/caffe-int8-convert-tools

Generate a quantization parameter file for ncnn framework int8 inference

Primary LanguagePythonBSD 3-Clause "New" or "Revised" LicenseBSD-3-Clause

Caffe-Int8-Convert-Tools

This convert tools is base on TensorRT 2.0 Int8 calibration tools, which use the KL algorithm to find the suitable threshold to quantize the activions from Float32 to Int8(-127 - 127).

We provide the Classification(SqueezeNet_v1.1) and Detection(MobileNet_v1 SSD 300) demo based on ncnn(a high-performance neural network inference framework optimized for the mobile platform) and the community ready to support this implementation.

The pull request in ncnn

NCNN have a new convert tool to support Post-Training-Quantization

Using this new ncnn-quantization-tools, you can convert your ncnn model to ncnn int8 model directly. If you just want to deploy your model with ncnn,I suggest you use it.

Reference

For details, please read the following PDF:

8-bit Inference with TensorRT

MXNet quantization implementation:

Quantization module for generating quantized (INT8) models from FP32 models

An introduction to the principles of a Chinese blog written by my friend(bruce.zhang):

The implement of Int8 quantize base on TensorRT

HowTo

The purpose of this tool(caffe-int8-convert-tool-dev.py) is to test new features, such as mulit-channels quantization depend on group num.

This format is already supported in the ncnn latest version. I will do my best to transform some common network models into classification-dev

python caffe-int8-convert-tool-dev-weight.py -h
usage: caffe-int8-convert-tool-dev-weight.py [-h] [--proto PROTO] [--model MODEL]
                                  [--mean MEAN MEAN MEAN] [--norm NORM]
                                  [--images IMAGES] [--output OUTPUT]
                                  [--group GROUP] [--gpu GPU]

find the pretrained caffemodel int8 quantize scale value

optional arguments:
  -h, --help            show this help message and exit
  --proto PROTO         path to deploy prototxt.
  --model MODEL         path to pretrained caffemodel
  --mean MEAN           value of mean
  --norm NORM           value of normalize(scale value or std value)
  --images IMAGES       path to calibration images
  --output OUTPUT       path to output calibration table file
  --group GROUP         enable the group scale(0:disable,1:enable,default:1)
  --gpu GPU             use gpu to forward(0:disable,1:enable,default:0)
python caffe-int8-convert-tool-dev-weight.py --proto=test/models/mobilenet_v1.prototxt --model=test/models/mobilenet_v1.caffemodel --mean 103.94 116.78 123.68 --norm=0.017 --images=test/images/ output=mobilenet_v1.table --group=1 --gpu=1

How to use the output file(calibration-dev.table)

For example in MobileNet_v1_dev.table

conv1_param_0 0.0 3779.48337933 482.140562772 1696.53814502
conv2_1/dw_param_0 0 72.129143 149.919382 // the convdw layer's weight scale every group is 0.0 72.129 149.919 ......
......
conv1 49.466518
conv2_1/dw 123.720796 // the convdw layer's bottom blobchannel scale is 123.720
......

Three steps to implement the conv1 layer int8 convolution:

  1. Quantize the bottom_blob and weight:

    bottom_blob_int8 = bottom_blob_float32 * data_scale(49.466518)
    weight_int8 = weight_float32 * weight_scale(156.639840)
    
  2. Convolution_Int8:

    top_blob_int32 = bottom_blob_int8 * weight_int8
    
  3. Dequantize the TopBlob_Int32 and add the bias:

    top_blob_float32 = top_blob_int32 / [data_scale(49.466518) * weight_scale(156.639840)] + bias_float32
    

How to use with ncnn

quantized int8 inference

Accuracy and Performance

We use ImageNet2012 Dataset to complete some classification test.

Type Detail
Calibration Dataset ILSVRC2012_img_test   1k
Test Dataset ILSVRC2012_img_val    5k
Framework ncnn
Support Layer Convolution,ConvolutionDepthwise,ReLU

The following table show the Top1 and Top5 different between Float32 and Int8 inference.

Models FP32 INT8 Loss
Top1 Top5 Top1 Top5 Diff Top1 Diff Top5
SqueezeNet v1.1 57.78% 79.88% 57.82% 79.84% +0.04% -0.04%
MobileNet v1 67.26% 87.92% 66.74% 87.43% -0.52% -0.49%
GoogleNet 68.50% 88.84% 68.62% 88.68% +0.12% -0.16%
ResNet18 65.49% 86.56% 65.30% 86.52% -0.19% -0.04%
ResNet50 71.80% 89.90% 71.76% 90.06% -0.04% +0.16%

We use VOC0712,MSCOCO Dataset to complete some detection test.

Type Detail
Test Dataset VOC2007
Unit mAP (Class 20)
Models FP32 INT8 Loss
SqueezeNet SSD 61.80 61.27 -0.53
MobileNet_v1 SSD 70.49 68.92 -1.57

Speed up

The following table show the speedup between Float32 and Int8 inference. It should be noted that the winograd algorithm is enable in the Float32 and Int8 inference. The Hardware Platform is Hisi3519(Cortex-A17@880MHz)

Uint(ms) SqueezeNet v1.1 MobileNet v1 GoogleNet ResNet18 MobileNetv1 SSD SqueezeNet SSD
Float32 282 490 1107 985 970 610
Int8 192 369 696 531 605 498
Ratio x1.46 x1.33 x1.59 x1.85 x1.60 x1.22

Memory reduce

Runtime Memory : mbytes

Models fp32-wino63 int8-wino23 int8-wino43
squeezenet_v1_1 50 30 32
mobilenet_v1 61 35 35
mobilenet_v1_ssd 90 45 45
squeezenet_v1_ssd 210 70 94
resnet18 335 77 130
googlenet_v1 154 72 89

Storage Memory : mbytes

Models fp32 int8
squeezenet_v1_1 4.71 1.20
mobilenet_v1 16.3 4.31
mobilenet_v1_ssd 22.0 5.60
squeezenet_v1_ssd 21.1 5.37
resnet18 44.6 11.2
googlenet_v1 26.6 6.72

Contributor

Thanks to NVIDIA for providing the principle of correlation entropy and ncnn's author nihui sharing his neural network inference framework.

Thanks to the help from the following friends:

Optimization Instructor : Fugangping, bruce.zhang

Algorithm : xupengfeixupf, JansonZhu, wangxinwei, lengmm

Python : daquexian

License

BSD 3 Clause