Unverified Commit 94dae690 authored by Myle Ott's avatar Myle Ott Committed by GitHub
Browse files

Merge pull request #77 from facebookresearch/oss-merge-internal

parents d74f200a 0a836276

fairseq/models/fairseq_decoder.py

+3 −0
Changes for fairseq/models/fairseq_decoder.py: 3 added lines, 0 removed lines.
Original line number Diff line number Diff line
@@ -18,3 +18,6 @@ class FairseqDecoder(nn.Module): 
    def max_positions(self):

        """Maximum input length supported by the decoder."""

        raise NotImplementedError



    def upgrade_state_dict(self, state_dict):

        return state_dict

fairseq/models/fairseq_encoder.py

+3 −0
Changes for fairseq/models/fairseq_encoder.py: 3 added lines, 0 removed lines.
Original line number Diff line number Diff line
@@ -18,3 +18,6 @@ class FairseqEncoder(nn.Module): 
    def max_positions(self):

        """Maximum input length supported by the encoder."""

        raise NotImplementedError



    def upgrade_state_dict(self, state_dict):

        return state_dict

fairseq/models/fairseq_model.py

+5 −0
Changes for fairseq/models/fairseq_model.py: 5 added lines, 0 removed lines.
Original line number Diff line number Diff line
@@ -43,6 +43,11 @@ class FairseqModel(nn.Module): 
        """Maximum output length supported by the decoder."""

        return self.decoder.max_positions()



    def upgrade_state_dict(self, state_dict):

        state_dict = self.encoder.upgrade_state_dict(state_dict)

        state_dict = self.decoder.upgrade_state_dict(state_dict)

        return state_dict



    def make_generation_fast_(self, **kwargs):

        """Optimize model for faster generation."""

        if self._is_generation_fast:

fairseq/models/fconv.py

+13 −2
Changes for fairseq/models/fconv.py: 13 added lines, 2 removed lines.
Original line number Diff line number Diff line
@@ -58,7 +58,7 @@ class FConvEncoder(FairseqEncoder): 
        self.projections = nn.ModuleList()

        self.convolutions = nn.ModuleList()

        for (out_channels, kernel_size) in convolutions:

            pad = (kernel_size - 1) // 2

            pad = (kernel_size - 1) / 2

            self.projections.append(Linear(in_channels, out_channels)

                                    if in_channels != out_channels else None)

            self.convolutions.append(
@@ -154,6 +154,7 @@ class FConvDecoder(FairseqIncrementalDecoder): 
                 max_positions=1024, convolutions=((512, 3),) * 20,

                 attention=True, dropout=0.1):

        super().__init__()

        self.register_buffer('version', torch.Tensor([2]))

        self.dictionary = dictionary

        self.dropout = dropout
@@ -265,6 +266,16 @@ class FConvDecoder(FairseqIncrementalDecoder): 
        """Maximum output length supported by the decoder."""

        return self.embed_positions.num_embeddings - self.dictionary.pad() - 1



    def upgrade_state_dict(self, state_dict):

        if state_dict.get('decoder.version', torch.Tensor([1]))[0] < 2:

            # old models use incorrect weight norm dimension

            for i, conv in enumerate(self.convolutions):

                # reconfigure weight norm

                nn.utils.remove_weight_norm(conv)

                self.convolutions[i] = nn.utils.weight_norm(conv, dim=0)

            state_dict['decoder.version'] = torch.Tensor([1])

        return state_dict



    def _split_encoder_out(self, encoder_out):

        """Split and transpose encoder outputs.
@@ -307,7 +318,7 @@ def LinearizedConv1d(in_channels, out_channels, kernel_size, dropout=0, **kwargs 
    std = math.sqrt((4 * (1.0 - dropout)) / (m.kernel_size[0] * in_channels))

    m.weight.data.normal_(mean=0, std=std)

    m.bias.data.zero_()

    return nn.utils.weight_norm(m)

    return nn.utils.weight_norm(m, dim=2)





def ConvTBC(in_channels, out_channels, kernel_size, dropout=0, **kwargs):

fairseq/modules/conv_tbc.py

+1 −1
Changes for fairseq/modules/conv_tbc.py: 1 added line, 1 removed line.
Original line number Diff line number Diff line
@@ -59,7 +59,7 @@ class ConvTBCFunction(Function): 
        kernel_size = weight_size[0]



        output = input.new(

            input_size[0] - kernel_size + 1 + pad * 2,

            input_size[0] - kernel_size + 1 + int(pad * 2),

            input_size[1],

            weight_size[2])