-
Notifications
You must be signed in to change notification settings - Fork 2
/
mnist_draw_4.py
250 lines (184 loc) · 8.64 KB
/
mnist_draw_4.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
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchbearer
import torchvision
from torchbearer import Trial, callbacks
from torchvision import transforms
import visualise
from memory import Memory
import tb_modules as tm
MU = torchbearer.state_key('mu')
LOGVAR = torchbearer.state_key('logvar')
STAGES = torchbearer.state_key('stages')
class Block(nn.Module):
def __init__(self, in_planes, out_planes, stride=1, padding=0):
super(Block, self).__init__()
self.conv = nn.Conv2d(in_planes, out_planes, kernel_size=3, padding=padding, stride=stride, bias=False)
self.bn = nn.BatchNorm2d(out_planes)
torch.nn.init.xavier_uniform_(self.conv.weight)
def forward(self, x):
out = F.relu(self.bn(self.conv(x)))
return out
class InverseBlock(nn.Module):
def __init__(self, in_planes, out_planes, stride=1, last=False, output_padding=0):
super(InverseBlock, self).__init__()
self.last = last
self.conv = nn.ConvTranspose2d(in_planes, out_planes, kernel_size=3, output_padding=output_padding, stride=stride, bias=False)
self.bn = nn.BatchNorm2d(out_planes)
torch.nn.init.xavier_uniform_(self.conv.weight)
def forward(self, x):
if not self.last:
out = F.relu(self.bn(self.conv(x)))
else:
out = self.bn(self.conv(x))
return out
class ContextNet(nn.Module):
def __init__(self):
super(ContextNet, self).__init__()
self.conv1 = Block(1, 64, stride=2)
self.conv2 = Block(64, 128, stride=2)
self.conv3 = Block(128, 256, stride=2)
def forward(self, x):
x = self.conv1(x)
x = self.conv2(x)
x = self.conv3(x)
x = x.view(x.size(0), -1)
return x
class GlimpseNet(nn.Module):
def __init__(self):
super(GlimpseNet, self).__init__()
self.conv1 = Block(1, 128)
def forward(self, x):
x = self.conv1(x)
x = x.view(x.size(0), -1)
return x
class GlimpseDecoder(nn.Module):
def __init__(self, h, w):
super(GlimpseDecoder, self).__init__()
self.h = h
self.w = w
self.conv1 = InverseBlock(128, 1, last=True)
def forward(self, x):
x = x.view(x.size(0), 128, self.h, self.w)
x = self.conv1(x)
return x
class MnistDraw(nn.Module):
def __init__(self, count, memory_size, output_stages=False):
super(MnistDraw, self).__init__()
self.output_stages = output_stages
self.memory = Memory(
output_inverse=True,
hidden_size=512,
memory_size=memory_size,
glimpse_size=4,
g_down=512,
c_down=1024,
context_net=ContextNet(),
glimpse_net=GlimpseNet()
)
self.decoder = GlimpseDecoder(2, 2)
self.count = count
self.qdown = nn.Linear(1024, memory_size)
self.soft = nn.LogSoftmax(dim=1)
self.drop = nn.Dropout(0.3)
self.mu = nn.Linear(memory_size, 4)
self.var = nn.Linear(memory_size, 4)
self.sup = nn.Linear(4, 512)
self.onehots = nn.Parameter(torch.eye(count), requires_grad=False)
if output_stages:
self.square = visualise.red_square(4, width=1).unsqueeze(0).cuda()
def sample(self, mu, logvar):
std = logvar.div(2).exp_()
eps = std.data.new(std.size()).normal_()
return mu + std * eps
def forward(self, x, state=None):
image = x
canvas = torch.zeros_like(x.data) - 6.0
x, context = self.memory.init(image)
c_data = context.data
query = F.relu6(self.qdown(c_data))
mu = []
var = []
stages = []
for i in range(self.count):
x, inverse = self.memory.glimpse(x, image)
out = self.memory(query)
o_mu = self.mu(out)
o_var = self.var(out)
mu.append(o_mu)
var.append(o_var)
out = self.sample(o_mu, o_var)
out = F.relu(self.sup(out))
out = self.decoder(out)
inverse = inverse.view(out.size(0), 2, 3)
grid = F.affine_grid(inverse, torch.Size((out.size(0), out.size(1), image.size(2), image.size(3))))
out = F.grid_sample(out, grid)
canvas += out
if self.output_stages:
square = self.square.clone().repeat(out.size(0), 1, 1, 1)
square = F.grid_sample(square, grid)
stage_image = canvas.data.clone().sigmoid().repeat(1, 3, 1, 1)
stage_image = stage_image + square
stage_image = stage_image.clamp(0, 1)
stages.append(stage_image.unsqueeze(1))
if state is not None:
state[torchbearer.Y_TRUE] = image
state[MU] = torch.cat(mu, dim=1)
state[LOGVAR] = torch.cat(var, dim=1)
if self.output_stages:
stages.append(image.clone().repeat(1, 3, 1, 1).unsqueeze(1))
state[STAGES] = torch.cat(stages, dim=1)
return F.sigmoid(canvas)
def draw(count, memory_size, file, device='cuda'):
testtransform = transforms.Compose([transforms.ToTensor()])
testset = torchvision.datasets.MNIST(root='./data/mnist', train=False,
download=True, transform=testtransform)
testloader = torch.utils.data.DataLoader(testset, pin_memory=True, batch_size=128,
shuffle=True, num_workers=10)
base_dir = os.path.join('mnist_' + str(memory_size), "4")
model = MnistDraw(count, memory_size, output_stages=True)
optimizer = optim.SGD(filter(lambda p: p.requires_grad, model.parameters()), lr=0)
from datetime import datetime
current_time = datetime.now().strftime('%b%d_%H-%M-%S')
from visualise import StagesGrid
trial = Trial(model, optimizer, nn.MSELoss(reduction='sum'), ['loss'], pass_state=True, callbacks=[
callbacks.TensorBoardImages(comment=current_time, nrow=10, num_images=20, name='Prediction', write_each_epoch=True,
key=torchbearer.Y_PRED, pad_value=1),
callbacks.TensorBoardImages(comment=current_time + '_mnist', nrow=10, num_images=20, name='Target', write_each_epoch=False,
key=torchbearer.Y_TRUE, pad_value=1),
StagesGrid('mnist_stages.png', STAGES, 20)
]).load_state_dict(torch.load(os.path.join(base_dir, file)), resume=False).with_generators(train_generator=testloader, val_generator=testloader).for_train_steps(1).for_val_steps(1).to(device)
trial.run() # Evaluate doesn't work with tensorboard in torchbearer, seems to have been fixed in most recent version
def run(count, memory_size, iteration, device='cuda'):
traintransform = transforms.Compose([transforms.RandomRotation(20), transforms.ToTensor()])
trainset = torchvision.datasets.MNIST(root='./data/mnist', train=True,
download=True, transform=traintransform)
trainloader = torch.utils.data.DataLoader(trainset, pin_memory=True, batch_size=128,
shuffle=True, num_workers=10)
testtransform = transforms.Compose([transforms.ToTensor()])
testset = torchvision.datasets.MNIST(root='./data/mnist', train=False,
download=True, transform=testtransform)
testloader = torch.utils.data.DataLoader(testset, pin_memory=True, batch_size=128,
shuffle=True, num_workers=10)
base_dir = os.path.join('mnist_' + str(memory_size), "4")
model = MnistDraw(count, memory_size)
optimizer = optim.Adam(filter(lambda p: p.requires_grad, model.parameters()), lr=1e-3)
from datetime import datetime
current_time = datetime.now().strftime('%b%d_%H-%M-%S')
trial = Trial(model, optimizer, nn.MSELoss(reduction='sum'), ['loss'], pass_state=True, callbacks=[
tm.kl_divergence(MU, LOGVAR),
callbacks.MostRecent(os.path.join(base_dir, 'iter_' + str(iteration) + '.{epoch:02d}.pt')),
callbacks.GradientClipping(5),
callbacks.ExponentialLR(0.99),
callbacks.TensorBoardImages(comment=current_time, name='Prediction', write_each_epoch=True,
key=torchbearer.Y_PRED),
callbacks.TensorBoardImages(comment=current_time + '_mnist', name='Target', write_each_epoch=True,
key=torchbearer.Y_TRUE)
]).with_generators(train_generator=trainloader, val_generator=testloader).to(device)
trial.run(100)
if __name__ == "__main__":
run(12, 256, 0)
draw(12, 256, 'iter_0.99.pt')