qnq 0.1.8

qnq's not quantization

QNQ -- QNQ's not quantization

version 0.1.8 2020.10.10

Description

The toolkit is for Techart algorithm team to quantize their custom neural network's pretrained model. The toolkit is beta now, you can contact me with email([e-mail]) for adding ops and fixing bugs.

How to install

pip install qnq

How to quantize

you can visit https://git.zwdong.com/zhiwei.dong/qnq_tutorial for more examples for QNQ.

1. Prepare your model.

   1. Check if your model contains non-class operator, like torch.matmul.
   2. If True, add from qnq.operators.torchfunc_ops import * to your code.
   3. Then use class replace non-class operator, you can refer fellow #! add by dongz

   class BasicBlock(nn.Module):
       expansion = 1
   
       def __init__(self, inplanes, planes, stride=1, downsample=None):
           super(BasicBlock, self).__init__()
           self.conv1 = conv3x3(inplanes, planes, stride)
           self.bn1 = nn.BatchNorm2d(planes)
           self.relu1 = nn.ReLU(inplace=True)
           self.relu2 = nn.ReLU(inplace=True)
           self.conv2 = conv3x3(planes, planes)
           self.bn2 = nn.BatchNorm2d(planes)
           self.downsample = downsample
           self.stride = stride
   
           #! add by dongz
           self.torch_add = TorchAdd()
   
       def forward(self, x):
           identity = x
   
           out = self.conv1(x)
           out = self.bn1(out)
           out = self.relu1(out)
   
           out = self.conv2(out)
           out = self.bn2(out)
   
           if self.downsample is not None:
               identity = self.downsample(x)
   
           #! add by dongz
           out = self.torch_add(out, identity)
           # out += identity
           out = self.relu2(out)
   
           return out
   

2. Prepare your loader.

   1. Your loader.__getitem__() should return a tuple like (data, label) or (data, index) or longer, but you should keep data(img, audio or other) is the very first one, that because qnq will use loader.__getitem__()[0] to forward your model.
   2. You should choose at least 1k data to calibration your quantized model.

3. Prepare pretrained checkpoints.

   1. Train your model and use torch.save() to save your checkpoints.
   2. Use checkpoints = torch.load(checkpoints_path) and model.load_state_dict(checkpoints) to load your checkpoints.

4. Quantize

   1. Add from qnq import quantize
   2. Call quantize(model, bit_width, data_loader, path).

How to eval with quantization

1. quantize() will turn on 'eval mode' for model, that will automatically quantize activation, and weight already be fixed-point right now.
2. Just call your origin version eval()

How to debug

1. Call quantize(model, bit_width, data_loader, path, is_debug=True).
2. Debug mode will plot every layer's stats.

Convert to ONNX model

• Follow https://git.zwdong.com/zhiwei.dong/qnq_tutorial to learn how to convert.

Operators supported

• Convolution Layers
  • Conv
  • ConvTranspose
• Pooling Layers
  • AveragePool
  • AdaptiveAvgPool
• Activation
  • Relu、Relu6
  • PRelu、LeakyRelu
• Normalization Layers
  • BatchNorm
  • LayerNorm
• Linear Layers
  • Linear
• Vision Layers
  • Upsample
• Torch Function
  • Add, Minus, DotMul, MatMul, Div, Exp
  • Sin, Cos
  • SoftMax, Sigmoid
  • TorchTemplate, TorchDummy