created bilstm dropout condition

Summary: Only drops out embedded_tokens if there is a non-zero probability in the dropout layer. This prevents recomputing a dropout mask even when it isn't required.

Differential Revision: D16190467

fbshipit-source-id: c6468c916ba381ad0eb3614ba0898cc5a03fd867

Author: shreydesai

pytext/models/representations/bilstm.py

@@ -57,7 +57,8 @@ def __init__(
         super().__init__(config)
 
         self.padding_value: float = padding_value
-        self.dropout = nn.Dropout(config.dropout)
+        if config.dropout > 0.0:
+            self.dropout = nn.Dropout(config.dropout)
         self.lstm = nn.LSTM(
             embed_dim,
             config.lstm_dim,
@@ -95,7 +96,8 @@ def forward(
                 Shape of each state is (bsize x num_layers * num_directions x nhid).
 
         """
-        embedded_tokens = self.dropout(embedded_tokens)
+        if self.dropout.p > 0.0:
+            embedded_tokens = self.dropout(embedded_tokens)
 
         if states is not None:
             # convert (h0, c0) from (bsz x num_layers*num_directions x nhid) to