ExampleResearchProject

My resarch project initial point.

New dataset work flow

1. Create a new file lib/dataset/[NEW_DATASET].py.
2. Implement the dataset class in lib/dataset/[NEW_DATASET].py.
3. Update lib/dataset/__init__.py.

The implemented dataset should return a dictionary as the batch for training. All key in the batch is preserved for the network.

See lib/dataset/dataset_example.py for an example.

New model work flow:

1. Create a new file lib/model/[NEW_MODEL].py.
2. Implement the network in lib/model/[NEW_MODEL].py.

The implemented network should have a member function def compute_losses(self, batch: dict) -> dict: where the batch is implemented by yourself in lib/dataset/[NEW_DATASET].py. The compute_losses function should return a dictionary where all element will be accumulate and log at each epoch. The key total' will be backprob in train.py`.

See lib/model/Example.py for an example.

New experiment work flow

1. Copy config/ExampleTask/example.yaml and alter the field to match your implemented dataset/model and all other training detail.
2. python train.py --cfg [PATH/TO/YOUR.yaml]

Run python train.py --cfg config/ExampleTask/example.yaml for an example.