encoder-converter 0.2.1

Convert huggingface encoder to onnx format.

Encoder converter

:rocket: Easy way to convert huggingface encoder model to other formats.

Description

Encoder converter is a package that allows you to convert the huggingface encoder model to other formats (e.g. onnx).

Features

Unfinished features will be implemented in future versions.

• Convert encoder model to onnx.
• Convert encoder model to openvino.
• Convert model with custom wrapper.

Installation

pip install encoder-converter

Usage

Run

convertencoder --model-name project/huggingface_repo --format onnx --output-dir /my/output/dir --cache-dir /cache/dir --output-model-name t5_encoder

Parameters

Parameter	Description	Default
--model-name	Huggingface model name
--format	Compiled model format. Available: onnx, openvino
--output-dir	Path to save compiled model and tokenizer artifacts.
--cache-dir	Path to a directory in which a downloaded pretrained model configuration should be cached while compiling.	/tmp
--output-model-name	If not specified, the default output model name will be parsed depends on the --model_name parameter.
If --output-model-name is not specified, you can find the complied model at output_dir/huggingface_repo.extension:

1. output_dir - --ouput-dir parameter.
2. huggingface_repo - extracts from the --model-name parameter.
3. extension - depends on selected format parameter:
   • onnx - onnx
   • openvino - xml