Sitansh Mehta Autonomy Bootcamp

main.py

@@ -9,9 +9,207 @@
 * We suggest using a convolutional neural network.
 * TensorFlow Keras has the CIFAR-10 dataset as a module, so you don't need to manually download and unpack it.
 """
-
 # Import whatever libraries/modules you need
 
+import torch
 import numpy as np
+import matplotlib.pyplot as plt
+import torch.optim as optim
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision
+import torchvision.transforms as transforms
+from torchvision import datasets, transforms
 
 # Your working code here
+"""
+CIFAR-10 has 10 classes of images, and each image is 32 x 32 pixels and has 3 color values. 
+Training data set size is 50,000 images, and test dataset size is 10,000 images.
+"""
+# Transforming the images
+"""
+- transforms.Compose() takes in a sequence of transformations that are to be applied to the input data.
+- transforms.ToTensor() converts the input images to tensors with the shape of channels, height, and width. 
+- transforms.Normalize() takes in two arguments, mean and std. Since CIFAR10 is a 3 channel dataset, (0.5, 0.5, 0.5)
+represents the mean values and standard deviation for the RGB channels. 
+"""
+
+transform_data = transforms.Compose(
+    [transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
+
+# Loading the testing and training data sets
+# Batch size is the number of samples processed in one iteration, ie, forward and backward pass.
+batch_size = 16
+
+# Loading the dataset
+train_set = torchvision.datasets.CIFAR10(root='/data',
+                                         train=True,
+                                         download=True,
+                                         transform=transform_data)
+
+train_loader = torch.utils.data.DataLoader(train_set,
+                                           batch_size=batch_size,
+                                           shuffle=True)
+
+test_set = torchvision.datasets.CIFAR10(root='/data',
+                                        train=False,
+                                        download=True,
+                                        transform=transform_data)
+
+test_loader = torch.utils.data.DataLoader(test_set,
+                                          batch_size=batch_size,
+                                          shuffle=True)
+
+
+# Developing a Neural Network Architecture as a class
+# Subclassing nn.Module
+class Network (nn.Module):
+    def __init__(self):
+        # Network inherits its functions and behavior from super class nn.Module
+        super(Network, self).__init__()
+        self.conv1 = nn.Conv2d(
+            in_channels=3, out_channels=32, kernel_size=(3, 3))
+        self.conv2 = nn.Conv2d(
+            in_channels=32, out_channels=64, kernel_size=(3, 3))
+        self.maxpool = nn.MaxPool2d(kernel_size=2, stride=2)
+        self.relu = nn.ReLU()
+        self.fc1 = nn.Linear(64 * 6 * 6, 500)
+        self.fc2 = nn.Linear(500, 10)
+
+    def forward(self, x):
+        x = self.maxpool(F.relu(self.conv1(x)))
+        x = self.maxpool(F.relu(self.conv2(x)))
+        x = torch.flatten(x, 1)
+        x = F.relu(self.fc1(x))
+        x = self.fc2(x)
+
+        return x
+
+
+# Instance of the neural network class "Network"
+net = Network()
+
+train_loss = []
+train_accu = []
+
+test_loss = []
+test_accu = []
+
+# Defining the loss/cost function
+"""
+CrossEntropyLoss() is a loss function that computes the SoftMax Activation and then computes the cross entropy loss.
+The SoftMax Activation takes the output of the model in the form of raw data for each class and converts it 
+to a probability distribution over the classes. It normalises the probabilities to add up to 1.
+The cross entropy loss function measures the dissimilarity between the predicted probability and the true labels. 
+
+"optimizer" is an optimizer object used to train the neural network using SGD--stochastic gradient descent.
+"lr" is the learning rate and it determines the step size at each iteration of the optimizatio process
+"momentum" helps speed up the optimization process
+"""
+criterion = nn.CrossEntropyLoss()
+optimizer = optim.SGD(net.parameters(), lr=0.001, momentum=0.9)
+
+
+# Defining a function to training the neural network
+"""
+The models parameters are updated based on the feedback from the loss functions.
+The model performs forward propagation to make predictions, calculates the loss between 
+the predicted and actual values, and then performs backward propagation to update the models
+weights and biases using an optimization algo (optimizer variable from line 106).
+"""
+
+
+def train(network, data_loader, criterion, optimizer):
+    dataset_size = len(data_loader.dataset)
+    num_batches = len(data_loader)
+    net.train()
+    loss_in_one_epoch = 0
+    correct_pred = 0
+
+    for i, data in enumerate(data_loader, 0):
+        inputs, labels = data
+
+        optimizer.zero_grad()
+
+        model_outputs = net(inputs)
+        loss = criterion(model_outputs, labels)
+        loss.backward()
+        optimizer.step()
+
+        loss_in_one_epoch += loss.item()
+        correct_pred += (model_outputs.argmax(1) ==
+                         labels).type(torch.float).sum().item()
+
+        if i % 2000 == 1999:
+            print('[%d, %5d] Training loss: %.3f' %
+                  (epoch + 1, i + 1, loss_in_one_epoch / 2000))
+
+    loss_in_one_epoch /= num_batches
+    correct_pred /= dataset_size
+    accuracy = 100 * correct_pred
+
+    print(
+        f"Train Error: \n Accuracy: {accuracy:>0.1f}%, Avg loss: {loss_in_one_epoch:>8f} \n")
+    train_accu.append(correct_pred)
+    train_loss.append(loss_in_one_epoch)
+
+
+# Defining a function to test the neural network
+def test(network, data_loader, criterion):
+    dataset_size = len(data_loader.dataset)
+    num_batches = len(data_loader)
+    net.eval()
+    loss = 0
+    correct_pred = 0
+
+    with torch.no_grad():
+        for data in data_loader:
+            images, labels = data
+
+            model_output = network(images)
+
+            batch_loss = criterion(model_output, labels).item()
+            loss += batch_loss
+            correct_pred += (model_output.argmax(1) ==
+                             labels).type(torch.float).sum().item()
+
+        loss /= num_batches
+        correct_pred /= dataset_size
+        accuracy = 100 * correct_pred
+        print(
+            f"Test Error: \n Accuracy: {accuracy:>0.1f}%, Avg loss: {loss:>8f} \n")
+        test_accu.append(correct_pred)
+        test_loss.append(loss)
+
+# Definig functions to graph the train vs test loss and accuracy
+
+
+def graph_accu():
+    plt.plot(train_accu, '-o')
+    plt.plot(test_accu, '-o')
+    plt.title('Train vs Test Accuracy')
+    plt.xlabel('Epoch')
+    plt.ylabel('Accuracy')
+    plt.legend(['Train', 'Valid'])
+
+
+def graph_loss():
+    plt.plot(train_loss, '-o')
+    plt.plot(test_loss, '-o')
+    plt.title('Train vs Test Loss')
+    plt.xlabel('Epoch')
+    plt.ylabel('Loss')
+    plt.legend(['Training', 'Test'])
+
+
+# Epochs is the number of training iterations the neural network undertakes
+epoch = 10
+
+for epochs in range(epoch):
+    print(f"Epoch {epochs+1}\n-------------------------------")
+    train(net, train_loader, criterion, optimizer)
+    test(net, test_loader, criterion)
+
+print("Training and Testing completed")
+graph_loss()
+graph_accu()