cehrgpt 0.0.2__py3-none-any.whl → 0.1.0__py3-none-any.whl

cehrgpt/data/hf_cehrgpt_dataset_mapping.py

@@ -1,5 +1,5 @@
 import datetime
-from typing import Any, Dict
+from typing import Any, Dict, Generator, Optional
 
 import numpy as np
 import pandas as pd
@@ -8,8 +8,12 @@ from cehrbert.data_generators.hf_data_generator.hf_dataset_mapping import (
     INPATIENT_VISIT_TYPE_CODES,
     INPATIENT_VISIT_TYPES,
     DatasetMapping,
+    VisitObject,
+    get_value,
+    has_events_and_get_events,
     replace_escape_chars,
 )
+from cehrbert.med_extension.schema_extension import Event
 from cehrbert.runners.hf_runner_argument_dataclass import DataTrainingArguments
 from cehrbert_data.const.common import NA
 from cehrbert_data.decorators.patient_event_decorator_base import get_att_function
@@ -32,7 +36,6 @@ class MedToCehrGPTDatasetMapping(DatasetMapping):
     def __init__(
         self,
         data_args: DataTrainingArguments,
-        is_pretraining: bool = True,
         include_inpatient_hour_token: bool = True,
     ):
         self._time_token_function = get_att_function(data_args.att_function_type)
@@ -41,7 +44,6 @@ class MedToCehrGPTDatasetMapping(DatasetMapping):
             data_args.inpatient_att_function_type
         )
         self._include_demographic_prompt = data_args.include_demographic_prompt
-        self._is_pretraining = is_pretraining
         self._include_inpatient_hour_token = include_inpatient_hour_token
 
     """
@@ -57,14 +59,7 @@ class MedToCehrGPTDatasetMapping(DatasetMapping):
     """
 
     def remove_columns(self):
-        if self._is_pretraining:
-            return ["visits", "birth_datetime", "index_date"]
-        else:
-            return [
-                "visits",
-                "birth_datetime",
-                "visit_concept_ids",
-            ]
+        return ["patient_id", "visits", "birth_datetime"]
 
     @staticmethod
     def _update_cehrgpt_record(
@@ -99,38 +94,45 @@ class MedToCehrGPTDatasetMapping(DatasetMapping):
             birth_datetime = birth_datetime.to_pydatetime()
         gender = record["gender"]
         race = record["race"]
+        visits = record["visits"]
+        # This indicates this is columnar format
+        if isinstance(visits, dict):
+            visits = sorted(self.convert_visit_columnar_to_python(visits))
+        else:
+            visits = sorted(visits, key=lambda _: get_value(_, "visit_start_datetime"))
 
         # Add the demographic tokens
-        first_visit = record["visits"][0]
-        year_str = f'year:{str(first_visit["visit_start_datetime"].year)}'
-        age_str = f'age:{str(relativedelta(first_visit["visit_start_datetime"], birth_datetime).years)}'
+        first_visit = visits[0]
+        first_visit_start_datetime: datetime.datetime = get_value(
+            first_visit, "visit_start_datetime"
+        )
+        year_str = f"year:{str(first_visit_start_datetime.year)}"
+        age_str = f"age:{str(relativedelta(first_visit_start_datetime, birth_datetime).years)}"
         self._update_cehrgpt_record(cehrgpt_record, year_str)
         self._update_cehrgpt_record(cehrgpt_record, age_str)
         self._update_cehrgpt_record(cehrgpt_record, gender)
         self._update_cehrgpt_record(cehrgpt_record, race)
 
         # Use a data cursor to keep track of time
-        date_cursor = None
-
-        # Loop through all the visits excluding the first event containing the demographics
-        for i, visit in enumerate(
-            sorted(record["visits"], key=lambda e: e["visit_start_datetime"])
-        ):
-
-            events = visit["events"]
-
-            # Skip this visit if the number measurements in the event is zero
-            if events is None or len(events) == 0:
+        datetime_cursor: Optional[datetime.datetime] = None
+        visit: VisitObject
+        # Loop through all the visits
+        for i, visit in enumerate(visits):
+            events: Generator[Event, None, None] = get_value(visit, "events")
+            has_events, events = has_events_and_get_events(events)
+            if not has_events:
                 continue
 
-            visit_start_datetime = visit["visit_start_datetime"]
-            time_delta = (
-                (visit_start_datetime - date_cursor).days if date_cursor else None
+            visit_start_datetime: datetime.datetime = get_value(
+                visit, "visit_start_datetime"
+            )
+            # If visit_end_datetime is populated for the inpatient visit, we update the datetime_cursor
+            visit_end_datetime: Optional[datetime.datetime] = get_value(
+                visit, "visit_end_datetime"
             )
-            date_cursor = visit_start_datetime
 
             # We assume the first measurement to be the visit type of the current visit
-            visit_type = visit["visit_type"]
+            visit_type = get_value(visit, "visit_type")
             is_er_or_inpatient = (
                 visit_type in INPATIENT_VISIT_TYPES
                 or visit_type in INPATIENT_VISIT_TYPE_CODES
@@ -138,21 +140,15 @@ class MedToCehrGPTDatasetMapping(DatasetMapping):
             )
 
             # Add artificial time tokens to the patient timeline if timedelta exists
-            if time_delta is not None:
+            if datetime_cursor is not None:
+                time_delta = max((visit_start_datetime - datetime_cursor).days, 0)
                 # This generates an artificial time token depending on the choice of the time token functions
                 self._update_cehrgpt_record(
                     cehrgpt_record,
                     code=self._time_token_function(time_delta),
                 )
 
-            # Add the VS token to the patient timeline to mark the start of a visit
-            relativedelta(visit["visit_start_datetime"], birth_datetime).years
-            # Calculate the week number since the epoch time
-            date = (
-                visit["visit_start_datetime"]
-                - datetime.datetime(year=1970, month=1, day=1)
-            ).days // 7
-
+            datetime_cursor = visit_start_datetime
             # Add a [VS] token
             self._update_cehrgpt_record(
                 cehrgpt_record,
@@ -163,11 +159,22 @@ class MedToCehrGPTDatasetMapping(DatasetMapping):
                 cehrgpt_record,
                 code=visit_type,
             )
+            # We need to insert an inpatient hour token right after the visit type, we calculate the hour interval
+            # with respect to the midnight of the day
+            if is_er_or_inpatient and self._include_inpatient_hour_token:
+                if datetime_cursor.hour > 0:
+                    # This generates an artificial time token depending on the choice of the time token functions
+                    self._update_cehrgpt_record(
+                        cehrgpt_record,
+                        code=f"i-H{datetime_cursor.hour}",
+                    )
+
             # Keep track of the existing outpatient events, we don't want to add them again
-            existing_outpatient_events = list()
+            existing_duplicate_events = list()
             for e in events:
                 # If the event doesn't have a time stamp, we skip it
-                if not e["time"]:
+                event_time: datetime.datetime = e["time"]
+                if not event_time:
                     continue
 
                 # If numeric_value exists, this is a concept/value tuple, we indicate this using a concept_value_mask
@@ -181,40 +188,48 @@ class MedToCehrGPTDatasetMapping(DatasetMapping):
                 is_numeric_type = int(numeric_value is not None)
                 code = replace_escape_chars(e["code"])
 
+                # Create the event identity
+                event_identity = (
+                    (event_time, code, text_value, unit)
+                    if is_er_or_inpatient
+                    else (event_time.date(), code, text_value, unit)
+                )
+
                 # Add a medical token to the patient timeline
                 # If this is an inpatient visit, we use the event time stamps to calculate age and date
                 # because the patient can stay in the hospital for a period of time.
                 if is_er_or_inpatient:
-                    # Calculate the week number since the epoch time
-                    date = (
-                        e["time"] - datetime.datetime(year=1970, month=1, day=1)
-                    ).days // 7
                     # Calculate the time diff in days w.r.t the previous measurement
-                    meas_time_diff = (e["time"] - date_cursor).days
-                    # Update the date_cursor if the time diff between two neighboring measurements is greater than and
+                    time_diff_days = (event_time - datetime_cursor).days
+                    # Update the datetime_cursor if the time diff between two neighboring measurements is greater than and
                     # equal to 1 day
-                    if meas_time_diff > 0:
-                        date_cursor = e["time"]
-                        if self._inpatient_time_token_function:
+                    if self._inpatient_time_token_function and time_diff_days > 0:
+                        # This generates an artificial time token depending on the choice of the time token functions
+                        self._update_cehrgpt_record(
+                            cehrgpt_record,
+                            code=f"i-{self._inpatient_time_token_function(time_diff_days)}",
+                        )
+
+                    if self._include_inpatient_hour_token:
+                        # if the time difference in days is greater than 0, we calculate the hour interval
+                        # with respect to the midnight of the day
+                        time_diff_hours = (
+                            event_time.hour
+                            if time_diff_days > 0
+                            else int(
+                                (event_time - datetime_cursor).total_seconds() // 3600
+                            )
+                        )
+
+                        if time_diff_hours > 0:
                             # This generates an artificial time token depending on the choice of the time token functions
                             self._update_cehrgpt_record(
                                 cehrgpt_record,
-                                code=f"i-{self._inpatient_time_token_function(meas_time_diff)}",
+                                code=f"i-H{time_diff_hours}",
                             )
-                else:
-                    # For outpatient visits, we use the visit time stamp to calculate age and time because we assume
-                    # the outpatient visits start and end on the same day.
-                    # We check whether the date/code/value combination already exists in the existing events
-                    # If they exist, we do not add them to the patient timeline for outpatient visits.
-                    if (
-                        date,
-                        code,
-                        numeric_value,
-                        text_value,
-                        concept_value_mask,
-                        numeric_value,
-                    ) in existing_outpatient_events:
-                        continue
+
+                if event_identity in existing_duplicate_events:
+                    continue
 
                 self._update_cehrgpt_record(
                     cehrgpt_record,
@@ -227,33 +242,27 @@ class MedToCehrGPTDatasetMapping(DatasetMapping):
                     ),
                     is_numeric_type=is_numeric_type,
                 )
-                existing_outpatient_events.append(
-                    (
-                        date,
-                        code,
-                        numeric_value,
-                        text_value,
-                        concept_value_mask,
-                        numeric_value,
-                    )
-                )
+                existing_duplicate_events.append(event_identity)
+                # we only want to update the time stamp when data_cursor is less than the event time
+                if datetime_cursor < event_time or datetime_cursor is None:
+                    datetime_cursor = event_time
+                    # We need to bound the datetime_cursor if the current visit is an admission type of visit
+                    # as the associated events could be generated after the visits are complete
+                    if is_er_or_inpatient and visit_end_datetime is not None:
+                        datetime_cursor = min(datetime_cursor, visit_end_datetime)
 
             # For inpatient or ER visits, we want to discharge_facility to the end of the visit
             if is_er_or_inpatient:
-                # If visit_end_datetime is populated for the inpatient visit, we update the date_cursor
-                visit_end_datetime = visit.get("visit_end_datetime", None)
-                if visit_end_datetime:
-                    date_cursor = visit_end_datetime
+                # If visit_end_datetime is populated for the inpatient visit, we update the datetime_cursor
+                if visit_end_datetime is not None:
+                    datetime_cursor = visit_end_datetime
 
                 if self._include_auxiliary_token:
                     # Reuse the age and date calculated for the last event in the patient timeline for the discharge
                     # facility event
-                    discharge_facility = (
-                        visit["discharge_facility"]
-                        if ("discharge_facility" in visit)
-                        and visit["discharge_facility"]
-                        else "0"
-                    )
+                    discharge_facility = get_value(visit, "discharge_facility")
+                    if not discharge_facility:
+                        discharge_facility = "0"
 
                     self._update_cehrgpt_record(
                         cehrgpt_record,
@@ -273,11 +282,13 @@ class MedToCehrGPTDatasetMapping(DatasetMapping):
 
         # Add some count information for this sequence
         cehrgpt_record["num_of_concepts"] = len(cehrgpt_record["concept_ids"])
-        cehrgpt_record["num_of_visits"] = len(record["visits"])
+        cehrgpt_record["num_of_visits"] = len(visits)
 
-        if "label" in record:
+        if record.get("index_date", None):
+            cehrgpt_record["index_date"] = record["index_date"]
+        if record.get("label", None):
             cehrgpt_record["label"] = record["label"]
-        if "age_at_index" in record:
+        if record.get("age_at_index", None):
             cehrgpt_record["age_at_index"] = record["age_at_index"]
 
         return cehrgpt_record

cehrgpt/data/sample_packing_sampler.py

@@ -0,0 +1,151 @@
+from typing import Iterator, List, Optional
+
+import torch
+import torch.distributed as dist
+from torch.utils.data import Sampler
+from transformers import logging
+
+LOG = logging.get_logger("transformers")
+
+
+class SamplePlacerHolder:
+    def __init__(self):
+        self.epoch = 0
+
+    def set_epoch(self, epoch):
+        self.epoch = epoch
+
+
+class SamplePackingBatchSampler(Sampler[List[int]]):
+    """
+    A batch sampler that creates batches by packing samples together.
+
+    to maximize GPU utilization, ensuring the total tokens per batch
+    doesn't exceed max_tokens.
+    """
+
+    def __init__(
+        self,
+        lengths: List[int],
+        max_tokens_per_batch: int,
+        max_position_embeddings: int,
+        num_replicas: Optional[int] = None,
+        rank: Optional[int] = None,
+        seed: int = 0,
+        drop_last: bool = False,
+    ):
+        """
+        Args:
+
+            lengths: List of sequence lengths for each sample
+            max_tokens: Maximum number of tokens in a batch
+            drop_last: Whether to drop the last incomplete batch
+        """
+        super().__init__()
+
+        if num_replicas is None:
+            if dist.is_available() and dist.is_initialized():
+                num_replicas = dist.get_world_size()
+                LOG.info(
+                    "torch.distributed is initialized and there are %s of replicas",
+                    num_replicas,
+                )
+            else:
+                num_replicas = 1
+                LOG.info(
+                    "torch.dist is not initialized and therefore default to 1 for num_replicas"
+                )
+
+        if rank is None:
+            if dist.is_available() and dist.is_initialized():
+                rank = dist.get_rank()
+                LOG.info(
+                    "torch.distributed is initialized and the current rank is %s", rank
+                )
+            else:
+                rank = 0
+                LOG.info(
+                    "torch.distributed is not initialized and therefore default to 0 for rank"
+                )
+
+        if not (0 <= rank < num_replicas):
+            raise ValueError(
+                f"Invalid rank {rank}, rank should be in the interval [0, {num_replicas - 1}]"
+            )
+
+        self.lengths = lengths
+        self.max_tokens_per_batch = max_tokens_per_batch
+        self.max_position_embeddings = max_position_embeddings
+        self.num_replicas = num_replicas
+        self.rank = rank
+        self.seed = seed
+        self.drop_last = drop_last
+        # Trainer https://github.com/huggingface/transformers/blame/main/src/transformers/trainer.py#L2470
+        # http://github.com/huggingface/accelerate/blob/v0.31.0/src/accelerate/data_loader.py#L482
+        # the huggingface trainer will call the accelerate.data_loader.DataLoaderShard.set_epoch,
+        # which will call batch_sampler.sample.set_epoch
+        self.sampler = SamplePlacerHolder()
+
+    def __iter__(self) -> Iterator[List[int]]:
+
+        # deterministically shuffle based on epoch and seed
+        g = torch.Generator()
+        g.manual_seed(self.seed + self.sampler.epoch)
+        indices = torch.randperm(len(self.lengths), generator=g).tolist()
+
+        # Partition indices for this rank
+        indices = indices[self.rank :: self.num_replicas]
+
+        batch = []
+        current_batch_tokens = 0
+
+        for idx in indices:
+            # We take the minimum of the two because each sequence will be truncated to fit
+            # the context window of the model
+            sample_length = min(self.lengths[idx], self.max_position_embeddings)
+            # If adding this sample would exceed max_tokens_per_batch, yield the current batch
+            if (
+                current_batch_tokens + sample_length + 2 > self.max_tokens_per_batch
+                and batch
+            ):
+                yield batch
+                batch = []
+                current_batch_tokens = 0
+
+            # Add the sample to the current batch
+            batch.append(idx)
+            # plus extract one for the [END] and [PAD] tokens to separate samples
+            current_batch_tokens += sample_length + 2
+
+        # Yield the last batch if it's not empty and we're not dropping it
+        if batch and not self.drop_last:
+            yield batch
+
+    def __len__(self) -> int:
+        """
+        Estimates the number of batches that will be generated.
+
+        This is an approximation since the exact number depends on the specific
+        sequence lengths and their order.
+        """
+        if len(self.lengths) == 0:
+            return 0
+
+        # We need to truncate the lengths due to the context window limit imposed by the model
+        truncated_lengths = [
+            min(self.max_position_embeddings, length + 2) for length in self.lengths
+        ]
+
+        # Calculate average sequence length
+        avg_seq_length = sum(truncated_lengths) // len(truncated_lengths)
+
+        # Estimate average number of sequences per batch
+        seqs_per_batch = self.max_tokens_per_batch // avg_seq_length
+
+        # Estimate total number of batches
+        if self.drop_last:
+            # If dropping last incomplete batch
+            return len(truncated_lengths) // seqs_per_batch * self.num_replicas
+        else:
+            # If keeping last incomplete batch, ensure at least 1 batch
+            return max(1, len(truncated_lengths) // seqs_per_batch) * self.num_replicas

cehrgpt/generation/generate_batch_hf_gpt_sequence.py

@@ -93,9 +93,9 @@ def generate_single_batch(
             temperature=temperature,
             top_p=top_p,
             top_k=top_k,
-            bos_token_id=tokenizer.end_token_id,
-            eos_token_id=tokenizer.end_token_id,
-            pad_token_id=tokenizer.pad_token_id,
+            bos_token_id=model.generation_config.bos_token_id,
+            eos_token_id=model.generation_config.eos_token_id,
+            pad_token_id=model.generation_config.pad_token_id,
             do_sample=True,
             use_cache=True,
             return_dict_in_generate=True,
@@ -150,15 +150,11 @@ def main(args):
             attn_implementation=(
                 "flash_attention_2" if is_flash_attn_2_available() else "eager"
             ),
-            torch_dtype=(
-                torch.bfloat16
-                if is_flash_attn_2_available() and args.use_bfloat16
-                else torch.float32
-            ),
         )
         .eval()
         .to(device)
     )
+
     cehrgpt_model.generation_config.pad_token_id = cehrgpt_tokenizer.pad_token_id
     cehrgpt_model.generation_config.eos_token_id = cehrgpt_tokenizer.end_token_id
     cehrgpt_model.generation_config.bos_token_id = cehrgpt_tokenizer.end_token_id
@@ -192,6 +188,7 @@ def main(args):
     LOG.info(f"Top P {args.top_p}")
     LOG.info(f"Top K {args.top_k}")
     LOG.info(f"Loading demographic_info at {args.demographic_data_path}")
+    LOG.info(f"MEDS format: {args.meds_format}")
 
     dataset = load_parquet_as_dataset(args.demographic_data_path)
     total_rows = len(dataset)
@@ -199,6 +196,7 @@ def main(args):
     num_of_batches = args.num_of_patients // args.batch_size + 1
     sequence_to_flush = []
     current_person_id = 1
+    prompt_size = 2 if args.meds_format else START_TOKEN_SIZE
     for i in range(num_of_batches):
         LOG.info(f"{datetime.datetime.now()}: Batch {i} started")
 
@@ -215,7 +213,7 @@ def main(args):
                     <= max_seq_allowed
                 ):
                     random_prompts.append(
-                        cehrgpt_tokenizer.encode(row["concept_ids"][:START_TOKEN_SIZE])
+                        cehrgpt_tokenizer.encode(row["concept_ids"][:prompt_size])
                     )
                 iter += 1
                 if not random_prompts and iter > 10:
@@ -326,6 +324,11 @@ def create_arg_parser():
         dest="drop_long_sequences",
         action="store_true",
     )
+    base_arg_parser.add_argument(
+        "--meds_format",
+        dest="meds_format",
+        action="store_true",
+    )
     return base_arg_parser
 
 

cehrgpt/models/config.py

@@ -133,14 +133,17 @@ class CEHRGPTConfig(PretrainedConfig):
         n_pretrained_embeddings_layers=2,
         pretrained_embedding_dim=768,
         pretrained_token_ids: List[int] = None,
+        next_token_prediction_loss_weight=1.0,
         time_token_loss_weight=1.0,
         time_to_visit_loss_weight=1.0,
         causal_sfm=False,
         demographics_size=4,
         lab_token_penalty=False,
         lab_token_loss_weight=0.9,
+        value_prediction_loss_weight=1.0,
         entropy_penalty=False,
         entropy_penalty_alpha=0.01,
+        sample_packing_max_positions=None,
         **kwargs,
     ):
         if token_to_time_token_mapping is None:
@@ -150,6 +153,11 @@ class CEHRGPTConfig(PretrainedConfig):
         self.vocab_size = vocab_size
         self.time_token_vocab_size = time_token_vocab_size
         self.n_positions = n_positions
+        self.sample_packing_max_positions = (
+            sample_packing_max_positions
+            if sample_packing_max_positions
+            else n_positions
+        )
         self.n_embd = n_embd
         self.n_layer = n_layer
         self.n_head = n_head
@@ -178,6 +186,7 @@ class CEHRGPTConfig(PretrainedConfig):
         self.include_values = include_values
         self.value_vocab_size = value_vocab_size
 
+        self.next_token_prediction_loss_weight = next_token_prediction_loss_weight
         self.include_ttv_prediction = include_ttv_prediction
         self.use_sub_time_tokenization = use_sub_time_tokenization
         self._token_to_time_token_mapping = token_to_time_token_mapping
@@ -195,6 +204,7 @@ class CEHRGPTConfig(PretrainedConfig):
         self.lab_token_loss_weight = lab_token_loss_weight
         self.entropy_penalty = entropy_penalty
         self.entropy_penalty_alpha = entropy_penalty_alpha
+        self.value_prediction_loss_weight = value_prediction_loss_weight
 
         kwargs["tie_word_embeddings"] = not use_pretrained_embeddings