This project involves building and training a convolutional auto-encoder model for the tasks of image reconstruction and denoising using the CIFAR10 dataset.
Design and train a convolutional auto-encoder for image reconstruction on the CIFAR10 dataset.
- Utilized the CIFAR10 data loader from PyTorch.
- The auto-encoder model consists of 2 encoder blocks (2D convolution) and 2 decoder blocks (transpose 2D convolution).
- Channel configurations:
- Encoder: (3,8), (8,8)
- Decoder: (8,8), (8,3)
- Activation functions: Relu and Tanh.
- Loss: Mean Squared Error (MSE).
- Optimizer: Adam.
- The model converged at a loss of
0.005
by epoch 9. The plot can be seen below:
Train the convolutional auto-encoder for image denoising on CIFAR10.
- The input images were perturbed with Gaussian noise (mean=0, variance=0.1) using the
torch.randn
function. - Same model and training configurations as the image reconstruction task.
- The model converged at a loss of
0.017
by epoch 8. The plo can be seen below:
- PyTorch
- skimage (for PSNR and SSIM calculations)