This project was done and tested using Python 3.7 so it should work for Python 3.7+.
To create a conda virtual environment that would work, run
conda env create -f environment.yaml
Custom dataset would need to be implemented by:
- Creating a class deriving from
torch.utils.data.Dataset; - Modifying
datasets/__init__.pyto include your new class.
To train, run
python train.py --config path/to/config/file [--gpus gpu_id] [--debug]
Arguments:
--config: path to configuration file
--gpus: gpu id to be used
--debug: to save the weights or not
For example:
python train.py --config configs/train/debug_ircad.yaml --gpus 0 --debug
Modify the default configuration file (YAML format) to suit your need, the properties' name should be self-explanatory.
All the result will be stored in the runs folder in separate subfolders, one for each run. The result consists of the log file for Tensorboard, the network pretrained models (best metrics, best loss, and the latest iteration).
This project uses Tensorboard to plot training graph. To see it, run
tensorboard --logdir=logs
and access using the announced port (default is 6006, e.g http://localhost:6006).
The .pth files contains a dictionary:
{
'epoch': the epoch of the training where the weight is saved
'model_state_dict': model state dict (use model.load_state_dict to load)
'optimizer_state_dict': optimizer state dict (use opt.load_state_dict to load)
'log': full logs of that run
'config': full configuration of that run
}
To test a pretrained model, run
python eval.py --weight path/to/pretrained/model [--gpus 0] [--vis] [--output visualization/dir]
Arguments:
--weight: path to pretrained model
--gpus: gpu id to be used
--vis: whether to save visualization or not
--output: the directory where the visualization will be stored
For example:
python eval.py --weight runs/test_best_acc.pth --gpus 0 --vis --output vis/best_acc