Update trVAE 1.1.2

.github/workflows/test.yml

@@ -83,6 +83,7 @@ jobs:
             matrix:
                 python: [3.7, 3.9]
                 os: [ubuntu-latest]
+                method: [scanorama, scvi-tools, trvae]
         steps:
         -   uses: actions/checkout@v2
         -   name: Set up Python ${{ matrix.python }}
@@ -93,11 +94,11 @@ jobs:
         -   name: Install dependencies
             run: |
                 python -m pip install --upgrade pip
-                pip install '.[test,scanorama,scvi]'
+                pip install '.[test,${{ matrix.method }}]'
 
         -   name: Test with pytest
             run: |
-                python -m pytest --cov --cov-append --cov-report=term-missing -k integration -vv
+                python -m pytest --cov --cov-append --cov-report=term-missing -k ${{ matrix.method }} -vv
         -   name: Upload coverage
             env:
                 CODECOV_NAME: ${{ matrix.os }}-${{ matrix.python }}

scib/integration.py

@@ -55,39 +55,44 @@ def trvae(adata, batch, hvg=None):
         raise OptionalDependencyNotInstalled(e)
 
     utils.check_sanity(adata, batch, hvg)
-    n_batches = len(adata.obs[batch].cat.categories)
 
-    train_adata, valid_adata = trvae.utils.train_test_split(adata, train_frac=0.80)
+    batches = adata.obs[batch].unique().tolist()
 
-    condition_encoder = trvae.utils.create_dictionary(
-        adata.obs[batch].cat.categories.tolist(), []
-    )
-
-    network = trvae.archs.trVAEMulti(
-        x_dimension=train_adata.shape[1],
-        n_conditions=n_batches,
-        output_activation="relu",
+    network = trvae.models.trVAE(
+        x_dimension=adata.shape[1],
+        architecture=[256, 64],
+        z_dimension=10,
+        gene_names=adata.var_names.tolist(),
+        conditions=batches,
+        model_path="/localscratch/",
+        alpha=0.0001,
+        beta=50,
+        eta=100,
+        loss_fn="sse",
+        output_activation="linear",
     )
 
     network.train(
-        train_adata,
-        valid_adata,
-        condition_key=batch,
-        condition_encoder=condition_encoder,
-        verbose=0,
-    )
-
-    labels, _ = trvae.tl.label_encoder(
         adata,
-        condition_key=batch,
-        label_encoder=condition_encoder,
+        batch,
+        train_size=0.8,
+        n_epochs=50,
+        batch_size=512,
+        early_stop_limit=10,
+        lr_reducer=20,
+        verbose=5,
+        save=False,
     )
 
-    network.get_corrected(adata, labels, return_z=False)
+    latent_adata = network.get_latent(adata, batch)
+
+    target_batch = adata.obs[batch].value_counts().index[0]
+
+    corrected_data = network.predict(adata, batch, target_condition=target_batch)
 
-    adata.obsm["X_emb"] = adata.obsm["mmd_latent"]
-    del adata.obsm["mmd_latent"]
-    adata.X = adata.obsm["reconstructed"]
+    # Assign trVAE outputs
+    adata.obsm["X_emb"] = latent_adata
+    adata.X = corrected_data
 
     return adata
 

setup.cfg

@@ -79,8 +79,8 @@ bbknn = bbknn ==1.3.9
 scanorama = scanorama ==1.7.0
 mnn = mnnpy ==0.1.9.5
 scgen = scgen  >=2.1.0
-scvi = scvi-tools  >=0.16.1
-trvae = trvae  ==1.1.2
+scvi-tools = scvi-tools  >=0.16.1
+trvae = tensorflow ==2.5.3; trvae @ git+https://github.com/theislab/trVAE.git@6fd87fcd1fc47a6b93579dcb7caac0d1e85ed10e
 trvaep = trvaep  ==0.1.0
 desc = desc  ==2.0.3
 

tests/integration/test_scanorama.py

@@ -7,9 +7,17 @@
 def test_scanorama(adata_paul15_template):
     adata = scib.ig.scanorama(adata_paul15_template, batch="batch")
 
+    # check full feature output
     scib.pp.reduce_data(
-        adata, n_top_genes=200, neighbors=True, use_rep="X_emb", pca=True, umap=False
+        adata, pca=True, n_top_genes=200, neighbors=True, use_rep="X_pca", umap=False
     )
+    score = scib.me.graph_connectivity(adata, label_key="celltype")
+    LOGGER.info(f"\nscore: {score}")
+
+    # check embedding output
+    scib.pp.reduce_data(adata, pca=False, neighbors=True, use_rep="X_emb", umap=False)
+    score = scib.me.graph_connectivity(adata, label_key="celltype")
+    LOGGER.info(f"\nscore: {score}")
 
     # check NMI after clustering
     res_max, score_max, _ = scib.cl.opt_louvain(

tests/integration/test_scvi-tools.py

@@ -0,0 +1,22 @@
+import scib
+from tests.common import assert_near_exact
+
+
+def test_scvi(adata_paul15_template):
+    adata = scib.ig.scvi(adata_paul15_template, batch="batch", max_epochs=20)
+
+    scib.pp.reduce_data(adata, pca=False, neighbors=True, use_rep="X_emb", umap=False)
+
+    score = scib.me.graph_connectivity(adata, label_key="celltype")
+    assert_near_exact(score, 0.9684638088694193, 1e-2)
+
+
+def test_scanvi(adata_paul15_template):
+    adata = scib.ig.scanvi(
+        adata_paul15_template, batch="batch", labels="celltype", max_epochs=20
+    )
+
+    scib.pp.reduce_data(adata, pca=False, neighbors=True, use_rep="X_emb", umap=False)
+
+    score = scib.me.graph_connectivity(adata, label_key="celltype")
+    assert_near_exact(score, 0.9834078129657216, 1e-2)

tests/integration/test_trvae.py

@@ -0,0 +1,19 @@
+import scib
+from tests.common import LOGGER
+
+
+def test_trvae(adata_paul15_template):
+    adata = scib.ig.trvae(adata_paul15_template, batch="batch")
+
+    # check full feature output
+    scib.pp.reduce_data(
+        adata, pca=True, n_top_genes=200, neighbors=True, use_rep="X_pca", umap=False
+    )
+    score = scib.me.graph_connectivity(adata, label_key="celltype")
+    LOGGER.info(f"\nscore: {score}")
+
+    # check embedding output
+    scib.pp.reduce_data(adata, pca=False, neighbors=True, use_rep="X_emb", umap=False)
+    score = scib.me.graph_connectivity(adata, label_key="celltype")
+    LOGGER.info(f"\nscore: {score}")
+    # assert_near_exact(score, 0.9834078129657216, 1e-2)