Improved compliance by removing creation of shards.json file and adding patient_splits.parquet file.

src/MEDS_transforms/transforms/tokenization.py

@@ -9,8 +9,6 @@
 columns of concern here thus are `patient_id`, `time`, `code`, `numeric_value`.
 """
 
-import json
-import random
 from pathlib import Path
 
 import hydra
@@ -19,7 +17,7 @@
 from omegaconf import DictConfig, OmegaConf
 
 from MEDS_transforms import PREPROCESS_CONFIG_YAML
-from MEDS_transforms.mapreduce.mapper import rwlock_wrap
+from MEDS_transforms.mapreduce.utils import rwlock_wrap, shard_iterator
 from MEDS_transforms.utils import hydra_loguru_init, write_lazyframe
 
 SECONDS_PER_MINUTE = 60.0
@@ -230,18 +228,17 @@ def main(cfg: DictConfig):
         f"Stage config:\n{OmegaConf.to_yaml(cfg.stage_cfg)}"
     )
 
-    input_dir = Path(cfg.stage_cfg.data_input_dir)
     output_dir = Path(cfg.stage_cfg.output_dir)
+    shards_single_output, include_only_train = shard_iterator(cfg)
 
-    shards = json.loads((Path(cfg.input_dir) / "splits.json").read_text())
+    if include_only_train:
+        raise ValueError("Not supported for this stage.")
 
-    patient_splits = list(shards.keys())
-    random.shuffle(patient_splits)
+    for in_fp, out_fp in shards_single_output:
+        sharded_path = out_fp.relative_to(output_dir)
 
-    for sp in patient_splits:
-        in_fp = input_dir / f"{sp}.parquet"
-        schema_out_fp = output_dir / "schemas" / f"{sp}.parquet"
-        event_seq_out_fp = output_dir / "event_seqs" / f"{sp}.parquet"
+        schema_out_fp = output_dir / "schemas" / sharded_path
+        event_seq_out_fp = output_dir / "event_seqs" / sharded_path
 
         logger.info(f"Tokenizing {str(in_fp.resolve())} into schemas at {str(schema_out_fp.resolve())}")
 

tests/transform_tester_base.py

@@ -7,6 +7,7 @@
 import json
 import os
 import tempfile
+from collections import defaultdict
 from io import StringIO
 from pathlib import Path
 
@@ -56,13 +57,19 @@
 
 # Test MEDS data (inputs)
 
-SPLITS = {
+SHARDS = {
     "train/0": [239684, 1195293],
     "train/1": [68729, 814703],
     "tuning/0": [754281],
     "held_out/0": [1500733],
 }
 
+SPLITS = {
+    "train": [239684, 1195293, 68729, 814703],
+    "tuning": [754281],
+    "held_out": [1500733],
+}
+
 MEDS_TRAIN_0 = """
 patient_id,time,code,numeric_value
 239684,,EYE_COLOR//BROWN,
@@ -292,6 +299,8 @@ def single_stage_transform_tester(
     do_pass_stage_name: bool = False,
     file_suffix: str = ".parquet",
     do_use_config_yaml: bool = False,
+    input_shards_map: dict[str, list[int]] | None = None,
+    input_splits_map: dict[str, list[int]] | None = None,
 ):
     with tempfile.TemporaryDirectory() as d:
         MEDS_dir = Path(d) / "MEDS_cohort"
@@ -306,9 +315,23 @@ def single_stage_transform_tester(
         MEDS_metadata_dir.mkdir(parents=True)
         cohort_dir.mkdir(parents=True)
 
-        # Write the splits
-        splits_fp = MEDS_dir / "splits.json"
-        splits_fp.write_text(json.dumps(SPLITS))
+        # Write the shards map
+        if input_shards_map is None:
+            input_shards_map = SHARDS
+
+        shards_fp = MEDS_metadata_dir / ".shards.json"
+        shards_fp.write_text(json.dumps(input_shards_map))
+
+        # Write the splits parquet file
+        if input_splits_map is None:
+            input_splits_map = SPLITS
+        input_splits_as_df = defaultdict(list)
+        for split_name, patient_ids in input_splits_map.items():
+            input_splits_as_df["patient_id"].extend(patient_ids)
+            input_splits_as_df["split"].extend([split_name] * len(patient_ids))
+        input_splits_df = pl.DataFrame(input_splits_as_df)
+        input_splits_fp = MEDS_metadata_dir / "patient_splits.parquet"
+        input_splits_df.write_parquet(input_splits_fp, use_pyarrow=True)
 
         if input_shards is None:
             input_shards = MEDS_SHARDS