This code provides Pytorch implementation of FCN architecture on synthetic dataset and real world dataset from scratch as well as finetuning. There are some custom utils functions for joint transformation, evaluation, and average meter.
Tested on Python 3.6.x and Keras 2.3.0 with TF backend version 1.14.0.
- Numpy
- Torchvision
- PyTorch
- Matplotlib
- Pillow
- Install the required dependencies:
pip install -r requirements.txtfcn.py: Model architecture
train_games.py: Train on games dataset
train_cityscapes.py: Train on cityscapes dataset
ft_cityscapes.py: Finetuning on cityscapes
eval_cityscapes.py: Evaluate on test dataset of cityscapes
cityscapes.py: create dataloader for cityscapes dataset
games_data.py: create dataloader for games dataset
To train on games from scratch
python train_games.py
To train on cityscapes from scratch
python train_cityscapes.py
To train on cityscapes for finetuning
python ft_cityscapes.py
Evaluate on cityscapes
python eval_cityscapes.py