MetaDelta++: Improve Generalization of Few-shot System Through Multi-Scale Pretrained Models and Improved Training Strategies
News!
- Serve as a baseline adopted for the Cross-Domain MetaDL competition!
- We've removed all the multiprocessing and uncessary parts to make the logic nice and clean. For the full version, you can refer to the branch neurips2021 or aaai2021.
Please follow the official website to set-up environments.
Please follow the official website to run the codes.
For example, cd to the folder of the cd-metadl, and run the following command:
cd path/to/cd-metadl
python -m cdmetadl.run --input_data_dir=public_data --submission_dir=path/to/this/folder --output_dir_ingestion=ingestion_output --output_dir_scoring=scoring_output --verbose=False --overwrite_previous_results=True --test_tasks_per_dataset=10
Remember to replace the path/to/cd-metadl
and path/to/this/folder
to your settings.
@AutoMLGroup/Meta_Learners