machine-translation-nar-de-en-0002

Use Case and High-Level Description

This is a Deutsch-English machine translation model based on non-autoregressive Transformer topology.

Tokenization occurs using the SentencePieceBPETokenizer (see the demo code for implementation details) and the enclosed tokenizer_src and tokenizer_tgt folders.

Specification

Metric

Value

GOps

23.19

MParams

77.47

Source framework

PyTorch*

Accuracy

The quality metrics were calculated on the wmt19-en-de dataset (test split in lower case).

Metric

Value

BLEU

21.4 %

Input

name: tokens shape: 1, 150 description: sequence of tokens (integer values) representing the tokenized sentence. The sequence structure is as follows (<s>, </s> and <pad> should be replaced by corresponding token IDs as specified by the dictionary): <s> + tokenized sentence + </s> + (<pad> tokens to pad to the maximum sequence length of 150)

Output

name: pred shape: 1, 200 description: sequence of tokens (integer values) representing the tokenized translation. The sequence structure is as follows (<s>, </s> and <pad> should be replaced by corresponding token IDs as specified by the dictionary): <s> + tokenized sentence + </s> + (<pad> tokens to pad to the maximum sequence length of 150)