-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
179 lines (145 loc) · 6.38 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
import torch
import matplotlib.pyplot as plt
import numpy as np
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from tqdm import trange
from utils.ops import build_optimizer, build_lr_scheduler, build_loss
from utils.vis_utils import plot_loss
def train(net, train_data, time_stamp, args):
"""
Mini-batch train NN on provided data set.
Args:
net: initialized neural network
train_data: dataset containing training inputs and targets
time_stamp: time stamp for documentation purposes
args: arguments
Returns:
net: trained neural network
"""
# Load data
trainloader = DataLoader(train_data, batch_size=args.batch_size, shuffle=True, num_workers=0, drop_last=True)
# Construct loss function, optimization method, and learning rate scheduler
criterion = build_loss(args)
optimizer = build_optimizer(net, args)
scheduler = build_lr_scheduler(optimizer, args)
# Initialize important parameters
num_train = len(trainloader)
num_epochs = args.num_epochs
max_norm = args.max_norm
train_loss = np.zeros(num_epochs)
# Begin training
for epoch in trange(num_epochs):
running_loss = 0.0
# Mini-batch training step
net.train()
for X_train_batch, y_train_batch in enumerate(trainloader):
for param in net.parameters(): # clear out gradients from previous optimization step
param.grad = None
outputs = net(X_train_batch)
loss = criterion(outputs, y_train_batch)
loss.backward() # calculate gradients
nn.utils.clip_grad_norm_(net.parameters(), max_norm) # clip gradient norm
optimizer.step() # update weights
running_loss += loss.item()
# Print real-time results
if (epoch + 1) % args.print_every == 0:
print('Epoch: {:>6d}/{:d}, {}: {:.10f}'.format(
epoch + 1, num_epochs, criterion.__class__.__name__, running_loss / num_train,
),
)
train_loss[epoch] = running_loss / num_train
# Update learning rate if necessary
scheduler.step(loss)
# Plot learning curve if specified
if args.plot_loss.lower() in ['true', 't', 'yes', 'y']:
figdir = ''.join(['figures/training-curves/', net.__class__.__name__ + time_stamp, '.png'])
fig = plot_loss(
train_loss=train_loss.flatten(),
title=net.__class__.__name__,
loss_fn=criterion.__class__.__name__,
)
plt.show()
fig.savefig(figdir, dpi=300, bbox_inches='tight')
# Save trained network if specified
if args.save_net.lower() in ['true', 't', 'yes', 'y']:
print('Saving the trained model ...')
save_path = ''.join(['saved-networks/', net.__class__.__name__ + time_stamp, '.pth'])
torch.save(net.state_dict(), save_path)
print('Saved model')
return net
def train_valid(net, train_data, valid_data, time_stamp, args):
"""
Mini-batch train and validate NN on provided data set.
Args:
net: neural network
train_data: data set containing training inputs and targets
valid_data: data set containing validation inputs and targets
time_stamp: time stamp for documentation purposes
args: arguments
"""
# Load data
trainloader = DataLoader(train_data, batch_size=args.batch_size, shuffle=True, num_workers=0, drop_last=True)
validloader = DataLoader(valid_data, batch_size=args.batch_size, shuffle=True, num_workers=0, drop_last=True)
# Construct loss function, optimization method, and learning rate scheduler
criterion = build_loss(args)
optimizer = build_optimizer(net, args)
scheduler = build_lr_scheduler(optimizer, args)
# Initialize important parameters
num_train = len(trainloader)
num_valid = len(validloader)
num_epochs = args.num_epochs
max_norm = args.max_norm
train_loss = np.zeros(num_epochs)
valid_loss = np.zeros(num_epochs)
# Begin training
for epoch in trange(num_epochs):
t_loss = 0.0
v_loss = 0.0
# Mini-batch training step
net.train()
for X_train_batch, y_train_batch in trainloader:
for param in net.parameters(): # clear out gradients from previous optimization step
param.grad = None
outputs = net(X_train_batch) # make predictions
loss = criterion(outputs, y_train_batch) # calculate loss
loss.backward() # calculate gradients
nn.utils.clip_grad_norm_(net.parameters(), max_norm) # clip gradient norm
optimizer.step() # update weights
t_loss += loss.item()
# Validation step
net.eval()
with torch.no_grad():
for X_valid_batch, y_valid_batch in validloader:
outputs = net(X_valid_batch)
loss = criterion(outputs, y_valid_batch)
v_loss += loss.item()
# Print real-time results
if (epoch + 1) % args.print_every == 0:
print('Epoch: {:>6d}/{:d}, Train loss: {:.10f}, Valid loss: {:.10f}'.format(
epoch + 1, num_epochs, t_loss / num_train, v_loss / num_valid,
),
)
train_loss[epoch] = t_loss / num_train
valid_loss[epoch] = v_loss / num_valid
# Update learning rate if necessary
scheduler.step(v_loss) # LR scheduler step can be dependent upon either training or validation loss
# Plot learning curves if specified
if args.plot_loss.lower() in ['true', 't', 'yes', 'y']:
figdir = ''.join(['figures/training-curves/', net.__class__.__name__ + time_stamp, '.png'])
fig = plot_loss(
train_loss=train_loss.flatten(),
valid_loss=valid_loss.flatten(),
title=net.__class__.__name__,
loss_fn=criterion.__class__.__name__,
)
plt.show()
fig.savefig(figdir, dpi=300, bbox_inches='tight')
# Save trained network if specified
if args.save_net.lower() in ['true', 't', 'yes', 'y']:
print('Saving the trained model ...')
save_path = ''.join(['saved-networks/', net.__class__.__name__ + time_stamp, '.pth'])
torch.save(net.state_dict(), save_path)
print('Saved model')
return net