[5:-3]), train_history)
np.save("val_history_{}_{}".format(mode, model_save_name[5:-3]), val_history)
if __name__ == '__main__':
if len(sys.argv) > 1:
args = sys.argv[1:]
if len(args) == 5:
main(mode=args[0], num_epochs=int(args[1]), num_workers=int(args[2]),
lr=float(args[3]), sc=float(args[4]))
elif len(args) == 6:
main(mode=args[0], num_epochs=int(args[1]), num_workers=int(args[2]),
lr=float(args[3]), sc=float(args[4]), model_name=args[5])
else:
response = '''Wrong number of arguments, please enter the following arguments:
1. Mode ('BCE' or 'NB')
2. Max epochs (int)
3. Number of worker threads (int)
4. Learning rate (float)
5. Scaling constant (float)
6. Target model file name (optional)'''
print(response)
else:
main(mode='BCE', num_epochs=50, num_workers=16, lr=1e-3, sc=1e-3)