Similar to *forward* but only return features. Returns: tuple: - the decoder's features of shape `(batch

Filter examples that are too large Args: indices (np.array): original array of sample indices dataset (~fair

Args: tokens(Tensor): Input tokens(Bsz*beam, seq_len) lprobs(Tensor): likelihood probability, Expected to

Args: src_tokens (LongTensor): tokens in the source language of shape `(batch, src_len)` src_lengths

Get the data parallel group the caller rank belongs to.

Build sinusoidal embeddings. This matches the implementation in tensor2tensor, but differs slightly from the description in Section 3

Get mel-filter bank features via PyKaldi or TorchAudio. Prefer PyKaldi (faster CPP implementation) to TorchAudio (Python implementation). Note tha

Get speech features from .npy file or waveform from .wav/.flac file. The file may be inside an uncompressed ZIP file and is accessed via byte

Filters out the layers according to a regexp. Note that we omit biases. Args: - model: a nn.Module - filter_regexp: a re

Computes the log probabilities for all the words of the vocabulary, given a 2D tensor of hidden vectors.

Return set of symbol pairs in a word. Word is represented as tuple of symbols (symbols being variable-length strings).

Given a quantization configuration, get the right parameter for the module to be quantized. Args: - module: a nn.Module

Load all training state from a checkpoint file. rank = 0 will load the checkpoint, and then broadcast it to all other ranks.

Loads a checkpoint to CPU (with upgrading for backward compatibility). If doing single-GPU training or if the checkpoint is only being loaded by

Load the dictionary from the filename Args: filename (str): the filename

Legacy entry point to optimize model for faster generation. Prefer prepare_for_inference_.

Maximum input length supported by the encoder or decoder.

Multiplies grads by a constant ``c``.

Return a Tensor of `size` filled with a range function on the device of x. If size is empty, using the size of the variable x.

A context manager to disable gradient synchronization.

use the field in args to overrides those in cfg

Prune the given state_dict if desired for LayerDrop (https://arxiv.org/abs/1909.11556). Training with LayerDrop allows models to be robust to

Aggregate logging outputs from data parallel training.

Reorder encoder output according to *new_order*. Args: encoder_out: output from the ``forward()`` method new

Save all training state in a checkpoint file.

Helper to get index of sep symbol

Set the number of parameters updates.

Set the number of parameters updates.

Return an example's size as a float or tuple. This value is used when filtering a dataset with ``--max-positions``.