ptq

Step-by-Step

This document is used to list steps of reproducing TensorFlow Wide & Deep tuning zoo result.

Prerequisite

1. Installation

# Install Intel® Neural Compressor
pip install neural-compressor

2. Install Intel Tensorflow

pip install intel-tensorflow

Note: Supported Tensorflow Version.

3. Install Additional Dependency packages

cd examples/tensorflow/recommendation/wide_deep_large_ds/quantization/ptq
pip install -r requirements.txt

4. Prepare Dataset

Download training dataset: (8 million samples)

$ wget https://storage.googleapis.com/dataset-uploader/criteo-kaggle/large_version/train.csv

Download evaluation dataset (2 million samples)

$ wget https://storage.googleapis.com/dataset-uploader/criteo-kaggle/large_version/eval.csv

5. Process Dataset

Process calib dataset

python preprocess_csv_tfrecords.py \
        --inputcsv-datafile train.csv \
        --calibrationcsv-datafile eval.csv \
        --outputfile-name processed_data

Process eval dataset

python preprocess_csv_tfrecords.py \
        --inputcsv-datafile eval.csv \
        --calibrationcsv-datafile train.csv \
        --outputfile-name processed_data

Two .tfrecords files are generated and will be used later on:

1. train_processed_data.tfrecords
2. eval_processed_data.tfrecords

6. Download Frozen PB

wget https://storage.googleapis.com/intel-optimized-tensorflow/models/v1_6/wide_deep_fp32_pretrained_model.pb

7. Run Command

The cmd of running WnD

bash run_tuning.sh --dataset_location=/path/to/datasets --input_model=/path/to/wide_deep_fp32_pretrained_model.pb --output_model=./wnd_int8_opt.pb
bash run_benchmark.sh --dataset_location=/path/to/datasets --input_model=./wnd_int8_opt.pb --mode=accuracy --batch_size=500
bash run_benchmark.sh --dataset_location=/path/to/datasets --input_model=./wnd_int8_opt.pb --mode=benchmark --batch_size=500

Other

This example takes the reference from https://github.com/IntelAI/models/tree/master/benchmarks/recommendation/tensorflow/wide_deep_large_ds. The pretrained model was trained with preprocessed data from dataset Criteo.