40 lines
2.2 KiB
Python
40 lines
2.2 KiB
Python
import torch.nn as nn
|
|
|
|
class DogCatClassifier(nn.Module):
|
|
def __init__ (self):
|
|
super().__init__()
|
|
|
|
# 3 color (RGB) image, so tensor is of shape (B x 3 x H X W)
|
|
# if we take a look, we can see that the images are of size 32 * 32 if we look at them in a file explorer, so our H and W are 32 in this case
|
|
self.conv1 = nn.Conv2d(3, 32, 3, padding = 1) # passes conv kernel over batch and increases num channels from 3 (for RBG) to 32
|
|
self.relu = nn.ReLU(inplace = True) # relu to add nonlinearity
|
|
self.mp = nn.MaxPool2d(2) # reduces h and w of img by a factor of 2
|
|
self.bn1 = nn.BatchNorm2d(32) #normalizes over z distribution https://arxiv.org/abs/1502.03167
|
|
|
|
# tensor size is now (B x 32 x 32/2 = 16 x 32/2 = 16)
|
|
|
|
self.conv2 = nn.Conv2d(32, 64, 3, padding = 1) # 32 channels to 64 ch with 3x3 kernel
|
|
self.bn2 = nn.BatchNorm2d(64) #normalizes over z distribution https://arxiv.org/abs/1502.03167
|
|
|
|
# tensor size is now (B x 64 x 32/4 = 8 x 32/4 = 8)
|
|
|
|
self.conv3 = nn.Conv2d(64, 128, 3, padding = 1) # 64 channels to 128 ch with 3x3 kernelnn
|
|
self.bn3= nn.BatchNorm2d(128) # normalizes over z distribution https://arxiv.org/abs/1502.03167
|
|
|
|
# tensor size is now (B x 128 x 32/8 = 4 x 32/8 = 4)
|
|
# basically, we have B batches, 128 channels, and a 4x4 pixel representation of our initial image
|
|
self.fc1 = nn.Linear(128 * 4 * 4 , 512)# 2048 feats (ch x h x w), lowkey had to calculator it lol
|
|
self.dropout = 0.5 # tunable, removes half of the values and replaces them with 0s
|
|
self.fc2 = nn.Linear(512, 1) # 512 feats to 1 scalar output
|
|
|
|
def forward(self, x):
|
|
x = self.bn1(self.mp(self.relu(self.conv1(x))))
|
|
x = self.bn2(self.mp(self.relu(self.conv2(x))))
|
|
x = self.bn3(self.mp(self.relu(self.conv3(x))))
|
|
x = x.view(x.size(0), -1) # see model_ez for why we do this before linear layers
|
|
# reformats for use in linear layer
|
|
x = self.fc1(x)
|
|
x = nn.functional.relu(x) # relu to add nonlinearity
|
|
x = self.fc2(x)
|
|
x = nn.functional.sigmoid(x) # 1 / 1 + e ^(-x)
|
|
return x
|