Implemeted a Generational Adverserial Network for colorizing black & white images.
Dataset used : COCO-2017 training
Original paper : Image-to-Image Translation with Conditional Adversarial Networks
- Pre-trained the Generator separately:
- Used U-Net with a pretrained (on ImageNet) Resnet model as its backbone and then pretrained the whole generator model on the dataset.
- Tried with Resnet18 and Resnet34 models.
- Tried different Image Augmentations: both train-time (from Albumentation library) and test-time (from TTA library)
- Tried different optimisers for Generator and Discriminator.
- Training for more epochs.
- Hyper-parameter optimisations
- Create a web application and deploy the model over it.