Skip to content

Latest commit

 

History

History
119 lines (92 loc) · 4.04 KB

README.md

File metadata and controls

119 lines (92 loc) · 4.04 KB

Step-by-Step

This document is used to list steps of reproducing TensorFlow style transfer Intel® Low Precision Optimization Tool tuning zoo result.

Prerequisite

1. Installation

# Install Intel® Low Precision Optimization Tool
pip instal lpot

2. Install Intel Tensorflow 1.15/2.0/2.1

pip intel-tensorflow==1.15.2 [2.0,2.1]

3. Install Additional Dependency packages

cd examples/tensorflow/style_transfer && pip install -r requirements.txt

4. Prepare Dataset

There are two folders named style_images and content_images you can use these two folders to generated stylized images for test you can also prepare your own style_images or content_images

5. Prepare Pretrained model

Automated approach

Run the prepare_model.py script located in LowPrecisionInferenceTool/examples/tensorflow/style_transfer.

usage: prepare_model.py [-h] [--model_path MODEL_PATH]

optional arguments:
  -h, --help            show this help message and exit
  --model_path MODEL_PATH directory to put models, default is ./model

Manual approach

wget https://storage.googleapis.com/download.magenta.tensorflow.org/models/arbitrary_style_transfer.tar.gz
tar -xvzf arbitrary_style_transfer.tar.gz ./model

Run Command

python style_tune.py --output_dir=./result --style_images_paths=./style_images --content_images_paths=./content_images --model_dir=./model --precision=quantized

Details of enabling Intel® Low Precision Optimization Tool on style transfer for Tensorflow.

This is a tutorial of how to enable style_transfer model with Intel® Low Precision Optimization Tool.

User Code Analysis

  1. User specifies fp32 model, calibration dataset q_dataloader, evaluation dataset eval_dataloader and metric in tuning.metric field of model-specific yaml config file.

  2. User specifies fp32 model, calibration dataset q_dataloader and a custom eval_func which encapsulates the evaluation dataset and metric by itself.

For style_transfer, we applied the latter one because we don't have metric for style transfer model.The first one is to implement the q_dataloader and implement a fake eval_func. As lpot have implement a style_transfer dataset, so only eval_func should be prepared after load the graph

Evaluation Part Adaption

As style transfer don't have a metric to measure the accuracy, we only implement a fake eval_func

def eval_func(model):
    return 1.

Write Yaml config file

In examples directory, there is a conf.yaml. We could remove most of items and only keep mandatory item for tuning. We also implement a calibration dataloader

model:
  name: style_transfer
  framework: tensorflow
  inputs: import/style_input,import/content_input
  outputs: import/transformer/expand/conv3/conv/Sigmoid

quantization:
  calibration:
    dataloader:
      batch_size: 2
      dataset:
        style_transfer:
          content_folder: ./content_images/          # NOTE: modify to content images path if needed
          style_folder: ./style_images/              # NOTE: modify to style images path if needed

evaluation:
  accuracy:
    dataloader:
      batch_size: 2
      dataset:
        style_transfer:
          content_folder: ./content_images/          # NOTE: modify to content images path if needed
          style_folder: ./style_images/              # NOTE: modify to style images path if needed

tuning:
    accuracy_criterion:
      relative: 0.01
    exit_policy:
      timeout: 0
    random_seed: 9527

Here we set the input tensor and output tensors name into inputs and outputs field. In this case we only calibration and quantize the model without tune the accuracy

Code update

After prepare step is done, we just need add 2 lines to get the quantized model.

from lpot import Quantization

quantizer = Quantization(args.config)
q_model = quantizer(graph, eval_func=eval_func)

The Intel® Low Precision Optimization Tool quantizer() function will return a best quantized model during timeout constrain.