# SLIM Quantization-aware training (QAT) on INT8 MKL-DNN
This document describes how to use [Paddle Slim](https://github.com/PaddlePaddle/FluidDoc/blob/develop/doc/fluid/advanced_usage/paddle_slim/paddle_slim.md) to convert a quantization-aware trained model to an INT8 MKL-DNN runnable model which has almost the same accuracy as QAT on GoogleNet, MobileNet-V1, MobileNet-V2, ResNet-101, ResNet-50, VGG16 and VGG19. We provide the accuracy results compared with fake QAT accuracy by running the QAT trained model with MKL-DNN int8 kernel on above 7 models.
## 0. Prerequisite
You need to install at least PaddlePaddle-1.5 python package `pip install paddlepaddle==1.5`.
## 1. How to generate INT8 MKL-DNN QAT model
You can refer to the unit test in [test_quantization_mkldnn_pass.py](test_quantization_mkldnn_pass.py). Users firstly use PaddleSlim quantization strategy to get a saved fake QAT model by [QuantizationFreezePass](https://github.com/PaddlePaddle/models/tree/develop/PaddleSlim/quant_low_level_api), then use the `TransformForMkldnnPass` to get the graph which can be run with MKL-DNN INT8 kernel. In Paddle Release 1.5, this pass only supports `conv2d` and `depthwise_conv2d` with channel-wise quantization for weights.
> Notes: The above commands will cost maybe several hours in the prediction stage (include int8 prediction and fp32 prediction) since there have 50000 pictures need to be predicted in `int8_full_val.bin`. User can set `OMP_NUM_THREADS` to the max number of physical cores of the used server to accelerate the process.
* The above commands will cost maybe several hours in the prediction stage (include int8 prediction and fp32 prediction) since there have 50000 pictures need to be predicted in `int8_full_val.bin`
* Running the above command with environment variable `FLAGS_use_mkldnn=true` will make the FP32 part of the test running using MKL-DNN (the INT8 part uses MKL-DNN either way).
