scdataloader 0.0.2__py3-none-any.whl → 0.0.4__py3-none-any.whl

scdataloader/datamodule.py

@@ -0,0 +1,375 @@
+import numpy as np
+import pandas as pd
+import lamindb as ln
+
+from torch.utils.data.sampler import (
+    WeightedRandomSampler,
+    SubsetRandomSampler,
+    SequentialSampler,
+)
+import torch
+from torch.utils.data import DataLoader, Sampler
+import lightning as L
+
+from typing import Optional, Union, Sequence
+
+from .data import Dataset
+from .collator import Collator
+from .utils import getBiomartTable
+
+
+class DataModule(L.LightningDataModule):
+    def __init__(
+        self,
+        collection_name: str,
+        label_to_weight: list = ["organism_ontology_term_id"],
+        organisms: list = ["NCBITaxon:9606"],
+        weight_scaler: int = 10,
+        train_oversampling_per_epoch: float = 0.1,
+        validation_split: float = 0.2,
+        test_split: float = 0,
+        gene_embeddings: str = "",
+        use_default_col: bool = True,
+        gene_position_tolerance: int = 10_000,
+        # this is for the mappedCollection
+        label_to_pred: list = ["organism_ontology_term_id"],
+        all_labels: list = ["organism_ontology_term_id"],
+        hierarchical_labels: list = [],
+        # this is for the collator
+        how: str = "random expr",
+        organism_name: str = "organism_ontology_term_id",
+        max_len: int = 1000,
+        add_zero_genes: int = 100,
+        do_gene_pos: Union[bool, str] = True,
+        tp_name: Optional[str] = None,  # "heat_diff"
+        assays_to_drop: list = [
+            "EFO:0008853",
+            "EFO:0010961",
+            "EFO:0030007",
+            "EFO:0030062",
+        ],
+        **kwargs,
+    ):
+        """
+        DataModule a pytorch lighting datamodule directly from a lamin Collection.
+        it can work with bare pytorch too
+
+        It implements train / val / test dataloaders. the train is weighted random, val is random, test is one to many separated datasets.
+        This is where the mappedCollection, dataset, and collator are combined to create the dataloaders.
+
+        Args:
+            collection_name (str): The lamindb collection to be used.
+            weight_scaler (int, optional): how much more you will see the most present vs less present category.
+            gene_position_tolerance (int, optional): The tolerance for gene position. Defaults to 10_000.
+                any genes within this distance of each other will be considered at the same position.
+            gene_embeddings (str, optional): The path to the gene embeddings file. Defaults to "".
+                the file must have ensembl_gene_id as index.
+                This is used to subset the available genes further to the ones that have embeddings in your model.
+            organisms (list, optional): The organisms to include in the dataset. Defaults to ["NCBITaxon:9606"].
+            label_to_weight (list, optional): List of labels to weight in the trainer's weighted random sampler. Defaults to [].
+            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.
+            **other args: see @file data.py and @file collator.py for more details
+            **kwargs: Additional keyword arguments passed to the pytorch DataLoader.
+        """
+        if collection_name is not None:
+            mdataset = Dataset(
+                ln.Collection.filter(name=collection_name).first(),
+                organisms=organisms,
+                obs=all_labels,
+                clss_to_pred=label_to_pred,
+                hierarchical_clss=hierarchical_labels,
+            )
+            print(mdataset)
+        # and location
+        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
+                biomart = getBiomartTable(
+                    attributes=["start_position", "chromosome_name"]
+                ).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
+            mdataset.genedf = biomart.loc[mdataset.genedf.index]
+            self.gene_pos = mdataset.genedf["pos"].tolist()
+
+        if gene_embeddings != "":
+            mdataset.genedf = mdataset.genedf.join(
+                pd.read_parquet(gene_embeddings), how="inner"
+            )
+            if do_gene_pos:
+                self.gene_pos = mdataset.genedf["pos"].tolist()
+        self.labels = {k: len(v) for k, v in mdataset.class_topred.items()}
+        # we might want not to order the genes by expression (or do it?)
+        # we might want to not introduce zeros and
+        if use_default_col:
+            kwargs["collate_fn"] = Collator(
+                organisms=organisms,
+                how=how,
+                valid_genes=mdataset.genedf.index.tolist(),
+                max_len=max_len,
+                add_zero_genes=add_zero_genes,
+                org_to_id=mdataset.encoder[organism_name],
+                tp_name=tp_name,
+                organism_name=organism_name,
+                class_names=label_to_weight,
+            )
+        self.validation_split = validation_split
+        self.test_split = test_split
+        self.dataset = mdataset
+        self.kwargs = kwargs
+        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.label_to_weight = label_to_weight
+        self.train_weights = None
+        self.train_labels = None
+        super().__init__()
+
+    def __repr__(self):
+        return (
+            f"DataLoader(\n"
+            f"\twith a dataset=({self.dataset.__repr__()}\n)\n"
+            f"\tvalidation_split={self.validation_split},\n"
+            f"\ttest_split={self.test_split},\n"
+            f"\tn_samples={self.n_samples},\n"
+            f"\tweight_scaler={self.weight_scaler},\n"
+            f"\train_oversampling_per_epoch={self.train_oversampling_per_epoch},\n"
+            f"\tlabel_to_weight={self.label_to_weight}\n"
+            + (
+                "\twith train_dataset size of=("
+                + str((self.train_weights != 0).sum())
+                + ")\n)"
+            )
+            if self.train_weights is not None
+            else ")"
+        )
+
+    @property
+    def decoders(self):
+        """
+        decoders the decoders for any labels that would have been encoded
+
+        Returns:
+            dict[str, dict[int, str]]
+        """
+        decoders = {}
+        for k, v in self.dataset.encoder.items():
+            decoders[k] = {va: ke for ke, va in v.items()}
+        return decoders
+
+    @property
+    def cls_hierarchy(self):
+        """
+        cls_hierarchy the hierarchy of labels for any cls that would have a hierarchy
+
+        Returns:
+            dict[str, dict[str, str]]
+        """
+        cls_hierarchy = {}
+        for k, dic in self.dataset.class_groupings.items():
+            rdic = {}
+            for sk, v in dic.items():
+                rdic[self.dataset.encoder[k][sk]] = [
+                    self.dataset.encoder[k][i] for i in list(v)
+                ]
+            cls_hierarchy[k] = rdic
+        return cls_hierarchy
+
+    @property
+    def genes(self):
+        """
+        genes the genes used in this datamodule
+
+        Returns:
+            list
+        """
+        return self.dataset.genedf.index.tolist()
+
+    @property
+    def num_datasets(self):
+        return len(self.dataset.mapped_dataset.storages)
+
+    def setup(self, stage=None):
+        """
+        setup method is used to prepare the data for the training, validation, and test sets.
+        It shuffles the data, calculates weights for each set, and creates samplers for each set.
+
+        Args:
+            stage (str, optional): The stage of the model training process.
+            It can be either 'fit' or 'test'. Defaults to None.
+        """
+        if len(self.label_to_weight) > 0:
+            weights, labels = self.dataset.get_label_weights(
+                self.label_to_weight, scaler=self.weight_scaler
+            )
+        else:
+            weights = np.ones(1)
+            labels = np.zeros(self.n_samples)
+        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:
+            for i, a in enumerate(
+                self.dataset.mapped_dataset.get_merged_labels("assay_ontology_term_id")
+            ):
+                if a not in self.assays_to_drop:
+                    idx_full.append(i)
+            idx_full = np.array(idx_full)
+        else:
+            idx_full = np.arange(self.n_samples)
+        test_datasets = []
+        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
+            print("these files will be considered test datasets:")
+            for i, c in enumerate(self.dataset.mapped_dataset.n_obs_list):
+                if cs + c > len_test:
+                    break
+                else:
+                    print("    " + self.dataset.mapped_dataset.path_list[i].path)
+                    test_datasets.append(self.dataset.mapped_dataset.path_list[i].path)
+                    cs += c
+
+            len_test = cs
+            print("perc test: ", len_test / self.n_samples)
+            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()
+            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
+
+        self.train_weights = weights
+        self.train_labels = labels
+        self.idx_full = idx_full
+
+        return 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,
+        # )
+        train_sampler = LabelWeightedSampler(
+            self.train_weights,
+            self.train_labels,
+            num_samples=int(self.n_samples * self.train_oversampling_per_epoch),
+            # replacement=True,
+        )
+        return DataLoader(self.dataset, sampler=train_sampler, **self.kwargs, **kwargs)
+
+    def val_dataloader(self):
+        return (
+            DataLoader(
+                self.dataset, sampler=SubsetRandomSampler(self.valid_idx), **self.kwargs
+            )
+            if self.valid_idx is not None
+            else None
+        )
+
+    def test_dataloader(self):
+        return (
+            DataLoader(
+                self.dataset, sampler=SequentialSampler(self.test_idx), **self.kwargs
+            )
+            if self.test_idx is not None
+            else None
+        )
+
+    # 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]]
+    num_samples: int
+
+    # 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.
+    # NOTE: here we use replacement=True, you can change it if you don't upsample a class.
+    def __init__(
+        self, label_weights: Sequence[float], labels: Sequence[int], num_samples: int
+    ) -> None:
+        """
+
+        :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.
+        """
+
+        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.num_samples = num_samples
+        # list of tensor.
+        self.klass_indices = [
+            (self.labels == i_klass).nonzero().squeeze(1)
+            for i_klass in range(len(label_weights))
+        ]
+
+    def __iter__(self):
+        sample_labels = torch.multinomial(
+            self.label_weights, num_samples=self.num_samples, replacement=True
+        )
+        sample_indices = torch.empty_like(sample_labels)
+        for i_klass, klass_index in enumerate(self.klass_indices):
+            if klass_index.numel() == 0:
+                continue
+            left_inds = (sample_labels == i_klass).nonzero().squeeze(1)
+            right_inds = torch.randint(len(klass_index), size=(len(left_inds),))
+            sample_indices[left_inds] = klass_index[right_inds]
+        yield from iter(sample_indices.tolist())
+
+    def __len__(self):
+        return self.num_samples

scdataloader/mapped.py

@@ -80,38 +80,51 @@ class MappedDataset:
         join_vars: Optional[Literal["auto", "inner", "None"]] = "auto",
         encode_labels: Optional[Union[bool, List[str]]] = False,
         parallel: bool = False,
+        unknown_class: str = "unknown",
     ):
         self.storages = []
         self.conns = []
         self.parallel = parallel
+        self.unknown_class = unknown_class
+        self.path_list = path_list
         self._make_connections(path_list, parallel)
 
         self.n_obs_list = []
         for storage in self.storages:
             with _Connect(storage) as store:
                 X = store["X"]
+                index = (
+                    store["var"]["ensembl_gene_id"]
+                    if "ensembl_gene_id" in store["var"]
+                    else store["var"]["_index"]
+                )
+                if join_vars == "None":
+                    if not all(
+                        [
+                            i <= j
+                            for i, j in zip(
+                                index[:99],
+                                index[1:100],
+                            )
+                        ]
+                    ):
+                        raise ValueError("The variables are not sorted.")
                 if isinstance(X, ArrayTypes):  # type: ignore
                     self.n_obs_list.append(X.shape[0])
                 else:
                     self.n_obs_list.append(X.attrs["shape"][0])
         self.n_obs = sum(self.n_obs_list)
 
-        self.indices = np.hstack(
-            [np.arange(n_obs) for n_obs in self.n_obs_list]
-        )
-        self.storage_idx = np.repeat(
-            np.arange(len(self.storages)), self.n_obs_list
-        )
+        self.indices = np.hstack([np.arange(n_obs) for n_obs in self.n_obs_list])
+        self.storage_idx = np.repeat(np.arange(len(self.storages)), self.n_obs_list)
 
         self.join_vars = join_vars if len(path_list) > 1 else None
         self.var_indices = None
-        if self.join_vars is not None:
+        if self.join_vars != "None":
             self._make_join_vars()
 
         self.encode_labels = encode_labels
-        self.label_keys = (
-            [label_keys] if isinstance(label_keys, str) else label_keys
-        )
+        self.label_keys = [label_keys] if isinstance(label_keys, str) else label_keys
         if isinstance(encode_labels, bool):
             if encode_labels:
                 encode_labels = label_keys
@@ -122,6 +135,8 @@ class MappedDataset:
             for label in encode_labels:
                 cats = self.get_merged_categories(label)
                 self.encoders[label] = {cat: i for i, cat in enumerate(cats)}
+                if unknown_class in self.encoders[label]:
+                    self.encoders[label][unknown_class] = -1
         else:
             self.encoders = {}
         self._closed = False
@@ -157,9 +172,15 @@ class MappedDataset:
                 raise ValueError(
                     "The provided AnnData objects don't have shared varibales."
                 )
-            self.var_indices = [
-                vrs.get_indexer(self.var_joint) for vrs in var_list
-            ]
+            self.var_indices = [vrs.get_indexer(self.var_joint) for vrs in var_list]
+
+    def _check_aligned_vars(self, vars: list):
+        i = 0
+        for storage in self.storages:
+            with _Connect(storage) as store:
+                if vars == _safer_read_index(store["var"]).tolist():
+                    i += 1
+        print("{}% are aligned".format(i * 100 / len(self.storages)))
 
     def __len__(self):
         return self.n_obs
@@ -172,20 +193,17 @@ class MappedDataset:
         else:
             var_idxs = None
         with _Connect(self.storages[storage_idx]) as store:
-            out = [self.get_data_idx(store, obs_idx, var_idxs)]
+            out = {"x": self.get_data_idx(store, obs_idx, var_idxs)}
             if self.label_keys is not None:
-                for i, label in enumerate(self.label_keys):
+                for _, label in enumerate(self.label_keys):
                     label_idx = self.get_label_idx(store, obs_idx, label)
                     if label in self.encoders:
-                        out.append(self.encoders[label][label_idx])
+                        out.update({label: self.encoders[label][label_idx]})
                     else:
-                        out.append(label_idx)
+                        out.update({label: label_idx})
+                out.update({"dataset": storage_idx})
         return out
 
-    def uns(self, idx, key):
-        storage = self.storages[self.storage_idx[idx]]
-        return storage["uns"][key]
-
     def get_data_idx(
         self,
         storage: StorageType,
@@ -240,13 +258,13 @@ class MappedDataset:
             if i == 0:
                 labels = self.get_merged_labels(val)
             else:
-                labels += "_" + self.get_merged_labels(val).astype(str).astype(
-                    "O"
-                )
+                labels += "_" + self.get_merged_labels(val).astype(str).astype("O")
         counter = Counter(labels)  # type: ignore
-        counter = np.array([counter[label] for label in labels])
+        rn = {n: i for i, n in enumerate(counter.keys())}
+        labels = np.array([rn[label] for label in labels])
+        counter = np.array(list(counter.values()))
         weights = scaler / (counter + scaler)
-        return weights
+        return weights, labels
 
     def get_merged_labels(self, label_key: str):
         """Get merged labels."""
@@ -255,9 +273,7 @@ class MappedDataset:
         for storage in self.storages:
             with _Connect(storage) as store:
                 codes = self.get_codes(store, label_key)
-                labels = (
-                    decode(codes) if isinstance(codes[0], bytes) else codes
-                )
+                labels = decode(codes) if isinstance(codes[0], bytes) else codes
                 cats = self.get_categories(store, label_key)
                 if cats is not None:
                     cats = decode(cats) if isinstance(cats[0], bytes) else cats
@@ -277,9 +293,7 @@ class MappedDataset:
                     cats_merge.update(cats)
                 else:
                     codes = self.get_codes(store, label_key)
-                    codes = (
-                        decode(codes) if isinstance(codes[0], bytes) else codes
-                    )
+                    codes = decode(codes) if isinstance(codes[0], bytes) else codes
                     cats_merge.update(codes)
         return cats_merge