To run the code, first install the requirements with
pip install -r requirements.txt
To train on SuMNIST dataset, run the following code on
python main.py --config configs/mnistsum_config.txt
To train on DiffSuMNist dataset, run the following code on
python main.py --config configs/diffmnistsum_config.txt
The MNIST dataset will be downloaded automatically to a folder specified in the config file datadir.
To view the logs from tensorboard, please run
tensorboard --logdir runs.
The experiments will be logged to tensorboard with the experiment name from the config expname.
The checkpoints will be logged to the basedir into the specific experiment folder.
The frequency of checkpoint saving and logging to tensorboard can be changed by setting
i_print and i_weight parameters.
To start the training from an existing checkpoint run the following
python main.py --config configs/mnistsum_config.txt --ckpt {ckpt_path}
The trained model can be tested by adding --test parameter and the checkpoint.
Example:
python main.py --config configs/mnistsum_config.txt --ckpt {ckpt_path} --test