You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
* implemented sample packing for cehrbert
* clamped concept_values between -10 and 10; fixed a bug when generating time_embeddings
* added compute_cehrbert_features.py for extract features
* removed the sort patient sequence mapping
* updated sample_packing_sampler.py
* updated the index_date and age_at_index data types in hf_dataset_collator.py
* Synchronized bert flash attention layer with the eager implementation
* removed torch_dtype when loading the model
* put back upad_input in src/cehrbert/models/hf_models/hf_cehrbert.py
* removed squeeze in compute_cehrbert_features
* switched to CehrBertForPreTraining for computing features
* added an option to get the features by averaging over the entire sequence for each sample
* added train_with_cehrbert_features
* added device to start_indices and end_indices
* swtiched to a vectorized implementation
* corrected the logic for extract CLS embeddings in sample packing
* implemented sample packing for cehrbert sample packing
* fixed the integration test
* adjusted finetuning model for sample packing
* do not use sample packing for running predicitons
* updated transformers version
* updated the integration tests
* fixed a data loading bug in streaming