-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
204 lines (174 loc) · 6.6 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
import torch
from indent import *
from torch.nn import functional, Module
import numpy as np
import tensorboardX
import torchvision
run_count = 0
opt_mapping = {
'Adam': torch.optim.Adam,
'SGD': torch.optim.SGD,
'RMSprop': torch.optim.RMSprop
}
@indenting
def train(net, generator_factory, max_batches, *,
optimizer='Adam', lr=0.01,
log_dir='runs', title=None, batch_size=64,
hyper_net=None, params=None, flatten=True):
global run_count
if params is None:
# if a hypernetwork was provided, we don't directly train via the network parameters,
# we train via the hypernetwork parameters (which produce the network parameters)
if hyper_net:
params = hyper_net.parameters()
else:
params = net.parameters()
params = list(params)
weight_shapes = [tuple(p.shape) for p in params]
weight_params = sum(np.prod(shape) for shape in weight_shapes)
title_str = f'Training "{title}"' if title else 'Training'
shape_str = ' '.join('×'.join(map(str, x)) for x in weight_shapes)
print(f"{title_str} (arrays: {len(params)}, params: {weight_params}, shapes: {shape_str})")
if log_dir:
if title is None:
run_count += 1
title = f"{run_count:03d}"
writer = tensorboardX.SummaryWriter(log_dir + '/' + str(title), flush_secs=2)
else:
writer = None
opt = opt_mapping[optimizer](params, lr=lr)
time = 0
accuracy_generator = generator_factory(batch_size, is_training=True)
training_generator = generator_factory(batch_size, is_training=False)
running_loss = None
loss_history = []
acc_history = []
# main training loop
for inputs, labels, *rest in training_generator:
time += 1
opt.zero_grad()
if flatten:
inputs = torch.flatten(inputs, start_dim=1)
# if we have a hyper network, we should use its .forward method to
# derive the parameters for our network
if hyper_net:
net.zero_grad()
hyper_net.zero_grad()
hyper_net.forward()
# apply the net to the input batch
res = net(inputs, *rest)
if not isinstance(res, tuple):
res = res, 0
# the second returned value should be additional losses (if any)
labels_prime, extra_loss = res
# calculate the loss
loss = functional.cross_entropy(labels_prime, labels)
loss += extra_loss
# obtain gradients of the loss
loss.backward()
# if we have a hypernetwork, we need to propogate those gradients
# from the network back into the hypernetwork
if hyper_net:
hyper_net.backward()
# do one step of optimization
opt.step()
# report losses, etc.
loss = loss.item()
loss_history.append(loss)
if running_loss is None:
running_loss = loss
else:
running_loss = 0.95 * running_loss + 0.05 * loss
if time % 10 == 0 and writer:
writer.add_scalar("loss", running_loss, time)
if time % 1000 == 0:
if time % 2500 == 0:
acc = test_accuracy(net, accuracy_generator, flatten=flatten)
acc_history.append((time, acc))
if writer: writer.add_scalar("accuracy", acc, time)
print(f"{time:>5d}\t{running_loss:.3f}\t{acc:.3f}")
else:
print(f"{time:>5d}\t{running_loss:.3f}")
# stop training when we hit max_batches
if time > max_batches:
break
# report final test accuracy
acc = test_accuracy(net, accuracy_generator, max_batches=10000, flatten=flatten)
if writer:
writer.add_scalar("accuracy", acc, time)
writer.close()
print(f"Done training, final accuracy = {acc:.3f}")
history = {'loss': loss_history, 'accuracy': acc_history}
# return a bunch of statistics about the training run
return {'loss': running_loss, # final loss
'accuracy': acc, # final test accuracy
'history': history, # dictionary containing history of losses and test accuracies over time
'weight_shapes': weight_shapes, # list of shapes of trained arrays
'weight_params': weight_params, # list of names of trained arrays
'batch_size': batch_size, # batch size
'batches': max_batches} # number of batches to train for
def test_accuracy(net, generator, max_batches=5000, flatten=True):
total = correct_total = 0
for img, label, *rest in generator:
if flatten:
img = torch.flatten(img, start_dim=1)
label_prime = net(img, *rest).argmax(1)
correct = sum(label == label_prime).item()
correct_total += correct
total += img.shape[0]
if total > max_batches:
break
return correct_total / total
'''
def visualize_weights(layer, n_colors=1):
with torch.no_grad():
if isinstance(layer, torch.nn.Linear):
weight = layer.weight
else:
weight, _ = layer.calculate_weight()
# ^ for XOXLinear and friends
output_size, input_size = weight.shape
input_height = input_width = int(np.sqrt(input_size))
size = (output_size, n_colors, input_height, input_width)
weight_reshaped = weight.reshape(size).clone()
img = batched_to_flat_image(weight_reshaped)
return img
'''
def batched_to_flat_image(t):
shape = t.shape
n = shape[0]
rank = len(shape)
red_blue = True
if rank == 2:
w = shape[1]
if w > 8:
h = np.ceil(np.sqrt(w))
w = w // h
else:
h = w
w = 1
shape = [n, 1, h, w]
elif rank == 3:
shape = [n, 1, shape[1], shape[2]]
elif rank == 4:
shape = list(shape)
if shape[1] > 1:
red_blue = False
t = t.view(*shape)
t_min = t.min()
t_max = t.max()
if red_blue and t_min < 0 < t_max:
#sorted, _ = t.flatten().sort()
#n = sorted.numel()
#t_min = sorted[int(n / 5)]
#t_max = sorted[int(n * 4 / 5)]
scale = max(-t_min, t_max)
# for positive, shift the green and blue down
# for negative, shift the red and green down
scaled = t / scale
r = 1 + torch.clamp(scaled, -1, 0)
g = 1 - abs(scaled)
b = 1 - torch.clamp(scaled, 0, 1)
t = torch.cat([r, g, b], dim=1)
grid = torchvision.utils.make_grid(t, 10, normalize=(not red_blue), padding=1)
return grid