Delegate squeeze to avoid modifying :func:`onmt.translate.translator.Translator.translate_batch()`

Assert all arguments have the same value

Tiles x on dimension dim count times.

compute accuracy

Log translation.

compute perplexity

Builds the optimizer from options. Args: cls: The ``Optimizer`` class to instantiate. model: The model to optimize.

Zero the gradients of optimized parameters.

Detect old-style vocabs (``List[Tuple[str, torchtext.data.Vocab]]``). Args: vocab: some object loaded from a *.vocab.pt file Returns

Sets the random seed.

yield a `list` containing `shard_size` line of `path`, or repeatly generate `default` if `path` is None.

Line-by-line read a file as bytes.

This returns user-defined train/validate data iterator for the trainer to iterate over. We implement simple ordered iterator strategy here,

Initialize decoder state with last state of the encoder.

Read data into dicts. Args: vecs (str or Iterable[str]): Sequence of feature vector paths or path to file containi

rnn factory, Use pytorch version when available.

Creates a boolean mask from sequence lengths.

Update the model parameters based on current gradients. Optionally, will employ gradient modification or update learning rate.

Detect old-style nesting (``dict[str, List[Tuple[str, Field]]]``).

Raise missing dep exception with standard error message.

Yield elements from data in chunks of batch_size, where each chunk size is a multiple of batch_size_multiple. This is an extended version of

Return True if check pass; if check fails and abort is True, raise an Exception, othereise return False.

utility for retrieving polyak averaged params

Returns the current learning rate.

Prune statistics dict for efficiency of max() The frequency of a symbol pair never increases, so pruning is generally safe (until we the most

Translate `inputs` using this model Args: inputs (List[dict[str, str]]): [{"src": "..."},{"src": ...}] Returns:

Tokenize a single sequence. Args: sequence (str): The sequence to tokenize. Returns: tok (str): The tokenize

The main training loop by iterating over `train_iter` and possibly running validation on `valid_iter`. Args: tra

Update statistics by suming values with another `Statistics` object Args: stat: another statistic object upd

Creates a boolean if gpu used

Initializes the controller. Args: optimizer: A ``torch.optim.Optimizer`` instance. learning_rate: The initial learning rate.

Compute the loss. Subclass must define this method. Args: batch: the current batch. output: the predict out

A naive forward pass for transformer decoder. # TODO: change 1 to T as T could be 1 or tgt_len Args: inputs (FloatTensor)

Detect old-style text fields. Not old style vocab, old nesting, and text-type fields not using ``TextMultiField``. Args: vocab:

Args: loss (:obj:`FloatTensor`): the loss computed by the loss criterion. scores (:obj:`FloatTensor`): a score for ea

Sparsemax building block: compute the threshold Args: input: any dimension dim: dimension along which to apply the sparsemax

error handler

Gather a `Statistics` object accross multiple process/nodes Args: stat(:obj:Statistics): the statistics object to gather

Wrapper for backward pass. Some optimizer requires ownership of the backward pass.

Convert valid alignment matrix to i-j Pharaoh format.(0 indexed)

Build a model from opts. Args: model_opt: the option loaded from checkpoint. It's important that the opts have been updated a

Args: opt: the option in current environment. text_field(TextMultiField): word and feats field. for_encoder(bool): build

Given scores from an expanded dictionary corresponeding to a batch, sums together copies, with a dictionary word when it is ambiguous.