Fixes (#442)

Summary:
minor fixes:
1- adding fairseq logo
2- encoder padding for fconv self att
3- legacy ddp change
Pull Request resolved: https://github.com/pytorch/fairseq/pull/442

Differential Revision: D13651715

Pulled By: myleott

fbshipit-source-id: ac93c80f1dbffdfe03fbd4b8a8ea527aecb576a7

README.md

@@ -1,4 +1,4 @@
-# Introduction
+# Introduction <img src="fairseq_logo.png" width="50"> 
 
 Fairseq(-py) is a sequence modeling toolkit that allows researchers and
 developers to train custom models for translation, summarization, language

fairseq/legacy_distributed_data_parallel.py

@@ -141,6 +141,8 @@ class LegacyDistributedDataParallel(nn.Module):
             for param in self.module.parameters():
                 if not param.requires_grad:
                     continue
+                if param.grad is None:
+                    param.grad = torch.zeros_like(param)
                 if param.grad.requires_grad:
                     raise RuntimeError("DistributedDataParallel only works "
                                        "with gradients that don't require "

fairseq/models/fconv_self_att.py

@@ -195,6 +195,10 @@ class FConvEncoder(FairseqEncoder):
         # project to size of convolution
         x = self.fc1(x)
 
+        encoder_padding_mask = src_tokens.eq(self.padding_idx).t()  # -> T x B
+        if not encoder_padding_mask.any():
+            encoder_padding_mask = None
+
         # B x T x C -> T x B x C
         x = x.transpose(0, 1)
 
@@ -202,6 +206,9 @@ class FConvEncoder(FairseqEncoder):
         for proj, conv, attention in zip(self.projections, self.convolutions, self.attention):
             residual = x if proj is None else proj(x)
 
+            if encoder_padding_mask is not None:
+                x = x.masked_fill(encoder_padding_mask.unsqueeze(-1), 0)
+
             x = F.dropout(x, p=self.dropout, training=self.training)
             padding_l = (conv.kernel_size[0] - 1) // 2
             padding_r = conv.kernel_size[0] // 2
@@ -218,6 +225,10 @@ class FConvEncoder(FairseqEncoder):
         # project back to size of embedding
         x = self.fc2(x)
 
+        if encoder_padding_mask is not None:
+            encoder_padding_mask = encoder_padding_mask.t()  # -> B x T
+            x = x.masked_fill(encoder_padding_mask.unsqueeze(-1), 0)
+
         # scale gradients (this only affects backward, not forward)
         x = GradMultiply.apply(x, 1.0 / (2.0 * self.num_attention_layers))
 
@@ -226,6 +237,7 @@ class FConvEncoder(FairseqEncoder):
 
         return {
             'encoder_out': (x, y),
+            'encoder_padding_mask': encoder_padding_mask,  # B x T
         }
 
     def reorder_encoder_out(self, encoder_out, new_order):
@@ -233,6 +245,10 @@ class FConvEncoder(FairseqEncoder):
             eo.index_select(0, new_order) for eo in encoder_out['encoder_out']
         )
 
+        encoder_out['encoder_padding_mask'] = tuple(
+            eo.index_select(0, new_order) for eo in encoder_out['encoder_padding_mask']
+        )
+
         if 'pretrained' in encoder_out:
             encoder_out['pretrained']['encoder_out'] = tuple(
                 eo.index_select(0, new_order)