Add fixes for benchmark results (#14)

* update submodule

* add sharedmem to alra

* Add subsample to cellxgene datasets

* add shared mem to knn_smoothing

* update changelog

* use hard cutoff for batch selection

* update process_dataset to include random subsample

* script

* [WIP] remove shared mem

* Update changelog

* update changelog

CHANGELOG.md

@@ -12,6 +12,8 @@
 
 * Add `CHANGELOG.md` (PR #7).
 
+* Update `process_dataset` component to subsample large datasets (PR #14).
+
 ## MAJOR CHANGES
 
 * Revamp `scripts` directory (PR #13).

src/api/file_common_dataset.yaml

@@ -10,6 +10,12 @@ info:
         name: counts
         description: Raw counts
         required: true
+    obs:
+      - type: string
+        name: batch
+        description: Batch information
+        required: false
+
     uns:
       - type: string
         name: dataset_id

src/data_processors/process_dataset/config.vsh.yaml

@@ -19,6 +19,10 @@ arguments:
     type: "integer"
     description: "A seed for the subsampling."
     example: 123
+  - name: "--n_obs_limit"
+    type: "integer"
+    description: "The maximum number of cells the dataset may have before subsampling according to `obs.batch`."
+    default: 20000
 resources:
   - type: python_script
     path: script.py

src/data_processors/process_dataset/script.py

@@ -1,4 +1,5 @@
 import sys
+import random
 import anndata as ad
 import numpy as np
 
@@ -26,13 +27,30 @@
 print(">> Load Data", flush=True)
 adata = ad.read_h5ad(par["input"])
 
+# limit to max number of observations
+adata_output = adata.copy()
+if adata.n_obs > par["n_obs_limit"]:
+    print(">> Subsampling the observations", flush=True)
+    print(f">> Setting seed to {par['seed']}")
+    random.seed(par["seed"])
+    if "batch" not in adata.obs:
+        obs_filt = np.ones(dtype=np.bool_, shape=adata.n_obs)
+        obs_index = np.random.choice(np.where(obs_filt)[0], par["n_obs_limit"], replace=False)
+        adata_output = adata[obs_index].copy()
+    else:
+        batch_counts = adata.obs.groupby('batch').size()
+        filtered_batches = batch_counts[batch_counts <= par["n_obs_limit"]]
+        sorted_filtered_batches = filtered_batches.sort_values(ascending=False)
+        selected_batch = sorted_filtered_batches.index[0]
+        adata_output = adata[adata.obs["batch"]==selected_batch,:].copy()
+
 # remove all layers except for counts
-for key in list(adata.layers.keys()):
+for key in list(adata_output.layers.keys()):
     if key != "counts":
-        del adata.layers[key]
+        del adata_output.layers[key]
 
 # round counts and convert to int
-counts = np.array(adata.layers["counts"]).round().astype(int)
+counts = np.array(adata_output.layers["counts"]).round().astype(int)
 
 print(">> process and split data", flush=True)
 train_data, test_data = split_molecules(
@@ -49,16 +67,16 @@
 # copy adata to train_set, test_set
 output_train = ad.AnnData(
     layers={"counts": X_train},
-    obs=adata.obs[[]],
-    var=adata.var[[]],
-    uns={"dataset_id": adata.uns["dataset_id"]}
+    obs=adata_output.obs[[]],
+    var=adata_output.var[[]],
+    uns={"dataset_id": adata_output.uns["dataset_id"]}
 )
 test_uns_keys = ["dataset_id", "dataset_name", "dataset_url", "dataset_reference", "dataset_summary", "dataset_description", "dataset_organism"]
 output_test = ad.AnnData(
     layers={"counts": X_test},
-    obs=adata.obs[[]],
-    var=adata.var[[]],
-    uns={key: adata.uns[key] for key in test_uns_keys}
+    obs=adata_output.obs[[]],
+    var=adata_output.var[[]],
+    uns={key: adata_output.uns[key] for key in test_uns_keys}
 )
 
 # add additional information for the train set