Merge pull request cgpotts#23 from zijwang/fix_model_eval

fix model.eval() and explicitly add model.train()

torch_autoencoder.py

@@ -68,6 +68,7 @@ def fit(self, X):
         # Graph
         self.model = self.define_graph()
         self.model.to(self.device)
+        self.model.train()
         # Optimization:
         loss = nn.MSELoss()
         optimizer = self.optimizer(self.model.parameters(), lr=self.eta)
@@ -105,6 +106,7 @@ def predict(self, X):
         This will have the same shape as `X`.
 
         """
+        self.model.eval()
         with torch.no_grad():
             X_tensor = self.convert_input_to_tensor(X)
             X_pred = self.model(X_tensor)

torch_rnn_classifier.py

@@ -216,6 +216,7 @@ def fit(self, X, y, **kwargs):
         # Graph:
         self.model = self.build_graph()
         self.model.to(self.device)
+        self.model.train()
         # Make sure this value is up-to-date; self.`model` might change
         # it if it creates an embedding:
         self.embed_dim = self.model.embed_dim
@@ -257,6 +258,7 @@ def predict_proba(self, X):
         np.array with shape (len(X), self.n_classes_)
 
         """
+        self.model.eval()
         with torch.no_grad():
             X, seq_lengths = self._prepare_dataset(X)
             preds = self.model(X, seq_lengths)

torch_shallow_neural_classifier.py

@@ -83,6 +83,7 @@ def fit(self, X, y, **kwargs):
         # Graph:
         self.model = self.define_graph()
         self.model.to(self.device)
+        self.model.train()
         # Optimization:
         loss = nn.CrossEntropyLoss()
         optimizer = self.optimizer(
@@ -122,6 +123,7 @@ def predict_proba(self, X):
         np.array with shape (len(X), self.n_classes_)
 
         """
+        self.model.eval()
         with torch.no_grad():
             X = torch.tensor(X, dtype=torch.float).to(self.device)
             preds = self.model(X)

torch_subtree_nn.py

@@ -86,6 +86,7 @@ def predict_proba(self, X):
         """Returns a list of lists of prediction vectors, one list of
         vectors per tree in `X`.
         """
+        self.model.eval()
         with torch.no_grad():
             preds = []
             for tree in X:

torch_tree_nn.py

@@ -113,6 +113,7 @@ def fit(self, X, **kwargs):
         # Model:
         self.build_graph()
         self.model.to(self.device)
+        self.model.train()
         # Optimization:
         loss = nn.CrossEntropyLoss()
         optimizer = self.optimizer(self.model.parameters(), lr=self.eta)
@@ -179,6 +180,7 @@ def predict_proba(self, X):
         np.array with shape (len(X), self.n_classes_)
 
         """
+        self.model.eval()
         with torch.no_grad():
             preds = []
             for tree in X: