Question 4 - Assignment 2

About the Code

Finetuning AlexNet's and VGG16 layers where the classes are reduced from 1000 to the number of classes in dataset(8) in the classifier layer.

Installing Dependencies

1. Install PyTorch by following the instructions given on the link here.
2. Run pip3 install torchvision tensorflow==1.4 tensorboardX in command line.

Running the Code

For AlexNet

1. First, get the dataset in two folders: train and test by modifying paths in data_process.py. Then, execute the command python3 data_process.py.
2. Then modify the dataset path in alexnet.py, where the newly created folders are saved.
3. Then run python3 alexnet.py.

For VGG16

1. Repeat the 2 and 3 steps as mentioned for AlexNet with the following file: vgg16.py.

Visualizing the Results

To visualize the current results, open a terminal session, and run the following command: tensorboard --logdir runs

Obtained Results

For AlexNet (Epochs = 150)

1. Training Accuracy = 92.53%
2. Training Loss = 0.26
3. Testing Accuracy = 92.5

For VGG16 (Epochs = 16)

1. Training Accuracy = 89.63%
2. Training Loss = 0.34
3. Testing Accuracy = 88.38

Graphs

For AlexNet (Epochs = 150)

For VGG16 (Epochs = 16)