epwalsh

class CopyNetSeq2Seq(Model):

    # snip...
    
    def _decoder_step(self,
                      last_predictions: torch.Tensor,
                      selective_weights: torch.Tensor,
                      state: Dict[str, torch.Tensor]) -> Dict[str, torch.Tensor]:
        # shape: (group_size, max_input_sequence_length, encoder_output_dim)
        encoder_outputs_mask = state["source_mask"].float()

epwalsh

class CopyNetSeq2Seq(Model):

    # snip...
    
    def _get_copy_scores(self, state: Dict[str, torch.Tensor]) -> torch.Tensor:
        # NOTE: here `trimmed_source_length` refers to the input sequence length minus 2,
        # so that the special START and END tokens in the source are ignored. We also need to
        # ignore PAD tokens, but that happens elsewhere using a mask.
        # shape: (batch_size, trimmed_source_length, encoder_output_dim)
        trimmed_encoder_outputs = state["encoder_outputs"][:, 1:-1]

epwalsh

class CopyNetSeq2Seq(Model):
    
    # snip...
    
    def _get_generation_scores(self, state: Dict[str, torch.Tensor]) -> torch.Tensor:
        # `self._output_generation_layer` is just a PyTorch linear layer with an input
        # dimension equal to the decoder hidden state size, and an output dimension
        # equal to the size of the target vocabulary.
        return self._output_generation_layer(state["decoder_hidden"])

epwalsh

{    
    "trainer": {
        "cuda_device": 0,
        "learning_rate_scheduler": {
            "type": "triangular",
            // total number of epochs, should match the trainner param `num_epochs` below
            "num_epochs": 80,
            // increase LR linearly for 20 epochs
            "warm_up": 20,
            // then decrease LR linearly for 30 epochs

epwalsh

Initial setup

Step 1: Fork the repo.

Step 2: Clone your fork locally.

git clone https://github.com/USERNAME/REPO.git

epwalsh

find ~/ISU-DMC/dmc2016 -name '*.h' -o -name '*.R' > RFILES
find ~/ISU-DMC/dmc2016 -name '*.h' -o -name '*.py' > PYFILES