add General attention classes

The modifications that I introduced in a previous commit did break
Bert's internal API. I reverted these changes and added more general
classes to handle the encoder-decoder attention case.

There may be a more elegant way to deal with retro-compatibility (I am
not comfortable with the current state of the code), but I cannot see it
right now.

transformers/modeling_bert.py

@@ -174,9 +174,9 @@ class BertEmbeddings(nn.Module):
         return embeddings
 
 
-class BertSelfAttention(nn.Module):
+class BertGeneralAttention(nn.Module):
     def __init__(self, config):
-        super(BertSelfAttention, self).__init__()
+        super(BertGeneralAttention, self).__init__()
         if config.hidden_size % config.num_attention_heads != 0:
             raise ValueError(
                 "The hidden size (%d) is not a multiple of the number of attention "
@@ -235,6 +235,67 @@ class BertSelfAttention(nn.Module):
         return outputs
 
 
+class BertSelfAttention(nn.Module):
+    def __init__(self, config):
+        super(BertSelfAttention, self).__init__()
+        if config.hidden_size % config.num_attention_heads != 0:
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, config.num_attention_heads))
+        self.output_attentions = config.output_attentions
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+        mixed_query_layer = self.query(hidden_states)
+        mixed_key_layer = self.key(hidden_states)
+        mixed_value_layer = self.value(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+        key_layer = self.transpose_for_scores(mixed_key_layer)
+        value_layer = self.transpose_for_scores(mixed_value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if self.output_attentions else (context_layer,)
+        return outputs
+
+
 class BertSelfOutput(nn.Module):
     def __init__(self, config):
         super(BertSelfOutput, self).__init__()
@@ -279,12 +340,49 @@ class BertAttention(nn.Module):
         self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
         self.pruned_heads = self.pruned_heads.union(heads)
 
-    def forward(self, query_tensor, key_tensor, value_tensor, attention_mask=None, head_mask=None):
-        self_outputs = self.self(query_tensor, key_tensor, value_tensor, attention_mask, head_mask)
+    def forward(self, hidden_states, attention_mask=None, head_mask=None):
+        self_outputs = self.self(hidden_states, attention_mask, head_mask)
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+class BertDecoderAttention(nn.Module):
+    def __init__(self, config):
+        super(BertAttention, self).__init__()
+        self.self = BertGeneralAttention(config)
+        self.output = BertSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        mask = torch.ones(self.self.num_attention_heads, self.self.attention_head_size)
+        heads = set(heads) - self.pruned_heads  # Convert to set and emove already pruned heads
+        for head in heads:
+            # Compute how many pruned heads are before the head and move the index accordingly
+            head = head - sum(1 if h < head else 0 for h in self.pruned_heads)
+            mask[head] = 0
+        mask = mask.view(-1).contiguous().eq(1)
+        index = torch.arange(len(mask))[mask].long()
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(self, query, key, value, attention_mask=None, head_mask=None):
+        self_outputs = self.self(query, key, value, attention_mask, head_mask)
         # in encoder-decoder attention we use the output of the previous decoder stage as the query
         # in the Multi-Head Attention. We thus pass query_tensor as the residual in BertOutput.
         # This shows the limits of the current code architecture, which may benefit from some refactoring.
-        attention_output = self.output(self_outputs[0], query_tensor)
+        attention_output = self.output(self_outputs[0], query)
         outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
         return outputs
 
@@ -326,11 +424,7 @@ class BertEncoderLayer(nn.Module):
         self.output = BertOutput(config)
 
     def forward(self, hidden_states, attention_mask=None, head_mask=None):
-        attention_outputs = self.attention(query_tensor=hidden_states,
-                                           key_tensor=hidden_states,
-                                           value_tensor=hidden_states,
-                                           attention_mask=attention_mask,
-                                           head_mask=head_mask)
+        attention_outputs = self.attention(hidden_states, attention_mask, head_mask)
         attention_output = attention_outputs[0]
         intermediate_output = self.intermediate(attention_output)
         layer_output = self.output(intermediate_output, attention_output)
@@ -342,20 +436,16 @@ class BertDecoderLayer(nn.Module):
     def __init__(self, config):
         super(BertDecoderLayer, self).__init__()
         self.self_attention = BertAttention(config)
-        self.attention = BertAttention(config)
+        self.attention = BertDecoderAttention(config)
         self.intermediate = BertIntermediate(config)
         self.output = BertOutput(config)
 
     def forward(self, hidden_states, encoder_outputs, attention_mask=None, head_mask=None):
-        self_attention_outputs = self.self_attention(query_tensor=hidden_states,
-                                                     key_tensor=hidden_states,
-                                                     value_tensor=hidden_states,
-                                                     attention_mask=attention_mask,
-                                                     head_mask=head_mask)
+        self_attention_outputs = self.self_attention(hidden_states, attention_mask, head_mask)
         self_attention_output = self_attention_outputs[0]
-        attention_outputs = self.attention(query_tensor=self_attention_output,
-                                           key_tensor=encoder_outputs,
-                                           value_tensor=encoder_outputs,
+        attention_outputs = self.attention(query=self_attention_output,
+                                           key=encoder_outputs,
+                                           value=encoder_outputs,
                                            attention_mask=attention_mask,
                                            head_mask=head_mask)
         attention_output = attention_outputs[0]
@@ -399,10 +489,34 @@ class BertEncoder(nn.Module):
 
 class BertDecoder(nn.Module):
     def __init__(self, config):
-        raise NotImplementedError
+        super(BertDecoder, self).__init__()
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.layers = nn.ModuleList([BertEncoderLayer(config) for _ in range(config.num_hidden_layers)])
 
-    def forward(self, encoder_output):
-        raise NotImplementedError
+    def forward(self, hidden_states, encoder_outputs, attention_mask=None, head_mask=None):
+        all_hidden_states = ()
+        all_attentions = ()
+        for i, layer_module in enumerate(self.layer):
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_outputs = layer_module(hidden_states, attention_mask, head_mask[i])
+            if self.output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+            hidden_states = layer_outputs[0]
+
+        # Add last layer
+        if self.output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (all_attentions,)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions)
 
 
 class BertPooler(nn.Module):