# 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 INT8 MKL-DNN quantized model. In **Release 1.5**, we have released the QAT1.0 MKL-DNN which enabled the INT8 MKL-DNN kernel for QAT trained model within 0.05% accuracy diff on GoogleNet, MobileNet-V1, MobileNet-V2, ResNet-101, ResNet-50, VGG16 and VGG19. In **Release 1.6**, QAT2.0 MKL-DNN, we did the performance optimization based on fake QAT models: ResNet50, ResNet101, Mobilenet-v1, Mobilenet-v2, VGG16 and VGG19 with the minor accuracy drop. Compared with Release 1.5, the QAT2.0 MKL-DNN got better performance gain on inference compared with fake QAT models but got a little bit bigger accuracy diff. We provide the accuracy benchmark both for QAT1.0 MKL-DNN and QAT2.0 MKL-DNN, and performance benchmark on QAT2.0 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 INT8 MKL-DNN quantized model. In **Release 1.5**, we have released the QAT1.0 MKL-DNN which enabled the INT8 MKL-DNN kernel for QAT trained model within 0.05% accuracy diff on GoogleNet, MobileNet-V1, MobileNet-V2, ResNet-101, ResNet-50, VGG16 and VGG19. In **Release 1.6**, QAT2.0 MKL-DNN, we did the performance optimization based on fake QAT models: ResNet50, ResNet101, Mobilenet-v1, Mobilenet-v2, VGG16 and VGG19 with the minor accuracy drop. Compared with Release 1.5, the QAT2.0 MKL-DNN got better performance gain on inference compared with fake QAT models but got a little bit bigger accuracy diff. In **Release 1.7**, a support for [Ernie (NLP) QAT trained model](https://github.com/PaddlePaddle/benchmark/tree/master/Inference/c%2B%2B/ernie) was added to the QAT2.0 MKL-DNN. We provide the accuracy benchmark both for QAT1.0 MKL-DNN and QAT2.0 MKL-DNN, and performance benchmark on QAT2.0 MKL-DNN.
Notes:
* MKL-DNN and MKL are required. The performance gain can only be obtained with AVX512 series CPU servers.
* INT8 accuracy is best on CPU servers supporting AVX512 VNNI extension.
## 0. Prerequisite
You need to install at least PaddlePaddle-1.6 python package `pip install paddlepaddle==1.6`.
You need to install at least PaddlePaddle-1.7 python package `pip install paddlepaddle==1.7`.
## 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 `FakeQAT2MkldnnINT8KernelPass` to get the graph which can be run with MKL-DNN INT8 kernel. In Paddle Release 1.6, this pass supports `conv2d` and `depthwise_conv2d` ops with channel-wise quantization for weights. Apart from it, another pass called FakeQAT2MkldnnINT8PerfPass is available for use. This pass allows users to transform their QAT INT8 model into a highly performance-optimized model that is ran using INT8 MKL-DNN kernels.
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 `QatInt8MkldnnPass` (from QAT1.0 MKL-DNN) to get a graph which can be run with MKL-DNN INT8 kernel. In Paddle Release 1.6, this pass supports `conv2d` and `depthwise_conv2d` ops with channel-wise quantization for weights. Apart from it, another pass called `Qat2Int8MkldnnPass` (from QAT2.0 MKL-DNN) is available for use. In Release 1.6, this pass additionally supports `pool2d` op and allows users to transform their QAT model into a highly performance-optimized INT8 model that is ran using INT8 MKL-DNN kernels. In Release 1.7, a support for `fc`, `reshape2` and `transpose2` ops was added to the pass.
* FP32 Optimized Throughput (images/s) is from [int8_mkldnn_quantization.md](https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/inference/tests/api/int8_mkldnn_quantization.md).
>**IV. Ernie QAT2.0 MKL-DNN Accuracy on Intel(R) Xeon(R) Gold 6271**
Three steps to reproduce the above-mentioned accuracy results, and we take ResNet50 benchmark as an example:
* ### Prepare dataset
Three steps are needed to reproduce the above-mentioned accuracy and performance results. Below we explain the steps taking ResNet50 as an example of image classification models. In order to reproduce NLP results, please follow [this guide](https://github.com/PaddlePaddle/benchmark/tree/master/Inference/c%2B%2B/ernie/mkldnn/README.md).
### Prepare dataset
#### Image classification
In order to download the dataset for image classification models benchmarking, execute:
The converted data binary file is saved by default in `$HOME/.cache/paddle/dataset/int8/download/int8_full_val.bin`
* ### Prepare model
### Prepare model
#### Image classification
You can run the following commands to download ResNet50 model. The exemplary code snippet provided below downloads a ResNet50 QAT model. The reason for having two different versions of the same model originates from having two different QAT training strategies: One for an non-optimized and second for an optimized graph transform which correspond to QAT1.0 and QAT2.0 respectively.
You can run `qat_int8_comparison.py` with the following arguments to reproduce the accuracy result on ResNet50. The difference of command line between the QAT1.0 MKL-DNN and QAT2.0 MKL-DNN is that we use argument `qat2` to enable QAT2.0 MKL-DNN. To perform QAT2.0 MKL-DNN the performance test, the environmental variable `OMP_NUM_THREADS=1` and `batch_size=1` parameter should be set.
### Commands to reproduce benchmark
#### Image classification
You can use the `qat_int8_image_classification_comparison.py` script to reproduce the accuracy result on ResNet50. The difference between commands usedin the QAT1.0 MKL-DNN and QAT2.0 MKL-DNN is that for QAT2.0 MKL-DNN two additional options are required: the `--qat2` option to enable QAT2.0 MKL-DNN, and the `--quantized_ops` option with a comma-separated list of operators to be quantized. To perform the QAT2.0 MKL-DNN performance test, the environment variable `OMP_NUM_THREADS=1` and `--batch_size=1` option should be set.
1. Save QAT2.0 INT8 model. You can use the script `save_qat_model.py` for this purpose. It also requires the option `--quantized_ops` to indicate which operators are to be quantized.
> Notes: Due to a large amount of images contained in `int8_full_val.bin` dataset (50 000), the accuracy benchmark which includes comparison of unoptimized and optimized QAT model may last long (even several hours). To accelerate the process, it is recommended to set `OMP_NUM_THREADS` to the max number of physical cores available on the server.
