Converter

Rather than implementing a training code for ProgressiveNet, we provide a converter for generating progressive model from static TensorFlowJS model.

How to install

$ npm install -g progressivenet-converter

How to use

Prerequisites

To convert a model, you have to prepare TFJS models for conversion. TensorFlow Hub provides various types of pretrained models in free. Or you can generate your own TFJS models by converting the TF models.

You can filter the model format in TFHub, thus you can easily find the TFJS models in here. The TFJS model has the structure like below:

group1-shard1of5.bin
group1-shard2of5.bin
group1-shard3of5.bin
group1-shard4of5.bin
group1-shard5of5.bin
model.json

Convert model

To convert a TFJS model into a progressive model, run command like below:

$ progressivenet-converter <TFJS_MODEL_PATH> <OUT_DIR> [<INTERFACE>]

INTERFACE indicates the bit-width list in the paper. For example, if you set it to 4,4,4,4, then the model is divided to four parts, allowing [4,8,12,16]-bit models progressively.

Here are the examples:

$ progressivenet-converter ./mobilenet_v2 ./mobilenet_v2_2222 2,2,2,2
$ progressivenet-converter ./mobilenet_v2 ./mobilenet_v2_44816 4,4,8,16

Running in local

To build and run convert in local, run commands below:

# Link progressivenet package
$ cd ../progressivenet
$ npm link
$ cd ../converter
$ npm link progressivenet

# Install packages
$ npm install

# Compile typescript
$ npm run build

# Register command line
$ npm install -g .