Skip to content
Snippets Groups Projects
Commit 7d7a2508 authored by Julien Dejasmin's avatar Julien Dejasmin
Browse files

new archi omniglot CNN

parent 753028ef
Branches
No related tags found
No related merge requests found
Show whitespace changes
Inline Side-by-side
Showing
with 66 additions and 55 deletions
utils/models.py
+ 66
− 55
View file @ 7d7a2508
...@@ -211,48 +211,53 @@ class NoBinaryNetOmniglotClassification(Net): ...@@ -211,48 +211,53 @@ class NoBinaryNetOmniglotClassification(Net):
def __init__(self): def __init__(self):
super(NoBinaryNetOmniglotClassification, self).__init__() super(NoBinaryNetOmniglotClassification, self).__init__()
self.layer1 = nn.Conv2d(1, 128, kernel_size=3, padding=1, stride=2) self.layer1 = nn.Conv2d(1, 64, kernel_size=3, padding=1, stride=1)
self.batchNorm1 = nn.BatchNorm2d(128) self.batchNorm1 = nn.BatchNorm2d(64)
self.dropout1 = nn.Dropout(0.5) #50 % probability # self.dropout1 = nn.Dropout(0.5) #50 % probability
# self.maxPool1 = nn.MaxPool2d(kernel_size=2, stride=2) self.maxPool1 = nn.MaxPool2d(kernel_size=2, stride=2)
self.act_layer1 = nn.ReLU() self.act_layer1 = nn.ReLU()
self.layer2 = nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2) self.layer2 = nn.Conv2d(64, 128, kernel_size=3, padding=1, stride=1)
self.batchNorm2 = nn.BatchNorm2d(256) self.batchNorm2 = nn.BatchNorm2d(128)
self.dropout2 = nn.Dropout(0.5) # self.dropout2 = nn.Dropout(0.5)
# self.maxPool2 = nn.MaxPool2d(kernel_size=2, stride=2) self.maxPool2 = nn.MaxPool2d(kernel_size=2, stride=2)
self.act_layer2 = nn.ReLU() self.act_layer2 = nn.ReLU()
self.layer3 = nn.Conv2d(256, 256, kernel_size=3, padding=1, stride=2) self.layer3 = nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=1)
self.batchNorm3 = nn.BatchNorm2d(256) self.batchNorm3 = nn.BatchNorm2d(256)
self.dropout3 = nn.Dropout(0.5) # self.dropout3 = nn.Dropout(0.5)
# self.maxPool3 = nn.MaxPool2d(kernel_size=2, stride=2) self.maxPool3 = nn.MaxPool2d(kernel_size=2, stride=2)
self.act_layer3 = nn.ReLU() self.act_layer3 = nn.ReLU()
self.layer4 = nn.Conv2d(256, 256, kernel_size=3, padding=1, stride=2) self.layer4 = nn.Conv2d(256, 512, kernel_size=3, padding=1, stride=1)
self.batchNorm4 = nn.BatchNorm2d(256) self.batchNorm4 = nn.BatchNorm2d(512)
self.dropout4 = nn.Dropout(0.5) # self.dropout4 = nn.Dropout(0.5)
self.maxPool4 = nn.MaxPool2d(kernel_size=2, stride=2)
self.act_layer4 = nn.ReLU() self.act_layer4 = nn.ReLU()
self.layer5 = nn.Conv2d(256, 256, kernel_size=3, padding=1, stride=2) self.layer5 = nn.Conv2d(512, 512, kernel_size=3, padding=1, stride=1)
self.batchNorm5 = nn.BatchNorm2d(256) self.batchNorm5 = nn.BatchNorm2d(512)
# self.maxPool4 = nn.MaxPool2d(kernel_size=2, stride=2) self.maxPool5 = nn.MaxPool2d(kernel_size=2, stride=2)
self.act_layer5 = nn.ReLU() self.act_layer5 = nn.ReLU()
self.fc = nn.Linear(3 * 3 * 256, 1623) self.fc1 = nn.Linear(3 * 3 * 512, 4096)
self.act_fc1 = nn.ReLU()
self.dropout1 = nn.Dropout(0.5)
self.fc2 = nn.Linear(4096, 1623)
def forward(self, input): def forward(self, input):
x = input x = input
slope = 1.0 slope = 1.0
""" """maxPool1
x_layer1 = self.act_layer1(self.maxPool1(self.batchNorm1(self.layer1(x) * slope))) x_layer1 = self.act_layer1(self.maxPool1(self.batchNorm1(self.layer1(x) * slope)))
x_layer2 = self.act_layer2(self.maxPool2(self.batchNorm2(self.layer2(x_layer1)))) x_layer2 = self.act_layer2(self.maxPool2(self.batchNorm2(self.layer2(x_layer1))))
x_layer3 = self.act_layer3(self.maxPool3(self.batchNorm3(self.layer3(x_layer2)))) x_layer3 = self.act_layer3(self.maxPool3(self.batchNorm3(self.layer3(x_layer2))))
x_layer4 = self.act_layer4(self.maxPool4(self.batchNorm4(self.layer4(x_layer3)))) x_layer4 = self.act_layer4(self.maxPool4(self.batchNorm4(self.layer4(x_layer3))))
""" """
x_layer1 = self.act_layer1(self.dropout1(self.batchNorm1(self.layer1(x) * slope))) x_layer1 = self.act_layer1(self.maxPool1(self.batchNorm1(self.layer1(x) * slope)))
x_layer2 = self.act_layer2(self.dropout2(self.batchNorm2(self.layer2(x_layer1)))) x_layer2 = self.act_layer2(self.maxPool2(self.batchNorm2(self.layer2(x_layer1))))
x_layer3 = self.act_layer3(self.dropout3(self.batchNorm3(self.layer3(x_layer2)))) x_layer3 = self.act_layer3(self.maxPool3(self.batchNorm3(self.layer3(x_layer2))))
x_layer4 = self.act_layer4(self.dropout4(self.batchNorm4(self.layer4(x_layer3)))) x_layer4 = self.act_layer4(self.maxPool4(self.batchNorm4(self.layer4(x_layer3))))
x_layer5 = self.act_layer5(self.batchNorm5(self.layer5(x_layer4))) x_layer5 = self.act_layer5(self.maxPool5(self.batchNorm5(self.layer5(x_layer4))))
x_layer5 = x_layer5.view(x_layer5.size(0), -1) x_layer5 = x_layer5.view(x_layer5.size(0), -1)
x_fc = self.fc(x_layer5) x_fc1 = self.dropout1(self.act_fc1(self.fc1(x_layer5)))
x_out = F.log_softmax(x_fc, dim=1) x_fc2 = self.fc2(x_fc1)
x_out = F.log_softmax(x_fc2, dim=1)
return x_out return x_out
...@@ -274,10 +279,10 @@ class BinaryNetOmniglotClassification(Net): ...@@ -274,10 +279,10 @@ class BinaryNetOmniglotClassification(Net):
self.third_conv_layer = third_conv_layer self.third_conv_layer = third_conv_layer
self.fourth_conv_layer = fourth_conv_layer self.fourth_conv_layer = fourth_conv_layer
self.layer1 = nn.Conv2d(1, 128, kernel_size=3, padding=1, stride=2) self.layer1 = nn.Conv2d(1, 64, kernel_size=3, padding=1, stride=1)
self.batchNorm1 = nn.BatchNorm2d(128) self.batchNorm1 = nn.BatchNorm2d(64)
self.dropout1 = nn.Dropout(0.5) # self.dropout1 = nn.Dropout(0.5)
# self.maxPool1 = nn.MaxPool2d(kernel_size=2, stride=2) self.maxPool1 = nn.MaxPool2d(kernel_size=2, stride=2)
if self.first_conv_layer: if self.first_conv_layer:
if self.mode == 'Deterministic': if self.mode == 'Deterministic':
self.act_layer1 = DeterministicBinaryActivation(estimator=estimator) self.act_layer1 = DeterministicBinaryActivation(estimator=estimator)
...@@ -285,10 +290,10 @@ class BinaryNetOmniglotClassification(Net): ...@@ -285,10 +290,10 @@ class BinaryNetOmniglotClassification(Net):
self.act_layer1 = StochasticBinaryActivation(estimator=estimator) self.act_layer1 = StochasticBinaryActivation(estimator=estimator)
else: else:
self.act_layer1 = nn.ReLU() self.act_layer1 = nn.ReLU()
self.layer2 = nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=2) self.layer2 = nn.Conv2d(64, 128, kernel_size=3, padding=1, stride=1)
self.batchNorm2 = nn.BatchNorm2d(256) self.batchNorm2 = nn.BatchNorm2d(128)
self.dropout2 = nn.Dropout(0.5) # self.dropout2 = nn.Dropout(0.5)
# self.maxPool2 = nn.MaxPool2d(kernel_size=2, stride=2) self.maxPool2 = nn.MaxPool2d(kernel_size=2, stride=2)
if self.second_conv_layer: if self.second_conv_layer:
if self.mode == 'Deterministic': if self.mode == 'Deterministic':
self.act_layer2 = DeterministicBinaryActivation(estimator=estimator) self.act_layer2 = DeterministicBinaryActivation(estimator=estimator)
...@@ -296,10 +301,10 @@ class BinaryNetOmniglotClassification(Net): ...@@ -296,10 +301,10 @@ class BinaryNetOmniglotClassification(Net):
self.act_layer2 = StochasticBinaryActivation(estimator=estimator) self.act_layer2 = StochasticBinaryActivation(estimator=estimator)
else: else:
self.act_layer2 = nn.ReLU() self.act_layer2 = nn.ReLU()
self.layer3 = nn.Conv2d(256, 256, kernel_size=3, padding=1, stride=2) self.layer3 = nn.Conv2d(128, 256, kernel_size=3, padding=1, stride=1)
self.batchNorm3 = nn.BatchNorm2d(256) self.batchNorm3 = nn.BatchNorm2d(256)
self.dropout3 = nn.Dropout(0.5) # self.dropout3 = nn.Dropout(0.5)
# self.maxPool3 = nn.MaxPool2d(kernel_size=2, stride=2) self.maxPool3 = nn.MaxPool2d(kernel_size=2, stride=2)
if self.third_conv_layer: if self.third_conv_layer:
if self.mode == 'Deterministic': if self.mode == 'Deterministic':
self.act_layer3 = DeterministicBinaryActivation(estimator=estimator) self.act_layer3 = DeterministicBinaryActivation(estimator=estimator)
...@@ -307,10 +312,10 @@ class BinaryNetOmniglotClassification(Net): ...@@ -307,10 +312,10 @@ class BinaryNetOmniglotClassification(Net):
self.act_layer3 = StochasticBinaryActivation(estimator=estimator) self.act_layer3 = StochasticBinaryActivation(estimator=estimator)
else: else:
self.act_layer3 = nn.ReLU() self.act_layer3 = nn.ReLU()
self.layer4 = nn.Conv2d(256, 256, kernel_size=3, padding=1, stride=2) self.layer4 = nn.Conv2d(256, 512, kernel_size=3, padding=1, stride=1)
self.batchNorm4 = nn.BatchNorm2d(256) self.batchNorm4 = nn.BatchNorm2d(512)
self.dropout4 = nn.Dropout(0.5) # self.dropout4 = nn.Dropout(0.5)
# self.maxPool4 = nn.MaxPool2d(kernel_size=2, stride=2) self.maxPool4 = nn.MaxPool2d(kernel_size=2, stride=2)
if self.fourth_conv_layer: if self.fourth_conv_layer:
if self.mode == 'Deterministic': if self.mode == 'Deterministic':
self.act_layer4 = DeterministicBinaryActivation(estimator=estimator) self.act_layer4 = DeterministicBinaryActivation(estimator=estimator)
...@@ -318,10 +323,15 @@ class BinaryNetOmniglotClassification(Net): ...@@ -318,10 +323,15 @@ class BinaryNetOmniglotClassification(Net):
self.act_layer4 = StochasticBinaryActivation(estimator=estimator) self.act_layer4 = StochasticBinaryActivation(estimator=estimator)
else: else:
self.act_layer4 = nn.ReLU() self.act_layer4 = nn.ReLU()
self.layer5 = nn.Conv2d(256, 256, kernel_size=3, padding=1, stride=2) self.layer5 = nn.Conv2d(512, 512, kernel_size=3, padding=1, stride=1)
self.batchNorm5 = nn.BatchNorm2d(256) self.batchNorm5 = nn.BatchNorm2d(512)
self.maxPool5 = nn.MaxPool2d(kernel_size=2, stride=2)
self.act_layer5 = nn.ReLU() self.act_layer5 = nn.ReLU()
self.fc = nn.Linear(3 * 3 * 256, 1623) self.fc1 = nn.Linear(3 * 3 * 512, 4096)
self.act_fc1 = nn.ReLU()
self.dropout1 = nn.Dropout(0.5)
self.fc2 = nn.Linear(4096, 1623)
def forward(self, input): def forward(self, input):
x = input x = input
...@@ -345,25 +355,26 @@ class BinaryNetOmniglotClassification(Net): ...@@ -345,25 +355,26 @@ class BinaryNetOmniglotClassification(Net):
x_layer4 = self.act_layer4(self.maxPool4(self.batchNorm4(self.layer4(x_layer3) * slope))) x_layer4 = self.act_layer4(self.maxPool4(self.batchNorm4(self.layer4(x_layer3) * slope)))
""" """
if self.first_conv_layer: if self.first_conv_layer:
x_layer1 = self.act_layer1(((self.dropout1(self.batchNorm1(self.layer1(x)))), slope)) x_layer1 = self.act_layer1(((self.maxPool1(self.batchNorm1(self.layer1(x)))), slope))
else: else:
x_layer1 = self.act_layer1(self.dropout1(self.batchNorm1(self.layer1(x) * slope))) x_layer1 = self.act_layer1(self.maxPool1(self.batchNorm1(self.layer1(x) * slope)))
if self.second_conv_layer: if self.second_conv_layer:
x_layer2 = self.act_layer2(((self.dropout2(self.batchNorm2(self.layer2(x_layer1)))), slope)) x_layer2 = self.act_layer2(((self.maxPool2(self.batchNorm2(self.layer2(x_layer1)))), slope))
else: else:
x_layer2 = self.act_layer2(self.dropout2(self.batchNorm2(self.layer2(x_layer1) * slope))) x_layer2 = self.act_layer2(self.maxPool2(self.batchNorm2(self.layer2(x_layer1) * slope)))
if self.third_conv_layer: if self.third_conv_layer:
x_layer3 = self.act_layer3(((self.dropout3(self.batchNorm3(self.layer3(x_layer2)))), slope)) x_layer3 = self.act_layer3(((self.maxPool3(self.batchNorm3(self.layer3(x_layer2)))), slope))
else: else:
x_layer3 = self.act_layer3(self.dropout3(self.batchNorm3(self.layer3(x_layer2) * slope))) x_layer3 = self.act_layer3(self.maxPool3(self.batchNorm3(self.layer3(x_layer2) * slope)))
if self.fourth_conv_layer: if self.fourth_conv_layer:
x_layer4 = self.act_layer4(((self.dropout4(self.batchNorm4(self.layer4(x_layer3)))), slope)) x_layer4 = self.act_layer4(((self.maxPool4(self.batchNorm4(self.layer4(x_layer3)))), slope))
else: else:
x_layer4 = self.act_layer4(self.dropout4(self.batchNorm4(self.layer4(x_layer3) * slope))) x_layer4 = self.act_layer4(self.maxPool4(self.batchNorm4(self.layer4(x_layer3) * slope)))
x_layer5 = self.act_layer5(self.batchNorm5(self.layer5(x_layer4) * slope)) x_layer5 = self.act_layer5(self.maxPool5(self.batchNorm5(self.layer5(x_layer4))))
x_layer5 = x_layer5.view(x_layer5.size(0), -1) x_layer5 = x_layer5.view(x_layer5.size(0), -1)
x_fc = self.fc(x_layer5) x_fc1 = self.dropout1(self.act_fc1(self.fc1(x_layer5)))
x_out = F.log_softmax(x_fc, dim=1) x_fc2 = self.fc2(x_fc1)
x_out = F.log_softmax(x_fc2, dim=1)
return x_out return x_out
......
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Please register or to comment