models.py

## TODO: define the convolutional neural network architecture

import torch
import torch.nn as nn
import torch.nn.functional as F
# can use the below import should you choose to initialize the weights of your Net
import torch.nn.init as I


class Net(nn.Module):

    def __init__(self):
        super(Net, self).__init__()

        ## TODO: Define all the layers of this CNN, the only requirements are:
        ## 1. This network takes in a square (same width and height), grayscale image as input
        ## 2. It ends with a linear layer that represents the keypoints
        ## it's suggested that you make this last layer output 136 values, 2 for each of the 68 keypoint (x, y) pairs

        # As an example, you've been given a convolutional layer, which you may (but don't have to) change:

        # 1 input image channel (grayscale), 32 output channels/feature maps, 3x3 square convolution kernel

        # Input 224X224
        self.conv1 = nn.Conv2d(1, 32, 5) # Convolutional layer 1
        # Output size = (Width-FKernel)/Stride +1 = (32, 220, 220)
        self.pool1 = nn.MaxPool2d(2, 2)
        # Layes output = (32, 110, 110)

        self.conv2 = nn.Conv2d(32, 64, 3) # Convolutional layer 2
        self.conv2_bn = nn.BatchNorm2d(64)
        # Output size = (Width-FKernel)/Stride +1 = (64, 108, 108)
        self.pool2 = nn.MaxPool2d(2, 2)
        # Layes output = (64, 54, 54)

        self.conv3 = nn.Conv2d(64, 128, 3) # Convolutional layer 3
        self.conv3_bn = nn.BatchNorm2d(128)
        # Output size = (Width-FKernel)/Stride +1 = (128, 52, 52)
        self.pool3 = nn.MaxPool2d(2, 2)
        # Layes output = (128, 26, 26)

        self.conv4 = nn.Conv2d(128, 256, 3) # Convolutional layer 4
        self.conv4_bn = nn.BatchNorm2d(256)
        # Output size = (Width-FKernel)/Stride +1 = (256, 24, 24)
        self.pool4 = nn.MaxPool2d(2, 2)
        # Layes output = (256, 12, 12)

        self.conv5 = nn.Conv2d(256, 512, 3) # Convolutional layer 4
        self.conv5_bn = nn.BatchNorm2d(512)
        # Output size = (Width-FKernel)/Stride +1 = (256, 10, 10)
        self.pool5 = nn.MaxPool2d(2, 2)
        # Layes output = (512, 5, 5)

        self.fc6 = nn.Linear(512*5*5, 1024) # Dense layer 1
        self.drop6 = nn.Dropout(p=0.2, inplace=True)

        self.fc7 = nn.Linear(1024, 512) # Dense layer 2
        self.drop7 = nn.Dropout(p=0.2, inplace=True)

        self.fc8 = nn.Linear(512, 136) # Dense layer output


        ## Note that among the layers to add, consider including:
        # maxpooling layers, multiple conv layers, fully-connected layers,
        # and other layers (such as dropout or batch normalization) to avoid overfitting



    def forward(self, x):
        ## TODO: Define the feedforward behavior of this model
        ## x is the input image and, as an example, here you may choose to include a pool/conv step:
        ## x = self.pool(F.relu(self.conv1(x)))
        x = self.pool1(F.relu(self.conv1(x)))

        x = self.pool2(F.relu(self.conv2_bn(self.conv2(x))))
        #x = self.pool2(F.relu(self.conv2(x)))

        x = self.pool3(F.relu(self.conv3_bn(self.conv3(x))))
        #x = self.pool3(F.relu(self.conv3(x)))

        x = self.pool4(F.relu(self.conv4_bn(self.conv4(x))))
        #x = self.pool4(F.relu(self.conv4(x)))

        x = self.pool5(F.relu(self.conv5_bn(self.conv5(x))))
        #x = self.pool5(F.relu(self.conv5(x)))

        x = x.view(x.size(0), -1) #reshape

        x = self.drop6(F.relu(self.fc6(x)))
        x = self.drop7(F.relu(self.fc7(x)))
        
        x = self.fc8(x)
        # a modified x, having gone through all the layers of your model, should be returned
        return x