ncnnqat

ncnnqat is a quantize aware training package for NCNN on pytorch.

Table of Contents

• ncnnqat
  • Table of Contents
  • Installation
  • Usage
  • Code Examples
  • Results
  • Todo

Installation

• Supported Platforms: Linux

• Accelerators and GPUs: NVIDIA GPUs via CUDA driver 10.1.

• Dependencies:

  • python >= 3.5, < 4
  • pytorch >= 1.6
  • numpy >= 1.18.1
  • onnx >= 1.7.0
  • onnx-simplifier >= 0.3.6

• Install ncnnqat via pypi:

  $ pip install ncnnqat (to do....)

  It is recommended to install from the source code

• or Install ncnnqat via repo：

  $ git clone https://github.com/ChenShisen/ncnnqat
  $ cd ncnnqat
  $ make install

Usage

• register_quantization_hook and merge_freeze_bn

  (suggest finetuning from a well-trained model, do it after a few epochs of training otherwise.)

  from ncnnqat import unquant_weight, merge_freeze_bn, register_quantization_hook
  ...
  ...
      for epoch in range(epoch_train):
          model.train()
      if epoch==well_epoch:
          register_quantization_hook(model)
      if epoch>=well_epoch:
          model = merge_freeze_bn(model)  #it will change bn to eval() mode during training
  ...

• Unquantize weight before update it

  ...
  ... 
      if epoch>=well_epoch:
          model.apply(unquant_weight)  # using original weight while updating
      optimizer.step()
  ...

• Save weight and save ncnn quantize table after train

  ...
  ...
      onnx_path = "./xxx/model.onnx"
      table_path="./xxx/model.table"
      dummy_input = torch.randn(1, 3, img_size, img_size, device='cuda')
      input_names = [ "input" ]
      output_names = [ "fc" ]
      torch.onnx.export(model, dummy_input, onnx_path, verbose=False, input_names=input_names, output_names=output_names)
      save_table(model,onnx_path=onnx_path,table=table_path)
  
  ...

  if use "model = nn.DataParallel(model)",pytorch unsupport torch.onnx.export,you should save state_dict first and prepare a new model with one gpu,then you will export onnx model.

  ...
  ...
      model_s = new_net() #
      model_s.cuda()
      register_quantization_hook(model_s)
      #model_s = merge_freeze_bn(model_s)
      onnx_path = "./xxx/model.onnx"
      table_path="./xxx/model.table"
      dummy_input = torch.randn(1, 3, img_size, img_size, device='cuda')
      input_names = [ "input" ]
      output_names = [ "fc" ]
      model_s.load_state_dict({k.replace('module.',''):v for k,v in model.state_dict().items()}) #model_s = model     model = nn.DataParallel(model)
            
      torch.onnx.export(model_s, dummy_input, onnx_path, verbose=False, input_names=input_names, output_names=output_names)
      save_table(model_s,onnx_path=onnx_path,table=table_path)
      
  
  ...

Code Examples

Cifar10 quantization aware training example.

python test/test_cifar10.py

SSD300 quantization aware training example.

   ln -s /your_coco_path/coco ./tests/ssd300/data

   python -m torch.distributed.launch \
    --nproc_per_node=4 \
    --nnodes=1 \
    --node_rank=0 \
    ./tests/ssd300/main.py \
    -d ./tests/ssd300/data/coco

    python ./tests/ssd300/main.py --onnx_save  #load model dict, export onnx and ncnn table

Results

• Cifar10

  result：

  net	fp32(onnx)	ncnnqat	ncnn aciq	ncnn kl
  mobilenet_v2	0.91	0.9066	0.9033	0.9066
  resnet18	0.94	0.93333	0.9367	0.937

• SSD300(resnet18|coco)

  fp32:
   Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.193
   Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.344
   Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.191
   Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.042
   Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.195
   Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.328
   Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.199
   Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.293
   Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.309
   Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.084
   Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.326
   Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.501
  Current AP: 0.19269
  
  ncnnqat:
   Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.192
   Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.342
   Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.194
   Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.041
   Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.194
   Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.327
   Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.197
   Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.291
   Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.307
   Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.082
   Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.325
   Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.497
  Current AP: 0.19202
  

Todo

....