70e4d7fe60a9658bb27b9f5fb67592a1222b2ec3,spotlight/sequence/representations.py,CNNNet,user_representation,#CNNNet#Any#,114

Before Change

        (batch_size, seq_len, dim) = sequence_embeddings.size()

        // Move embedding dimensions to channels and add a fourth dim.
        sequence_embeddings = (sequence_embeddings
                               .permute(0, 2, 1)
                               .contiguous()
                               .view(batch_size, dim, seq_len, 1))

        x = sequence_embeddings
        for cnn_layer in self.cnn_layers:
            x = cnn_layer(x)

        user_representations = x.view(batch_size, dim, -1)
        pooled_representations = (user_representations
                                  .max(-1)[0]
                                  .view(batch_size, dim))

After Change

        for i, cnn_layer in enumerate(self.cnn_layers):
            // Pad so that the CNN doesn"t have the future
            // of the sequence in its receptive field.
            x = F.pad(x, (0, 0, self.kernel_width - min(i, 1), 0))
            x = F.relu(cnn_layer(x))

        x = x.squeeze(3)

Instances