MasterThesis

Neural Sign Language By Learning Tokenization

Tokenization

CNN Training

They produce CNN models.

CNN Feature Extraction

They produce tfrecord and numpy features of the translation dataset. Their form can be right-hand, full frame etc.

NMT Module

It uses OpenNMT-tf library. There are four moddels: Bahdanau, Loung, Transformer, Multi-Source

Training a NMT Module

onmt train_and_eval --model_type [PythonModelFilePath] --config [ConfigYAMLFilePath]

Creating Test and Eval Translation Sentences

Average Models

onmt-average-checkpoints
--model_dir [ModelDirectory]
--output_dir [OutputDirectory]
--max_count 10

Infer Sentences

onmt-main infer
--config conf.yml
--features_file [TestFeatureFilePath]
--predictions_file [OutputPredictonFilePath]
--checkpoint_path [ModelPath(AverageModel)]

Evaluation

Use eval scripts evaluation_utils and pbs Code: evaluation_utils.evaluate(ref_file=,tran_file=, metric=)
Script: python3 pbs.py -reference-file REFERENCE_FILE --baseline-file BASELINE_FILE --sample-size SAMPLE_SIZE