student_code.py

# python imports
import os
from tqdm import tqdm

# torch imports
import torch
import torch.nn as nn
import torch.optim as optim

# helper functions for computer vision
import torchvision
import torchvision.transforms as transforms


class SimpleFCNet(nn.Module):
    """
    A simple neural network with fully connected layers
    """
    def __init__(self, input_shape=(28, 28), num_classes=10):
        super(SimpleFCNet, self).__init__()
        # create the model by adding the layers
        layers = []

        ###################################
        #     fill in the code here       #
        ###################################
        # Add a Flatten layer to convert the 2D pixel array to a 1D vector
        layers.append(nn.Flatten())
        # Add a fully connected / linear layer with 128 nodes
        layers.append(nn.Linear(784, 128))
        # Add ReLU activation
        layers.append(nn.ReLU())
        # Append a fully connected / linear layer with 64 nodes
        layers.append(nn.Linear(128, 64))
        # Add ReLU activation
        layers.append(nn.ReLU())
        # Append a fully connected / linear layer with num_classes (10) nodes
        layers.append(nn.Linear(64, 10))
        self.layers = nn.Sequential(*layers)

        self.reset_params()

    def forward(self, x):
        # the forward propagation
        out = self.layers(x)
        if self.training:
            # softmax is merged into the loss during training
            return out
        else:
            # attach softmax during inference
            out = nn.functional.softmax(out, dim=1)
            return out

    def reset_params(self):
        # to make our model a faithful replica of the Keras version
        for m in self.modules():
            if isinstance(m, nn.Linear):
                nn.init.xavier_uniform_(m.weight)
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0.0)


class SimpleConvNet(nn.Module):
    """
    A simple convolutional neural network
    """
    def __init__(self, input_shape=(32, 32), num_classes=100):
        super(SimpleConvNet, self).__init__()
        ####################################################
        # you can start from here and create a better model
        ####################################################
        # this is a simple implementation of LeNet-5
        layers = []
        fc_dim = 16 * (input_shape[0] // 4 - 3) * (input_shape[1] // 4 - 3)
        # 2 convs
        
        layers.append(nn.Conv2d(3, 64, kernel_size=2, stride=1))  # 64*32*32 
        layers.append(nn.BatchNorm2d(64))
        layers.append(nn.ReLU(inplace=True))
        layers.append(nn.MaxPool2d(kernel_size=2, stride=2))  # 64*16*16

        layers.append(nn.Conv2d(64, 128, kernel_size=2, stride=1))  # 128*15*15
        layers.append(nn.BatchNorm2d(128))
        layers.append(nn.ReLU(inplace=True))
        layers.append(nn.MaxPool2d(kernel_size=2, stride=2))  # 128*7*7
        
        layers.append(nn.Conv2d(128, 256, kernel_size=2, stride=1))  # 256 * 6 * 6
        layers.append(nn.BatchNorm2d(256))
        layers.append(nn.ReLU(inplace=True))
        layers.append(nn.MaxPool2d(kernel_size=2, stride=2)) # 256 * 3 * 3
        layers.append(nn.Flatten())
        # 3 FCs
        layers.append(nn.Linear(2304, 256))
        layers.append(nn.ReLU(inplace=True))
        layers.append(nn.Linear(256, 128))
        layers.append(nn.ReLU(inplace=True))
        layers.append(nn.Linear(128, num_classes))

        self.layers = nn.Sequential(*layers)

    def forward(self, x):
        #################################
        # Update the code here as needed
        #################################
        # the forward propagation
        out = self.layers(x)
        if self.training:
            # softmax is merged into the loss during training
            return out
        else:
            # attach softmax during inference
            out = nn.functional.softmax(out, dim=1)
            return out


def train_model(model, train_loader, optimizer, criterion, epoch):
    """
    model (torch.nn.module): The model created to train
    train_loader (pytorch data loader): Training data loader
    optimizer (optimizer.*): A instance of some sort of optimizer, usually SGD
    criterion (nn.CrossEntropyLoss) : Loss function used to train the network
    epoch (int): Current epoch number
    """
    model.train()
    train_loss = 0.0
    for input, target in tqdm(train_loader, total=len(train_loader)):
        ######################################################
        # fill in the standard training loop of forward pass,
        # backward pass, loss computation and optimizer step
        ######################################################

        # 1) zero the parameter gradients
        optimizer.zero_grad()
        # 2) forward + backward + optimize
        yhat = model(input)
        # Update the train_loss variable
        loss = criterion(yhat, target)
        loss.backward()
        optimizer.step()
        # .item() detaches the node from the computational graph
        # Uncomment the below line after you fill block 1 and 2
        train_loss += loss.item()

    train_loss /= len(train_loader)
    print('[Training set] Epoch: {:d}, Average loss: {:.4f}'.format(epoch+1, train_loss))

    return train_loss


def test_model(model, test_loader, epoch):
    model.eval()
    correct = 0
    with torch.no_grad():
        for input, target in test_loader:
            output = model(input)
            pred = output.max(1, keepdim=True)[1]
            correct += pred.eq(target.view_as(pred)).sum().item()

    test_acc = correct / len(test_loader.dataset)
    print('[Test set] Epoch: {:d}, Accuracy: {:.2f}%\n'.format(
        epoch+1, 100. * test_acc))

    return test_acc