Let's walk through the steps we're about to take.\n",
"\n",
"1. Extend the base Module class:\n",
" \n",
"
class ConvolutionalNetwork(nn.Module):
\n",
" def \\_\\_init\\_\\_(self):
\n",
" super().\\_\\_init\\_\\_()\n",
"\n",
"2. Set up the convolutional layers with
torch.nn.Conv2d()The first layer has one input channel (the grayscale color channel). We'll assign 6 output channels for feature extraction. We'll set our kernel size to 3 to make a 3x3 filter, and set the step size to 1.
\n",
"
self.conv1 = nn.Conv2d(1, 6, 3, 1)\n",
"The second layer will take our 6 input channels and deliver 16 output channels.
\n",
"
self.conv2 = nn.Conv2d(6, 16, 3, 1)\n",
"\n",
"3. Set up the fully connected layers with
torch.nn.Linear().
The input size of (5x5x16) is determined by the effect of our kernels on the input image size. A 3x3 filter applied to a 28x28 image leaves a 1-pixel edge on all four sides. In one layer the size changes from 28x28 to 26x26. We could address this with zero-padding, but since an MNIST image is mostly black at the edges, we should be safe ignoring these pixels. We'll apply the kernel twice, and apply pooling layers twice, so our resulting output will be\n",
"$\\;(((28-2)/2)-2)/2 = 5.5\\;$ which rounds down to 5 pixels per side.
\n",
"
self.fc1 = nn.Linear(5\\*5\\*16, 120)\n",
"
self.fc2 = nn.Linear(120, 84)\n",
"
self.fc3 = nn.Linear(84, 10)\n",
"See below for a more detailed look at this step.
\n",
"\n",
"4. Define the forward method.
Activations can be applied to the convolutions in one line using
F.relu() and pooling is done using
F.max_pool2d()\n",
"
def forward(self, X):
\n",
" X = F.relu(self.conv1(X))
\n",
" X = F.max_pool2d(X, 2, 2)
\n",
" X = F.relu(self.conv2(X))
\n",
" X = F.max_pool2d(X, 2, 2)
\n",
"Flatten the data for the fully connected layers:
\n",
" X = X.view(-1, 5\\*5\\*16)
\n",
" X = F.relu(self.fc1(X))
\n",
" X = self.fc2(X)
\n",
" return F.log_softmax(X, dim=1)\n",
"
"
]
},
{
"cell_type": "markdown",
"metadata": {
"id": "jItof2UxxkxN"
},
"source": [
"