Merge pull request #205 from jrzaurin/fix_mlp_hidden_dim_attribute_bug

fixed an issue related to the mlp_hidden_dims param/attribute #204

pytorch_widedeep/models/tabular/resnet/tab_resnet.py

@@ -400,7 +400,7 @@ def __init__(
 
         if self.mlp_hidden_dims is not None:
             self.mlp = MLP(
-                d_hidden=[self.mlp_first_hidden_dim] + self.mlp_hidden_dim,
+                d_hidden=[self.mlp_first_hidden_dim] + self.mlp_hidden_dims,
                 activation=(
                     "relu" if self.mlp_activation is None else self.mlp_activation
                 ),

pytorch_widedeep/models/tabular/transformers/ft_transformer.py

@@ -307,7 +307,7 @@ def __init__(
         # therefore all related params are optional
         if self.mlp_hidden_dims is not None:
             self.mlp = MLP(
-                d_hidden=[self.mlp_first_hidden_dim] + self.mlp_hidden_dim,
+                d_hidden=[self.mlp_first_hidden_dim] + self.mlp_hidden_dims,
                 activation=(
                     "relu" if self.mlp_activation is None else self.mlp_activation
                 ),

pytorch_widedeep/models/tabular/transformers/saint.py

@@ -288,7 +288,7 @@ def __init__(
         # therefore all related params are optional
         if self.mlp_hidden_dims is not None:
             self.mlp = MLP(
-                d_hidden=[self.mlp_first_hidden_dim] + self.mlp_hidden_dim,
+                d_hidden=[self.mlp_first_hidden_dim] + self.mlp_hidden_dims,
                 activation=(
                     "relu" if self.mlp_activation is None else self.mlp_activation
                 ),

pytorch_widedeep/models/tabular/transformers/tab_fastformer.py

@@ -324,7 +324,7 @@ def __init__(
         # therefore all related params are optional
         if self.mlp_hidden_dims is not None:
             self.mlp = MLP(
-                d_hidden=[self.mlp_first_hidden_dim] + self.mlp_hidden_dim,
+                d_hidden=[self.mlp_first_hidden_dim] + self.mlp_hidden_dims,
                 activation=(
                     "relu" if self.mlp_activation is None else self.mlp_activation
                 ),

pytorch_widedeep/models/tabular/transformers/tab_perceiver.py

@@ -321,7 +321,7 @@ def __init__(
         # Mlp
         if self.mlp_hidden_dims is not None:
             self.mlp = MLP(
-                d_hidden=[self.mlp_first_hidden_dim] + self.mlp_hidden_dim,
+                d_hidden=[self.mlp_first_hidden_dim] + self.mlp_hidden_dims,
                 activation=(
                     "relu" if self.mlp_activation is None else self.mlp_activation
                 ),

pytorch_widedeep/models/tabular/transformers/tab_transformer.py

@@ -311,7 +311,7 @@ def __init__(
 
         if self.mlp_hidden_dims is not None:
             self.mlp = MLP(
-                d_hidden=[self.mlp_first_hidden_dim] + self.mlp_hidden_dim,
+                d_hidden=[self.mlp_first_hidden_dim] + self.mlp_hidden_dims,
                 activation=(
                     "relu" if self.mlp_activation is None else self.mlp_activation
                 ),

requirements.txt

@@ -1,6 +1,6 @@
 pandas>=1.3.5
 numpy>=1.21.6
-scipy>=1.7.3
+scipy>=1.7.3,<=1.12.0
 scikit-learn>=1.0.2
 gensim
 spacy

tests/test_model_components/test_mc_transformers.py

@@ -208,6 +208,23 @@ def _build_model(model_name, params):
         return TabPerceiver(n_perceiver_blocks=2, n_latents=2, latent_dim=16, **params)
 
 
+# test out of a bug related to the mlp_hidden_dims attribute (https://github.com/jrzaurin/pytorch-widedeep/issues/204)
+@pytest.mark.parametrize(
+    "model_name", ["tabtransformer", "saint", "fttransformer", "tabfastformer"]
+)
+def test_mlphidden_dims(model_name):
+    params = {
+        "column_idx": {k: v for v, k in enumerate(colnames)},
+        "cat_embed_input": embed_input,
+        "continuous_cols": colnames[n_cols:],
+        "mlp_hidden_dims": [32, 16],
+    }
+
+    model = _build_model(model_name, params)
+    out = model(X_tab)
+    assert out.size(0) == 10 and out.size(1) == model.output_dim == 16
+
+
 @pytest.mark.parametrize(
     "embed_continuous, with_cls_token, model_name",
     [