TensorFlow.js Inference API

This package provides a cli tool for model inferencing in a Node env. Additionally, the package is compiled to a binary with the same functionality, which allows to use tfjs for ML tasks in any envs and platforms.

The tool can be used to validate TFJS model execution against python results.

Note: This package is under development.

How to use

Run the cli tool in Node env.

Checkout the code: git clone https://github.com/tensorflow/tfjs.git

Go to tfjs-inference directory: cd tfjs-inference

Install dependencies: yarn

Run the inference:

ts-node src/index.ts --model_path=MODEL_PATH --inputs_dir=INPUTS_DIR
 *   --outputs_dir=OUTPUTS_DIR

The script expects three required arguments: model_path, inputs_dir and outputs_dir. There are also optional arguments, see the options below.

Options

model_path: Directory to a tfjs model json file.

inputs_dir: Directory to read the input tensor info and output info files.

outputs_dir: Directory to write the output files. Output files include: data.json, shape.json and dtype.json. Additionally, name.json is written if the model returns a map. The order of the output tensors follow the same order as the tf_output_name_file.

inputs_data_file: (Optional) Filename of the input data file. Default to data.json.

inputs_shape_file: (Optional) Filename of the input shape file. Default to shape.json

inputs_dtype_file: (Optional) Filename of the input dtype file. Default to dtype.json

tf_input_name_file: (Optional) Filename of the input name of the tf model. The input names should match the names defined in the signatureDef of the model. Default to tf_input_name.json

tf_output_name_file: (Optional) Filename of the output name of the tf model. The output names should match the names defined in the signatureDef of the model. Default to tf_output_name.json

backend: Choose which tfjs backend to use. Supported backends: cpu|wasm. Default to cpu.

Notes:

• The program requires these files to exist in the inputs_dir: inputs_data_file, inputs_shape_file, inputs_dtype_file, and tf_input_name_file.
• For model_path, absolute path is preferred. It also supports relative path, which should be relative to tfjs-inference directory.
• About the input and output formats. They are represented as array of tensors. The data.json, shape.json, dtype.json and tf_input_name.json files together represent the array of tensors. Example