-
Notifications
You must be signed in to change notification settings - Fork 0
/
capsulenetwork.py
259 lines (243 loc) · 11.9 KB
/
capsulenetwork.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
251
252
253
254
255
256
257
258
259
import torch
from torchvision import datasets, transforms
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from make_graph import plot_graph, generate_reconstructions
import numpy as np
import time
# execfile("make_graph.py") -> For Google Colab
def squash(tensor, dim=-1):
# Square of Absolute Value
squared_norm = (tensor ** 2).sum(dim=dim, keepdim=True)
# scale = ||Vc||^2 / (1 + ||Vc||^2)
scale = squared_norm / (1 + squared_norm)
# squash = scale * (Vc / ||Vc||)
return scale * tensor / torch.sqrt(squared_norm)
def softmax(tensor, dim=1):
# Shape of tensor = [10 x 128 x 1152 x 1 x 16]
transposed_tensor = tensor.transpose(dim, len(tensor.shape)-1)
# Shape of tensor = [10 x 128 x 16 x 1 x 1152]
softmaxed_tensor = F.softmax(transposed_tensor.contiguous().view(-1, transposed_tensor.shape[-1]), dim=-1)
# Shape of softmaxed_tensor = [20480 x 1152]
output_tensor = softmaxed_tensor.view(*transposed_tensor.shape).transpose(dim, len(tensor.shape)-1)
# Shape of output_tensor = [10 x 128 x 1152 x 1 x 16]
return output_tensor
class PrimaryCapsuleLayer(nn.Module):
def __init__(self, num_capsules, in_channels, out_channels, kernel_size, stride):
super(PrimaryCapsuleLayer, self).__init__()
# Repeat Convolutions based on number of capsules (8 in this case).
self.capsules = nn.ModuleList([nn.Conv2d(in_channels=256, out_channels=32, kernel_size=9, stride=2, padding=0) for _ in range(num_capsules)])
def forward(self, x):
# Shape of x = [128 x 256 x 20 x 20]
# Apply Convolutions to Input x to generate capsules.
outputs = [capsule(x).view(x.shape[0], -1, 1) for capsule in self.capsules]
# [128 x 256 x 20 x 20] --> [128 x (32*6*6) x 1]
# outputs is now a list of length 8 (number of capsules) with each element as [128 x 1152 x 1]
outputs = torch.cat(outputs, dim=-1)
# Shape of outputs = [128 x 1152 x 8]
outputs = squash(outputs)
return outputs
class DigitCapsuleLayer(nn.Module):
def __init__(self, num_capsules, routing_nodes, in_channels, out_channels, routing_iterations=3):
super(DigitCapsuleLayer, self).__init__()
self.routing_iterations = routing_iterations
# Weights = [10 x 1152 x 8 x 16]
self.weights = torch.randn(num_capsules, routing_nodes, in_channels, out_channels).cuda()
def forward(self, x):
# Shape of x = [128 x 1152 x 8]
# Shape of x = [1 x 128 x 1152 x 1 x 8], Shape of Weights = [10 x 1 x 1152 x 8 x16]
x_hat = torch.matmul(x[None,:, :, None, :], self.weights[:, None, :, :, :])
# Shape of x_hat = [10 x 128 x 1152 x 1 x 16]
# b is a temporary variable that will store the value of routing weights c and will be gradually updated.
b = torch.zeros(*x_hat.shape).cuda()
# Dynamic Routing Algorithm
for i in range(self.routing_iterations):
# Routing weights for all capsules of layer l (i.e dim_2 = 1152)
c = softmax(b, dim=2)
# Weighted sum of x_hat and routing weights c across all capsules of layer l (i.e. dim_2 = 1152)
outputs = squash((x_hat * c).sum(dim=2, keepdim=True))
# Shape of outputs = [10 x 128 x 1 x 1 x 16]
if i != (self.routing_iterations-1):
# Weight Update Step: Update weight b using dot product similarity.
db = (x_hat * outputs).sum(dim=-1, keepdim=True)
# Shape of db = [10 x 128 x 1152 x 1 x 1]
b = b + db
return outputs
class CapsuleNetwork(nn.Module):
def __init__(self):
super(CapsuleNetwork, self).__init__()
self.conv1 = nn.Conv2d(in_channels=1, out_channels=256, kernel_size=9, stride=1, padding=0)
self.primary_capsules = PrimaryCapsuleLayer(num_capsules=8, in_channels=256, out_channels=32, kernel_size=9, stride=2)
self.digit_capsules = DigitCapsuleLayer(num_capsules=10, routing_nodes=32*6*6, in_channels=8, out_channels=16, routing_iterations=3)
self.decoder = nn.Sequential(nn.Linear(16*10, 512), nn.ReLU(inplace=True), nn.Linear(512, 1024), nn.ReLU(inplace=True), nn.Linear(1024, 784), nn.Sigmoid())
def forward(self, x, y=None, pose=False):
# Encoder Network
# Layer 1: Convolutional Layer
# Input = [128 x 1 x 28 x 28]
# Output = [128 x 256 x 20 x 20]
x = F.relu(self.conv1(x), inplace=True)
# Layer 2: Primary Capsule Layer
# Input = [128 x 256 x 20 x 20]
# Output = [128 x (32*6*6) x 8] = [128 x 1152 x 8]
x = self.primary_capsules(x)
# Layer 3: Digit Capsule Layer
# Input = [128 x 1152 x 8]
# Output = [10 x 128 x 1 x 1 x 16]
x = self.digit_capsules(x)
x = x.squeeze().transpose(0, 1)
# Shape of x = [128, 10, 16]
class_scores = torch.sqrt((x ** 2).sum(dim=-1))
# Shape of class_scores = [128 x 10]
class_probabilities = F.softmax(class_scores, dim=-1)
# Shape of class_probabilities = [128 x 10]
if y is None: # During Testing
# Generate one hot encoded labels from class probabilities
_, labels = class_probabilities.topk(k=1, dim=1)
y = torch.eye(10).cuda().index_select(dim=0, index=labels.squeeze())
if pose == True:
# Dive deeper into what each of the 16 dimensions represent by fluctuating each value while holding
#other values constant
dimension_list = []
for dim in range(0, x.shape[2]):
alpha = -0.25
interval_list = []
while(alpha <= 0.25):
temp = x.contiguous().view(1, -1, x.shape[2]).clone()
# Update one dimension while keeping other dimensions constant.
temp[0, :, dim] = temp[0, :, dim] + alpha
alpha = alpha + 0.05
interval_list.append(temp)
temp = torch.cat(interval_list, dim=0).unsqueeze(dim=0)
dimension_list.append(temp)
x = torch.cat(dimension_list, dim=0)
x = x.view(x.shape[0], x.shape[1], x.shape[2]/10, 10, x.shape[3])
repeat_y = x.shape[0] * x.shape[1]
x = x.view(-1, x.shape[3], x.shape[4])
y = y.unsqueeze(dim=0)
y_list = []
for _ in range(repeat_y):
y_list.append(y)
y = torch.cat(y_list, dim=0)
y = y.view(-1, y.shape[2])
# Decoder Network
# Layer 4: Fully Connected Layer 1
# Input = 16*10 = 160 (x[128 x 10 x 16] * y[128 x 10 x 1](one-hot encoded labels) is done to ignore incorrect vectors through 0 masking)
# Output = 512
# Layer 5: Fully Connected Layer 2
# Input = 512
# Output = 1024
# Layer 6: Fully Connected Layer 3
# Input = 1024
# Output = 28*28 = 784
reconstructions = self.decoder((x * y[:, :, None]).view(x.shape[0], -1))
# Shape of reconstructions = [128 x 784]
return class_probabilities, reconstructions
class CapsuleLoss(nn.Module):
def __init__(self):
super(CapsuleLoss, self).__init__()
self.reconstruction_loss = nn.MSELoss(size_average=False)
self.m_positive = 0.9
self.m_negative = 0.1
self.lambda_constant = 0.5
self.alpha_constant = 0.0005
def forward(self, images, reconstructions, labels, class_probabilities):
# Compute Margin Loss
# In this case ReLU works the same as the Max function used in the original paper i.e. Make sure everything is positive.
lhs = labels * F.relu(self.m_positive - class_probabilities, inplace=True)**2
rhs = (1. - labels) * F.relu(class_probabilities - self.m_negative, inplace=True)**2
margin_loss = lhs + self.lambda_constant*rhs
# Get scalar value
margin_loss = margin_loss.sum()
assert torch.numel(images) == torch.numel(reconstructions)
images = images.view(reconstructions.shape[0], -1)
# Compute Reconstruction Loss
reconstruction_loss = self.reconstruction_loss(images, reconstructions)
# Compute Overall Loss
overall_loss = (margin_loss + self.alpha_constant * reconstruction_loss) / images.shape[0]
return overall_loss
transform = transforms.Compose([transforms.ToTensor()])
trainset = datasets.MNIST("MNIST_data/", transform=transform, download=True, train=True)
testset = datasets.MNIST("MNIST_data/", transform=transform, download=True, train=False)
model = CapsuleNetwork()
model.cuda()
optimizer = optim.Adam(model.parameters())
loss_function = CapsuleLoss()
epochs = 100
overall_training_accuracy = []
overall_training_loss = []
mode = "Test"
if mode == "Train":
# Training
trainloader = torch.utils.data.DataLoader(trainset, shuffle=True, batch_size=128)
for epoch in range(epochs):
start_time = time.time()
running_loss = 0
training_accuracy = 0
count = 0
percent = 0
for images, raw_labels in iter(trainloader):
count += 1
# Generate One Hot encodings of labels
labels = torch.eye(10).index_select(dim=0, index=raw_labels)
class_probabilities, reconstructions = model(images.cuda(), labels.cuda())
# Compute Loss and Gradients
loss = loss_function(images.cuda(), reconstructions, labels.cuda(), class_probabilities)
optimizer.zero_grad()
loss.backward()
optimizer.step()
running_loss += loss.item()
_, logits = class_probabilities.topk(k=1, dim=1)
training_accuracy_tensor = logits.view(*raw_labels.shape) == raw_labels.cuda()
training_accuracy += torch.mean(training_accuracy_tensor.type(torch.FloatTensor))
if (count % (int(len(trainloader)/10)+1) == 0):
percent += 10
print "%d %% Complete" % (percent)
percent += 10
print "%d %% Complete" % (percent)
training_loss = running_loss / len(trainloader)
# Calculate Accuracy
training_accuracy = training_accuracy * 100 / len(trainloader)
elapsed_time = time.time() - start_time
print "Epoch %d / %d:" % (epoch+1, epochs)
print "Time Elapsed = %d s" % (elapsed_time)
print "Training Loss = %f" % (training_loss)
print "Training Accuracy = %0.2f %%" % (training_accuracy)
if ((epoch+1) % 10 == 0):
print "Generate Reconstructions:"
generate_reconstructions(reconstructions[:10].cpu(), images[:10].cpu(), epoch+1)
print "Reconstructions Generated"
overall_training_accuracy.append(training_accuracy)
overall_training_loss.append(training_loss)
history = {
"train_loss": overall_training_loss,
"train_accuracy": overall_training_accuracy
}
np.save("history.npy", history)
torch.save(model, "model.pth")
elif mode == "Test":
testloader = torch.utils.data.DataLoader(testset, shuffle=True, batch_size=100)
loaded_model = torch.load("model.pth")
loaded_model.cuda()
loaded_model.eval()
testing_accuracy = 0
# Testing
for images, labels in iter(testloader):
class_probabilities, reconstructions = loaded_model(images.cuda())
_, logits = class_probabilities.topk(k=1, dim=1)
testing_accuracy_tensor = logits.view(*labels.shape) == labels.cuda()
testing_accuracy += torch.mean(testing_accuracy_tensor.type(torch.FloatTensor))
testing_accuracy = testing_accuracy * 100 / len(testloader)
print "Test Accuracy = %0.2f %%" % (testing_accuracy)
generate_reconstructions(reconstructions.cpu(), images.cpu(), "Test")
plot_graph(testing_accuracy)
elif mode == "Pose":
loaded_model = torch.load("model.pth")
loaded_model.eval()
test_indices = list(range(0, 11))
test_sampler = torch.utils.data.SubsetRandomSampler(test_indices)
testloader = torch.utils.data.DataLoader(testset, sampler=test_sampler, batch_size=11)
# Pose Relationship
for images, _ in iter(testloader):
_, reconstructions = loaded_model(images.cuda(), pose=True)
generate_reconstructions(reconstructions.cpu())