mmcdermott/EventStreamGPT

Issues

• Setting min_seq_len=1 in PytorchDatasetConfig with task dataframes leads to ragged tensors index error

  #118 opened 2 months ago by juancq
  4

• Running out of memory during build_event_and_measurement_dfs and polars dataset

  #54 opened 2 months ago by juancq
  11

• build_dataset fails when aggregating timestamps into buckets

  #119 opened 2 months ago by juancq
  0

• Setting min_seq_len in PytorchDatasetConfig to < 2 with task dataframes leads to keyerror

  #109 opened 2 months ago by juancq
  5

• Subject ID splits can get messed up if subjects are not simple int types.

  #114 opened 4 months ago by mmcdermott
  4

• Could not override 'config.task_specific_params.pooling_method' when lauching wandb agent

  #108 opened 5 months ago by Rhett-Ying
  2

• Processing Synthetic Data with ESGPT

  #113 opened 4 months ago by sujaybanerjee
  1

• Support control for how events are aggregated/combined into compound events

  #111 opened 4 months ago by juancq
  1

• end_time in task dataframes is not counted as an event (right open interval)

  #110 opened 5 months ago by juancq
  1

• Task `to_int_index` does not pass tests with more recent versions of polars

  #112 opened 4 months ago by mmcdermott
  0

• nvalid series dtype: expected `Utf8`, got `datetime[ns]`

  #106 opened 5 months ago by rvandewater
  2

• Nested ragged tensors integration does not respect `seq_padding_side`

  #104 opened 6 months ago by mmcdermott
  1

• Getting started steps

  #96 opened 6 months ago by rvandewater
  1

• sklearn baseline pipe code should be using logger

  #95 opened 6 months ago by mmcdermott
  0

• PytorchDataset doesn't have a way to extract subject_ids, start_time, and end_time all together

  #75 opened a year ago by pargaw
  0

• Should be possible to resume pre-training and evaluate a pre-trained model post-hoc

  #91 opened 10 months ago by mmcdermott
  0

• May be able to make temporal loss more stable by directly computing log-prob from pre-activation function outputs

  #93 opened 8 months ago by mmcdermott
  2

• CATEGORICAL_ONLY v. NUMERICAL_ONLY v. CATEGORICAL_AND_NUMERICAL is confusing.

  #13 opened a year ago by mmcdermott
  0

• should use pin_memory to speed up dataloaders

  #6 opened a year ago by mmcdermott
  1

• ID_map should be saved in dataset construction

  #8 opened a year ago by mmcdermott
  0

• Should enable retrieval metrics for classification as well

  #9 opened a year ago by mmcdermott
  0

• Default model should be simplified and use the built-in PyTorch transformer layers (though still support structured attention)

  #10 opened a year ago by mmcdermott
  1

• Torch.compile doesn't work for some reason.

  #11 opened a year ago by mmcdermott
  0

• Dataset storage to disk should not rely on pickle.

  #14 opened a year ago by mmcdermott
  0

• Test coverage needs to be expanded and interface simplified

  #15 opened a year ago by mmcdermott
  0

• Upgrade to python 3.11 and eliminate the custom StrEnum class.

  #23 opened a year ago by mmcdermott
  0

• Possibly replace `JSONableMixin` with `OmegaConf` throughout.

  #24 opened a year ago by mmcdermott
  0

• DataLoader with num_workers > 0 increases memory consumption over time

  #73 opened a year ago by juancq
  2

• Event type vocabulary should be filtered to UNKs too.

  #43 opened a year ago by mmcdermott
  0

• Should allow use of NLL-Beta

  #26 opened a year ago by mmcdermott
  0

• conda environment file fails

  #32 opened 8 months ago by juancq
  6

• Misspelled measurement_configs in transformer/config.py

  #72 opened a year ago by juancq
  1

• Pretrain crashes during collating static and dynamic data in pytorch_dataset

  #66 opened a year ago by juancq
  4

• Support same variable name across different input sources

  #53 opened a year ago by juancq
  6

• Modeling Configs Should be Re-arranged for Clarity, Better Defaults, and Ease-of-use

  #61 opened a year ago by mmcdermott
  0

• Metrics in weights and biases should use `/` not `_` to separate name parts.

  #60 opened a year ago by mmcdermott
  0

• Setting maximum number of threads Polars will use

  #47 opened a year ago by juancq
  1

• Setting train_subset_size to float generates ValidationError

  #52 opened a year ago by juancq
  1

• Add streaming flag for calls to polars collect

  #46 opened a year ago by juancq
  4

• Option of using pyarrow when writing parquet

  #49 opened a year ago by juancq
  0

• ESGPT Pipeline Permits Duplicate dynamic measurement rows

  #36 opened a year ago by mmcdermott
  1

• Task dataframe will only keep one task schema row per subject.

  #37 opened a year ago by mmcdermott
  1

• `num_samples` is overloaded in zero shot lightning module

  #40 opened a year ago by mmcdermott
  1

• Pipeline for the UNIVARIATE_REGRESSION and MULTIVARIATE_REGRESSION tasks

  #39 opened a year ago by sujeongim
  2

• self has no task_df member.

  #34 opened a year ago by mmcdermott
  1

• Vocabulary.describe doesn't appear to work if wrap_lines is True.

  #44 opened a year ago by mmcdermott
  0

• Vocabulary serialization into configs should not include observation frequencies.

  #42 opened a year ago by mmcdermott
  0

• Benchmark results over standard EHR tasks like diagnosis prediction

  #41 opened a year ago by avani17101
  3

• 'ExprArrayNameSpace' object has no attribute 'eval'

  #30 opened a year ago by Guitaricet
  2

• task_df interface in pytorch dataset should use polars.

  #12 opened a year ago by mmcdermott
  0