Release Notes
MindSpore Golden Stick 0.3.0-alpha Release Notes
Major Features and Improvements
[stable] SLB(Searching for Low-Bit Weights in Quantized Neural Networks)QAT algorithm now support BatchNorm calibration. we can invoke
set_enable_bn_calibration
api to enable BatchNorm calibration. For a network with a BatchNorm layer, the BatchNorm calibration can reduces the decrease in network accuracy caused by the SLB quantization algorithm. (!150)[stable] We verified the quantization effect of SimQAT(Simulated Quantization Aware Training) algorithm and the SLB algorithm on the ResNet network and the Imagenet2012 dataset. For details, please refer to MindSpore Models readme.
[stable] SimQAT algorithm now support inference on MindSpore Lite backend. We quant the LeNet network with SimQAT and deploy it on ARM CPU. For details, please refer to Deployment Effect.
API Change
Backwards Compatible Change
SLB algorithm adds the
set_enable_bn_calibration
interface to enable or disable BatchNorm calibration.(!117)Add
convert
interface to the algorithm base class, which is configured to convert training network to inferring network. And the network will be exported to MindIR file for Deployment. For details, please refer to Model Deployment.(!176)Add
set_save_mindir
interface to the algorithm base class, which is configured to automatically export MindIR after training. For details, please refer to Model Deployment.(!168)
Bug fixes
[STABLE] Refactor SimQAT algorithm code, and solve bugs such as activation operator loss, pre-trained parameter loss, simulation quantization operators redundancy, etc.
Contributors
Thanks goes to these wonderful people:
liuzhicheng01, fuzhongqian, hangangqiang, yangruoqi713, kevinkunkun.
Contributions of any kind are welcome!