scdataloader 1.9.1__py3-none-any.whl → 2.0.0__py3-none-any.whl

scdataloader/datamodule.py

@@ -1,4 +1,9 @@
+import multiprocessing as mp
 import os
+import random
+import time
+from concurrent.futures import ProcessPoolExecutor, as_completed
+from functools import partial
 from typing import Optional, Sequence, Union
 
 import lamindb as ln
@@ -13,10 +18,11 @@ from torch.utils.data.sampler import (
     SubsetRandomSampler,
     WeightedRandomSampler,
 )
+from tqdm import tqdm
 
 from .collator import Collator
 from .data import Dataset
-from .utils import getBiomartTable, slurm_restart_count
+from .utils import fileToList, getBiomartTable, listToFile
 
 FILE_DIR = os.path.dirname(os.path.abspath(__file__))
 
@@ -26,9 +32,8 @@ class DataModule(L.LightningDataModule):
         self,
         collection_name: str,
         clss_to_weight: list = ["organism_ontology_term_id"],
-        organisms: list = ["NCBITaxon:9606"],
         weight_scaler: int = 10,
-        train_oversampling_per_epoch: float = 0.1,
+        n_samples_per_epoch: int = 2_000_000,
         validation_split: float = 0.2,
         test_split: float = 0,
         gene_embeddings: str = "",
@@ -43,7 +48,7 @@ class DataModule(L.LightningDataModule):
         max_len: int = 1000,
         add_zero_genes: int = 100,
         replacement: bool = True,
-        do_gene_pos: Union[bool, str] = True,
+        do_gene_pos: str = "",
         tp_name: Optional[str] = None,  # "heat_diff"
         assays_to_drop: list = [
             # "EFO:0008853", #patch seq
@@ -53,7 +58,10 @@ class DataModule(L.LightningDataModule):
         ],
         metacell_mode: float = 0.0,
         get_knn_cells: bool = False,
-        modify_seed_on_requeue: bool = True,
+        store_location: str = None,
+        force_recompute_indices: bool = False,
+        sampler_workers: int = None,
+        sampler_chunk_size: int = None,
         **kwargs,
     ):
         """
@@ -65,9 +73,8 @@ class DataModule(L.LightningDataModule):
 
         Args:
             collection_name (str): The lamindb collection to be used.
-            organisms (list, optional): The organisms to include in the dataset. Defaults to ["NCBITaxon:9606"].
             weight_scaler (int, optional): how much more you will see the most present vs less present category.
-            train_oversampling_per_epoch (float, optional): The proportion of the dataset to include in the training set for each epoch. Defaults to 0.1.
+            n_samples_per_epoch (int, optional): The number of samples to include in the training set for each epoch. Defaults to 2_000_000.
             validation_split (float, optional): The proportion of the dataset to include in the validation split. Defaults to 0.2.
             test_split (float, optional): The proportion of the dataset to include in the test split. Defaults to 0.
                 it will use a full dataset and will round to the nearest dataset's cell count.
@@ -88,61 +95,38 @@ class DataModule(L.LightningDataModule):
             hierarchical_clss (list, optional): List of hierarchical classes. Defaults to [].
             metacell_mode (float, optional): The probability of using metacell mode. Defaults to 0.0.
             clss_to_predict (list, optional): List of classes to predict. Defaults to ["organism_ontology_term_id"].
-            modify_seed_on_requeue (bool, optional): Whether to modify the seed on requeue. Defaults to True.
             get_knn_cells (bool, optional): Whether to get the k-nearest neighbors of each queried cells. Defaults to False.
+            store_location (str, optional): The location to store the sampler indices. Defaults to None.
+            force_recompute_indices (bool, optional): Whether to force recompute the sampler indices. Defaults to False.
+            sampler_workers (int, optional): The number of workers to use for the sampler. Defaults to None (auto-determined).
+            sampler_chunk_size (int, optional): The size of the chunks to use for the sampler. Defaults to None (auto-determined).
             **kwargs: Additional keyword arguments passed to the pytorch DataLoader.
             see @file data.py and @file collator.py for more details about some of the parameters
         """
-        if collection_name is not None:
-            mdataset = Dataset(
-                ln.Collection.filter(name=collection_name).first(),
-                organisms=organisms,
-                clss_to_predict=clss_to_predict,
-                hierarchical_clss=hierarchical_clss,
-                metacell_mode=metacell_mode,
-                get_knn_cells=get_knn_cells,
+        if "organism_ontology_term_id" not in clss_to_predict:
+            raise ValueError(
+                "need 'organism_ontology_term_id' in the set of classes at least"
             )
+        mdataset = Dataset(
+            ln.Collection.filter(name=collection_name, is_latest=True).first(),
+            clss_to_predict=clss_to_predict,
+            hierarchical_clss=hierarchical_clss,
+            metacell_mode=metacell_mode,
+            get_knn_cells=get_knn_cells,
+            store_location=store_location,
+            force_recompute_indices=force_recompute_indices,
+        )
         # and location
         self.metacell_mode = bool(metacell_mode)
         self.gene_pos = None
         self.collection_name = collection_name
         if do_gene_pos:
-            if type(do_gene_pos) is str:
-                print("seeing a string: loading gene positions as biomart parquet file")
-                biomart = pd.read_parquet(do_gene_pos)
-            else:
-                # and annotations
-                if organisms != ["NCBITaxon:9606"]:
-                    raise ValueError(
-                        "need to provide your own table as this automated function only works for humans for now"
-                    )
-                biomart = getBiomartTable(
-                    attributes=["start_position", "chromosome_name"],
-                    useCache=True,
-                ).set_index("ensembl_gene_id")
-                biomart = biomart.loc[~biomart.index.duplicated(keep="first")]
-                biomart = biomart.sort_values(by=["chromosome_name", "start_position"])
-                c = []
-                i = 0
-                prev_position = -100000
-                prev_chromosome = None
-                for _, r in biomart.iterrows():
-                    if (
-                        r["chromosome_name"] != prev_chromosome
-                        or r["start_position"] - prev_position > gene_position_tolerance
-                    ):
-                        i += 1
-                    c.append(i)
-                    prev_position = r["start_position"]
-                    prev_chromosome = r["chromosome_name"]
-                print(f"reduced the size to {len(set(c)) / len(biomart)}")
-                biomart["pos"] = c
+            biomart = pd.read_parquet(do_gene_pos)
             mdataset.genedf = mdataset.genedf.join(biomart, how="inner")
             self.gene_pos = mdataset.genedf["pos"].astype(int).tolist()
-
         if gene_embeddings != "":
             mdataset.genedf = mdataset.genedf.join(
-                pd.read_parquet(gene_embeddings), how="inner"
+                pd.read_parquet(gene_embeddings).loc[:, :2], how="inner"
             )
             if do_gene_pos:
                 self.gene_pos = mdataset.genedf["pos"].tolist()
@@ -151,7 +135,7 @@ class DataModule(L.LightningDataModule):
         # we might want to not introduce zeros and
         if use_default_col:
             kwargs["collate_fn"] = Collator(
-                organisms=organisms,
+                organisms=mdataset.organisms,
                 how=how,
                 valid_genes=mdataset.genedf.index.tolist(),
                 max_len=max_len,
@@ -159,7 +143,7 @@ class DataModule(L.LightningDataModule):
                 org_to_id=mdataset.encoder[organism_name],
                 tp_name=tp_name,
                 organism_name=organism_name,
-                class_names=clss_to_predict,
+                class_names=list(self.classes.keys()),
             )
         self.validation_split = validation_split
         self.test_split = test_split
@@ -171,16 +155,19 @@ class DataModule(L.LightningDataModule):
         self.assays_to_drop = assays_to_drop
         self.n_samples = len(mdataset)
         self.weight_scaler = weight_scaler
-        self.train_oversampling_per_epoch = train_oversampling_per_epoch
+        self.n_samples_per_epoch = n_samples_per_epoch
         self.clss_to_weight = clss_to_weight
         self.train_weights = None
         self.train_labels = None
-        self.modify_seed_on_requeue = modify_seed_on_requeue
+        self.sampler_workers = sampler_workers
+        self.sampler_chunk_size = sampler_chunk_size
+        self.store_location = store_location
         self.nnz = None
-        self.restart_num = 0
         self.test_datasets = []
+        self.force_recompute_indices = force_recompute_indices
         self.test_idx = []
         super().__init__()
+        print("finished init")
 
     def __repr__(self):
         return (
@@ -190,7 +177,7 @@ class DataModule(L.LightningDataModule):
             f"\ttest_split={self.test_split},\n"
             f"\tn_samples={self.n_samples},\n"
             f"\tweight_scaler={self.weight_scaler},\n"
-            f"\ttrain_oversampling_per_epoch={self.train_oversampling_per_epoch},\n"
+            f"\tn_samples_per_epoch={self.n_samples_per_epoch},\n"
             f"\tassays_to_drop={self.assays_to_drop},\n"
             f"\tnum_datasets={len(self.dataset.mapped_dataset.storages)},\n"
             f"\ttest datasets={str(self.test_datasets)},\n"
@@ -242,6 +229,44 @@ class DataModule(L.LightningDataModule):
         """
         return self.dataset.genedf.index.tolist()
 
+    @genes.setter
+    def genes(self, genes):
+        self.dataset.genedf = self.dataset.genedf.loc[genes]
+        self.kwargs["collate_fn"].genes = genes
+        self.kwargs["collate_fn"]._setup(
+            genedf=self.dataset.genedf,
+            org_to_id=self.kwargs["collate_fn"].org_to_id,
+            valid_genes=genes,
+        )
+
+    @property
+    def encoders(self):
+        return self.dataset.encoder
+
+    @encoders.setter
+    def encoders(self, encoders):
+        self.dataset.encoder = encoders
+        self.kwargs["collate_fn"].org_to_id = encoders[
+            self.kwargs["collate_fn"].organism_name
+        ]
+        self.kwargs["collate_fn"]._setup(
+            org_to_id=self.kwargs["collate_fn"].org_to_id,
+            valid_genes=self.genes,
+        )
+
+    @property
+    def organisms(self):
+        return self.dataset.organisms
+
+    @organisms.setter
+    def organisms(self, organisms):
+        self.dataset.organisms = organisms
+        self.kwargs["collate_fn"].organisms = organisms
+        self.kwargs["collate_fn"]._setup(
+            org_to_id=self.kwargs["collate_fn"].org_to_id,
+            valid_genes=self.genes,
+        )
+
     @property
     def num_datasets(self):
         return len(self.dataset.mapped_dataset.storages)
@@ -256,106 +281,191 @@ class DataModule(L.LightningDataModule):
             It can be either 'fit' or 'test'. Defaults to None.
         """
         SCALE = 10
-        if "nnz" in self.clss_to_weight and self.weight_scaler > 0:
-            self.nnz = self.dataset.mapped_dataset.get_merged_labels("nnz")
-            self.clss_to_weight.remove("nnz")
-            (
-                (self.nnz.max() / SCALE)
-                / ((1 + self.nnz - self.nnz.min()) + (self.nnz.max() / SCALE))
-            ).min()
-        if len(self.clss_to_weight) > 0 and self.weight_scaler > 0:
-            weights, labels = self.dataset.get_label_weights(
-                self.clss_to_weight,
-                scaler=self.weight_scaler,
-                return_categories=True,
+        print("setting up the datamodule")
+        start_time = time.time()
+        if (
+            self.store_location is None
+            or not os.path.exists(
+                os.path.join(self.store_location, "train_weights.npy")
             )
-        else:
-            weights = np.ones(1)
-            labels = np.zeros(self.n_samples, dtype=int)
-        if isinstance(self.validation_split, int):
-            len_valid = self.validation_split
-        else:
-            len_valid = int(self.n_samples * self.validation_split)
-        if isinstance(self.test_split, int):
-            len_test = self.test_split
-        else:
-            len_test = int(self.n_samples * self.test_split)
-        assert (
-            len_test + len_valid < self.n_samples
-        ), "test set + valid set size is configured to be larger than entire dataset."
-
-        idx_full = []
-        if len(self.assays_to_drop) > 0:
-            badloc = np.isin(
-                self.dataset.mapped_dataset.get_merged_labels("assay_ontology_term_id"),
-                self.assays_to_drop,
-            )
-            idx_full = np.arange(len(labels))[~badloc]
-        else:
-            idx_full = np.arange(self.n_samples)
-        if len_test > 0:
-            # this way we work on some never seen datasets
-            # keeping at least one
-            len_test = (
-                len_test
-                if len_test > self.dataset.mapped_dataset.n_obs_list[0]
-                else self.dataset.mapped_dataset.n_obs_list[0]
+            or self.force_recompute_indices
+        ):
+            if "nnz" in self.clss_to_weight and self.weight_scaler > 0:
+                self.nnz = self.dataset.mapped_dataset.get_merged_labels(
+                    "nnz", is_cat=False
+                )
+                self.clss_to_weight.remove("nnz")
+                (
+                    (self.nnz.max() / SCALE)
+                    / ((1 + self.nnz - self.nnz.min()) + (self.nnz.max() / SCALE))
+                ).min()
+            if len(self.clss_to_weight) > 0 and self.weight_scaler > 0:
+                weights, labels = self.dataset.get_label_weights(
+                    self.clss_to_weight,
+                    scaler=self.weight_scaler,
+                    return_categories=True,
+                )
+            else:
+                weights = np.ones(1)
+                labels = np.zeros(self.n_samples, dtype=int)
+            if isinstance(self.validation_split, int):
+                len_valid = self.validation_split
+            else:
+                len_valid = int(self.n_samples * self.validation_split)
+            if isinstance(self.test_split, int):
+                len_test = self.test_split
+            else:
+                len_test = int(self.n_samples * self.test_split)
+            assert len_test + len_valid < self.n_samples, (
+                "test set + valid set size is configured to be larger than entire dataset."
             )
-            cs = 0
-            for i, c in enumerate(self.dataset.mapped_dataset.n_obs_list):
-                if cs + c > len_test:
-                    break
-                else:
-                    self.test_datasets.append(
-                        self.dataset.mapped_dataset.path_list[i].path
-                    )
-                    cs += c
-            len_test = cs
-            self.test_idx = idx_full[:len_test]
-            idx_full = idx_full[len_test:]
-        else:
-            self.test_idx = None
 
-        np.random.shuffle(idx_full)
-        if len_valid > 0:
-            self.valid_idx = idx_full[:len_valid].copy()
-            # store it for later
-            idx_full = idx_full[len_valid:]
-        else:
-            self.valid_idx = None
-        weights = np.concatenate([weights, np.zeros(1)])
-        labels[~np.isin(np.arange(self.n_samples), idx_full)] = len(weights) - 1
-        # some labels will now not exist anymore as replaced by len(weights) - 1.
-        # this means that the associated weights should be 0.
-        # by doing np.bincount(labels)*weights this will be taken into account
-        self.train_weights = weights
-        self.train_labels = labels
-        self.idx_full = idx_full
+            idx_full = []
+            if len(self.assays_to_drop) > 0:
+                badloc = np.isin(
+                    self.dataset.mapped_dataset.get_merged_labels(
+                        "assay_ontology_term_id"
+                    ),
+                    self.assays_to_drop,
+                )
+                idx_full = np.arange(len(labels))[~badloc]
+            else:
+                idx_full = np.arange(self.n_samples)
+            if len_test > 0:
+                # this way we work on some never seen datasets
+                # keeping at least one
+                len_test = (
+                    len_test
+                    if len_test > self.dataset.mapped_dataset.n_obs_list[0]
+                    else self.dataset.mapped_dataset.n_obs_list[0]
+                )
+                cs = 0
+                d_size = list(enumerate(self.dataset.mapped_dataset.n_obs_list))
+                random.Random(42).shuffle(d_size)  # always same order
+                for i, c in d_size:
+                    if cs + c > len_test:
+                        break
+                    else:
+                        self.test_datasets.append(
+                            self.dataset.mapped_dataset.path_list[i].path
+                        )
+                        cs += c
+                len_test = cs
+                self.test_idx = idx_full[:len_test]
+                idx_full = idx_full[len_test:]
+            else:
+                self.test_idx = None
+
+            np.random.shuffle(idx_full)
+            if len_valid > 0:
+                self.valid_idx = idx_full[:len_valid].copy()
+                # store it for later
+                idx_full = idx_full[len_valid:]
+            else:
+                self.valid_idx = None
+            weights = np.concatenate([weights, np.zeros(1)])
+            labels[~np.isin(np.arange(self.n_samples), idx_full)] = len(weights) - 1
+            # some labels will now not exist anymore as replaced by len(weights) - 1.
+            # this means that the associated weights should be 0.
+            # by doing np.bincount(labels)*weights this will be taken into account
+            self.train_weights = weights
+            self.train_labels = labels
+            self.idx_full = idx_full
+        if self.store_location is not None:
+            if (
+                not os.path.exists(
+                    os.path.join(self.store_location, "train_weights.npy")
+                )
+                or self.force_recompute_indices
+            ):
+                os.makedirs(self.store_location, exist_ok=True)
+                if self.nnz is not None:
+                    np.save(os.path.join(self.store_location, "nnz.npy"), self.nnz)
+                np.save(
+                    os.path.join(self.store_location, "train_weights.npy"),
+                    self.train_weights,
+                )
+                np.save(
+                    os.path.join(self.store_location, "train_labels.npy"),
+                    self.train_labels,
+                )
+                np.save(
+                    os.path.join(self.store_location, "idx_full.npy"), self.idx_full
+                )
+                if self.test_idx is not None:
+                    np.save(
+                        os.path.join(self.store_location, "test_idx.npy"), self.test_idx
+                    )
+                if self.valid_idx is not None:
+                    np.save(
+                        os.path.join(self.store_location, "valid_idx.npy"),
+                        self.valid_idx,
+                    )
+                listToFile(
+                    self.test_datasets,
+                    os.path.join(self.store_location, "test_datasets.txt"),
+                )
+            else:
+                self.nnz = (
+                    np.load(os.path.join(self.store_location, "nnz.npy"), mmap_mode="r")
+                    if os.path.exists(os.path.join(self.store_location, "nnz.npy"))
+                    else None
+                )
+                self.train_weights = np.load(
+                    os.path.join(self.store_location, "train_weights.npy")
+                )
+                self.train_labels = np.load(
+                    os.path.join(self.store_location, "train_labels.npy")
+                )
+                self.idx_full = np.load(
+                    os.path.join(self.store_location, "idx_full.npy"), mmap_mode="r"
+                )
+                self.test_idx = (
+                    np.load(os.path.join(self.store_location, "test_idx.npy"))
+                    if os.path.exists(os.path.join(self.store_location, "test_idx.npy"))
+                    else None
+                )
+                self.valid_idx = (
+                    np.load(os.path.join(self.store_location, "valid_idx.npy"))
+                    if os.path.exists(
+                        os.path.join(self.store_location, "valid_idx.npy")
+                    )
+                    else None
+                )
+                self.test_datasets = fileToList(
+                    os.path.join(self.store_location, "test_datasets.txt")
+                )
+                print("loaded from store")
+        print(f"done setup, took {time.time() - start_time:.2f} seconds")
         return self.test_datasets
 
     def train_dataloader(self, **kwargs):
-        # train_sampler = WeightedRandomSampler(
-        #    self.train_weights[self.train_labels],
-        #    int(self.n_samples*self.train_oversampling_per_epoch),
-        #    replacement=True,
-        # )
-        try:
-            train_sampler = LabelWeightedSampler(
-                label_weights=self.train_weights,
-                labels=self.train_labels,
-                num_samples=int(self.n_samples * self.train_oversampling_per_epoch),
-                element_weights=self.nnz,
-                replacement=self.replacement,
-                restart_num=self.restart_num,
-                modify_seed_on_requeue=self.modify_seed_on_requeue,
-            )
-        except ValueError as e:
-            raise ValueError(e + "have you run `datamodule.setup()`?")
+        if len(self.clss_to_weight) > 0 and self.weight_scaler > 0:
+            try:
+                print("Setting up the parallel train sampler...")
+                # Create the optimized parallel sampler
+                print(f"Using {self.sampler_workers} workers for class indexing")
+                train_sampler = LabelWeightedSampler(
+                    label_weights=self.train_weights,
+                    labels=self.train_labels,
+                    num_samples=int(self.n_samples_per_epoch),
+                    element_weights=self.nnz,
+                    replacement=self.replacement,
+                    n_workers=self.sampler_workers,
+                    chunk_size=self.sampler_chunk_size,
+                    store_location=self.store_location,
+                    force_recompute_indices=self.force_recompute_indices,
+                )
+            except ValueError as e:
+                raise ValueError(str(e) + " Have you run `datamodule.setup()`?")
+        else:
+            train_sampler = SubsetRandomSampler(self.idx_full)
+        current_loader_kwargs = kwargs.copy()
+        current_loader_kwargs.update(self.kwargs)
         return DataLoader(
             self.dataset,
             sampler=train_sampler,
-            **self.kwargs,
-            **kwargs,
+            **current_loader_kwargs,
         )
 
     def val_dataloader(self):
@@ -385,115 +495,335 @@ class DataModule(L.LightningDataModule):
             **self.kwargs,
         )
 
-    # def teardown(self):
-    # clean up state after the trainer stops, delete files...
-    # called on every process in DDP
-    # pass
-
 
 class LabelWeightedSampler(Sampler[int]):
-    label_weights: Sequence[float]
-    klass_indices: Sequence[Sequence[int]]
+    """
+    A weighted random sampler that samples from a dataset with respect t o both class weights and element weights.
+
+    This sampler is designed to handle very large datasets efficiently, with optimizations for:
+    1. Parallel building of class indices
+    2. Chunked processing for large arrays
+    3. Efficient memory management
+    4. Proper handling of replacement and non-replacement sampling
+    """
+
+    label_weights: torch.Tensor
+    klass_indices: dict[int, torch.Tensor]
     num_samples: int
-    nnz: Optional[Sequence[int]]
+    element_weights: Optional[torch.Tensor]
     replacement: bool
-    restart_num: int
-    modify_seed_on_requeue: bool
-    # when we use, just set weights for each classes(here is: np.ones(num_classes)), and labels of a dataset.
-    # this will result a class-balanced sampling, no matter how imbalance the labels are.
 
     def __init__(
         self,
         label_weights: Sequence[float],
-        labels: Sequence[int],
+        labels: np.ndarray,
         num_samples: int,
         replacement: bool = True,
-        element_weights: Sequence[float] = None,
-        restart_num: int = 0,
-        modify_seed_on_requeue: bool = True,
+        element_weights: Optional[Sequence[float]] = None,
+        n_workers: int = None,
+        chunk_size: int = None,  # Process 10M elements per chunk
+        store_location: str = None,
+        force_recompute_indices: bool = False,
     ) -> None:
         """
+        Initialize the sampler with parallel processing for large datasets.
 
-        :param label_weights: list(len=num_classes)[float], weights for each class.
-        :param labels: list(len=dataset_len)[int], labels of a dataset.
-        :param num_samples: number of samples.
-        :param restart_num: if we are continuing a previous run, we need to restart the sampler from the same point.
+        Args:
+            label_weights: Weights for each class (length = number of classes)
+            labels: Class label for each dataset element (length = dataset size)
+            num_samples: Number of samples to draw
+            replacement: Whether to sample with replacement
+            element_weights: Optional weights for each element within classes
+            n_workers: Number of parallel workers to use (default: number of CPUs-1)
+            chunk_size: Size of chunks to process in parallel (default: 10M elements)
         """
-
+        print("Initializing optimized parallel weighted sampler...")
         super(LabelWeightedSampler, self).__init__(None)
-        # reweight labels from counter otherwsie same weight to labels that have many elements vs a few
-        label_weights = np.array(label_weights) * np.bincount(labels)
-
-        self.label_weights = torch.as_tensor(label_weights, dtype=torch.float32)
-        self.labels = torch.as_tensor(labels, dtype=torch.int)
-        self.element_weights = (
-            torch.as_tensor(element_weights, dtype=torch.float32)
-            if element_weights is not None
-            else None
-        )
+
+        # Compute label weights (incorporating class frequencies)
+        # Directly use labels as numpy array without conversion
+        label_weights = np.asarray(label_weights) * np.bincount(labels)
+        self.label_weights = torch.as_tensor(
+            label_weights, dtype=torch.float32
+        ).share_memory_()
+
+        # Store element weights if provided
+        if element_weights is not None:
+            self.element_weights = torch.as_tensor(
+                element_weights, dtype=torch.float32
+            ).share_memory_()
+        else:
+            self.element_weights = None
+
         self.replacement = replacement
         self.num_samples = num_samples
-        self.restart_num = slurm_restart_count(use_mine=True) + restart_num
-        self.modify_seed_on_requeue = modify_seed_on_requeue
-        # list of tensor.
-        self.klass_indices = [
-            (self.labels == i_klass).nonzero().squeeze(1)
-            for i_klass in range(len(label_weights))
-        ]
-        self.klass_sizes = [len(klass_indices) for klass_indices in self.klass_indices]
+        if (
+            store_location is None
+            or not os.path.exists(os.path.join(store_location, "klass_indices.pt"))
+            or force_recompute_indices
+        ):
+            # Set number of workers (default to CPU count - 1, but at least 1)
+            if n_workers is None:
+                # Check if running on SLURM
+                n_workers = min(20, max(1, mp.cpu_count() - 1))
+                if "SLURM_CPUS_PER_TASK" in os.environ:
+                    n_workers = min(
+                        20, max(1, int(os.environ["SLURM_CPUS_PER_TASK"]) - 1)
+                    )
+
+            # Try to auto-determine optimal chunk size based on memory
+            if chunk_size is None:
+                try:
+                    import psutil
+
+                    # Check if running on SLURM
+                    available_memory = psutil.virtual_memory().available
+                    for name in [
+                        "SLURM_MEM_PER_NODE",
+                        "SLURM_MEM_PER_CPU",
+                        "SLURM_MEM_PER_GPU",
+                        "SLURM_MEM_PER_TASK",
+                    ]:
+                        if name in os.environ:
+                            available_memory = (
+                                int(os.environ[name]) * 1024 * 1024
+                            )  # Convert MB to bytes
+                            break
+
+                    # Use at most 50% of available memory across all workers
+                    memory_per_worker = 0.5 * available_memory / n_workers
+                    # Rough estimate: each label takes 4 bytes, each index 8 bytes
+                    bytes_per_element = 12
+                    chunk_size = min(
+                        max(100_000, int(memory_per_worker / bytes_per_element / 3)),
+                        2_000_000,
+                    )
+                    print(f"Auto-determined chunk size: {chunk_size:,} elements")
+                except (ImportError, KeyError):
+                    chunk_size = 2_000_000
+                    print(f"Using default chunk size: {chunk_size:,} elements")
+
+            # Parallelize the class indices building
+            print(f"Building class indices in parallel with {n_workers} workers...")
+            klass_indices = self._build_class_indices_parallel(
+                labels, chunk_size, n_workers
+            )
+
+            # Convert klass_indices to a single tensor and offset vector
+            all_indices = []
+            offsets = []
+            current_offset = 0
+
+            # Sort keys to ensure consistent ordering
+            keys = klass_indices.keys()
+
+            # Build concatenated tensor and track offsets
+            for i in range(max(keys) + 1):
+                offsets.append(current_offset)
+                if i in keys:
+                    indices = klass_indices[i]
+                    all_indices.append(indices)
+                    current_offset += len(indices)
+
+            # Convert to tensors
+            self.klass_indices = torch.cat(all_indices).to(torch.int32).share_memory_()
+            self.klass_offsets = torch.tensor(offsets, dtype=torch.long).share_memory_()
+        if store_location is not None:
+            store_path = os.path.join(store_location, "klass_indices.pt")
+            if os.path.exists(store_path) and not force_recompute_indices:
+                self.klass_indices = torch.load(store_path).share_memory_()
+                self.klass_offsets = torch.load(
+                    store_path.replace(".pt", "_offsets.pt")
+                ).share_memory_()
+                print(f"Loaded sampler indices from {store_path}")
+            else:
+                torch.save(self.klass_indices, store_path)
+                torch.save(self.klass_offsets, store_path.replace(".pt", "_offsets.pt"))
+                print(f"Saved sampler indices to {store_path}")
+        print(f"Done initializing sampler with {len(self.klass_offsets)} classes")
 
     def __iter__(self):
+        # Sample classes according to their weights
+        print("sampling a new batch of size", self.num_samples)
+
         sample_labels = torch.multinomial(
             self.label_weights,
             num_samples=self.num_samples,
             replacement=True,
-            generator=None
-            if self.restart_num == 0 and not self.modify_seed_on_requeue
-            else torch.Generator().manual_seed(self.restart_num),
         )
-        sample_indices = torch.empty_like(sample_labels)
-        for i_klass, klass_index in enumerate(self.klass_indices):
+        # Get counts of each class in sample_labels
+        unique_samples, sample_counts = torch.unique(sample_labels, return_counts=True)
+
+        # Initialize result tensor
+        result_indices_list = []  # Changed name to avoid conflict if you had result_indices elsewhere
+
+        # Process only the classes that were actually sampled
+        for i, (label, count) in tqdm(
+            enumerate(zip(unique_samples.tolist(), sample_counts.tolist())),
+            total=len(unique_samples),
+            desc="Processing classes in sampler",
+        ):
+            klass_index = self.klass_indices[
+                self.klass_offsets[label] : self.klass_offsets[label + 1]
+            ]
+
             if klass_index.numel() == 0:
                 continue
-            left_inds = (sample_labels == i_klass).nonzero().squeeze(1)
-            if len(left_inds) == 0:
-                continue
+
+            # Sample elements from this class
             if self.element_weights is not None:
-                right_inds = torch.multinomial(
-                    self.element_weights[klass_index],
-                    num_samples=len(klass_index)
-                    if not self.replacement and len(klass_index) < len(left_inds)
-                    else len(left_inds),
-                    replacement=self.replacement,
-                    generator=None
-                    if self.restart_num == 0 and not self.modify_seed_on_requeue
-                    else torch.Generator().manual_seed(self.restart_num),
-                )
+                # This is a critical point for memory
+                current_element_weights_slice = self.element_weights[klass_index]
+
+                if self.replacement:
+                    right_inds = torch.multinomial(
+                        current_element_weights_slice,
+                        num_samples=count,
+                        replacement=True,
+                    )
+                else:
+                    num_to_sample = min(count, len(klass_index))
+                    right_inds = torch.multinomial(
+                        current_element_weights_slice,
+                        num_samples=num_to_sample,
+                        replacement=False,
+                    )
             elif self.replacement:
-                right_inds = torch.randint(
-                    len(klass_index),
-                    size=(len(left_inds),),
-                    generator=None
-                    if self.restart_num == 0 and not self.modify_seed_on_requeue
-                    else torch.Generator().manual_seed(self.restart_num),
-                )
+                right_inds = torch.randint(len(klass_index), size=(count,))
             else:
-                maxelem = (
-                    len(left_inds)
-                    if len(left_inds) < len(klass_index)
-                    else len(klass_index)
-                )
-                right_inds = torch.randperm(len(klass_index))[:maxelem]
-            sample_indices[left_inds[: len(right_inds)]] = klass_index[right_inds]
-            if len(right_inds) < len(left_inds):
-                sample_indices[left_inds[len(right_inds) :]] = -1
-        # drop all -1
-        sample_indices = sample_indices[sample_indices != -1]
-        # torch shuffle
-        sample_indices = sample_indices[torch.randperm(len(sample_indices))]
-        self.num_samples = len(sample_indices)
-        # raise Exception("stop")
-        yield from iter(sample_indices.tolist())
+                num_to_sample = min(count, len(klass_index))
+                right_inds = torch.randperm(len(klass_index))[:num_to_sample]
+
+            # Get actual indices
+            sampled_indices = klass_index[right_inds]
+            result_indices_list.append(sampled_indices)
+
+        # Combine all indices
+        if result_indices_list:  # Check if the list is not empty
+            final_result_indices = torch.cat(
+                result_indices_list
+            )  # Use the list with the appended new name
+
+            # Shuffle the combined indices
+            shuffled_indices = final_result_indices[
+                torch.randperm(len(final_result_indices))
+            ]
+            self.num_samples = len(shuffled_indices)
+            yield from shuffled_indices.tolist()
+        else:
+            self.num_samples = 0
+            yield from iter([])
 
     def __len__(self):
         return self.num_samples
+
+    def _merge_chunk_results(self, results_list):
+        """Merge results from multiple chunks into a single dictionary.
+
+        Args:
+            results_list: list of dictionaries mapping class labels to index arrays
+
+        Returns:
+            merged dictionary with PyTorch tensors
+        """
+        merged = {}
+
+        # Collect all labels across all chunks
+        all_labels = set()
+        for chunk_result in results_list:
+            all_labels.update(chunk_result.keys())
+
+        # For each unique label
+        for label in all_labels:
+            # Collect indices from all chunks where this label appears
+            indices_lists = [
+                chunk_result[label]
+                for chunk_result in results_list
+                if label in chunk_result
+            ]
+
+            if indices_lists:
+                # Concatenate all indices for this label
+                merged[label] = torch.tensor(
+                    np.concatenate(indices_lists), dtype=torch.long
+                )
+            else:
+                merged[label] = torch.tensor([], dtype=torch.long)
+
+        return merged
+
+    def _build_class_indices_parallel(self, labels, chunk_size, n_workers=None):
+        """Build class indices in parallel across multiple workers.
+
+        Args:
+            labels: array of class labels
+            n_workers: number of parallel workers
+            chunk_size: size of chunks to process
+
+        Returns:
+            dictionary mapping class labels to tensors of indices
+        """
+        n = len(labels)
+        results = []
+        # Create chunks of the labels array with proper sizing
+        n_chunks = (n + chunk_size - 1) // chunk_size  # Ceiling division
+        print(f"Processing {n:,} elements in {n_chunks} chunks...")
+
+        # Process in chunks to limit memory usage
+        with ProcessPoolExecutor(
+            max_workers=n_workers, mp_context=mp.get_context("spawn")
+        ) as executor:
+            # Submit chunks for processing
+            futures = []
+            for i in range(n_chunks):
+                start_idx = i * chunk_size
+                end_idx = min((i + 1) * chunk_size, n)
+                # We pass only chunk boundaries, not the data itself
+                # This avoids unnecessary copies during process creation
+                futures.append(
+                    executor.submit(
+                        self._process_chunk_with_slice,
+                        (start_idx, end_idx, labels),
+                    )
+                )
+
+            # Collect results as they complete with progress reporting
+            for future in tqdm(
+                as_completed(futures), total=len(futures), desc="Processing chunks"
+            ):
+                results.append(future.result())
+
+        # Merge results from all chunks
+        print("Merging results from all chunks...")
+        merged_results = self._merge_chunk_results(results)
+
+        return merged_results
+
+    def _process_chunk_with_slice(self, slice_info):
+        """Process a slice of the labels array by indices.
+
+        Args:
+            slice_info: tuple of (start_idx, end_idx, labels_array) where
+                       start_idx and end_idx define the slice to process
+
+        Returns:
+            dict mapping class labels to arrays of indices
+        """
+        start_idx, end_idx, labels_array = slice_info
+
+        # We're processing a slice of the original array
+        labels_slice = labels_array[start_idx:end_idx]
+        chunk_indices = {}
+
+        # Create a direct map of indices
+        indices = np.arange(start_idx, end_idx)
+
+        # Get unique labels in this slice for more efficient processing
+        unique_labels = np.unique(labels_slice)
+        # For each valid label, find its indices
+        for label in unique_labels:
+            # Find positions where this label appears (using direct boolean indexing)
+            label_mask = labels_slice == label
+            chunk_indices[int(label)] = indices[label_mask]
+
+        return chunk_indices