Creating a UNet Convolutional Neural Network for interactively colourising black and white photos.
python3 - m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
Uni-Freiberg: Olaf Ronneberger: 18 May 2015
Debugger Cafe: Sovit Ranjan Rath: 3 April 2023
Modified to take greyscale photos as input, then output LUV colorised photos
- No BatchNorm
- Image input and targets Normalized:
$\mu$ : 0.5,$\sigma$ 0.25 - Epochs: 5
- Mean Squared Error Loss
- No dropout
- Learning Rate: 5e-5
Training loss (smoothed) / Time
| Input | Epoch 1 | Epoch 2 | Epoch 3 | Epoch 4 | Epoch 5 | Ground Truth |
|---|
- Fix HSV normalization
- Run performance evaluation
- Pytorch and Rust GUI : https://medium.com/@heyamit10/loading-and-running-a-pytorch-model-in-rust-f10d2577d570
- Dropout
- No normalisation
- Improved Loss function: MSE + SSIM
- BatchNorm
- Skip Connection Convolution, then
- Hintegration: Convolve hints, then integrate then convolve colour hints alongside down convolutions.
- Dropout (not from school, from network)
- Variable learning rate