Project based on Pytorch-Lightning
Experiments are written in Yaml format and stored in experiments folder
EXPERIMENT_NAME: "baseline" - You need to specify experiment name, this name will be used to store logs
DATA: ... - This block is used to specify what samples to train on and where the data stored. Moreover, BATCH_SIZE and validation size are needed to be specified
OPTIMIZER: ... - Optimizer configuration: from what package import, what optimizer to use and arguments to this optimizer
To run training process You need to run python main.py --experiment_cfg experiments/train.yaml
MODEL: ... - What model to use
SCHEDULER: ... - Parameters of scheduler to use.
CRITERION: ... - What loss to use to optimize
AUGMENTATION: ... - The training augmentations. The augmentation need to be written in package specified in PY
EARLY_STOPPING: ... - Parameters of early stopping process
CHECKPOINT: ... - Parameters of checkpointing process
EPOCHS: ... - The number of epochs
GPUS: ... - GPUS to train on
To start training process with baseline experiment config run :
python main.py --experiment_cfg experiments/baseline.yaml
###Convert checkpoint of Pytorch-Lightning
To convert checkpoint of Pytorch-Lightning run:
python convert_checkpoint.py --checkpoint path_to_checkpoint --config_file path_to_config
###Inference
Inference config has the following structure:
MODEL: ... - What model to use
TEST: ... - This block is used to specify what samples to test on and where the data stored
To run inference
python inference.py --config_file test.yaml