背景:

量化PoseNet(hrnet开源代码中, 基于resnet backbone的模型)过程中, 在线模型量时, 如果开启分通道量化, 则在离线模型生成时, 会报错, 错误发生在nn.ConvTranspose2d算子上:

===============================================================

[cnmlError] Filter DIM_C is: 256, but the size of Pos and Scale is: 512.

[cnmlError] Deconv Op channel quant: the size of Pos and Scale should be equal with filter DIM_C

[ERROR][/pytorch/catch/torch_mlu/csrc/aten/operators/cnml/internal/convtranspose2d_internal.cpp][line:183][cnml_convtranspose2d_internal][thread:140613578876672][process:9298]:

CNML error: CNML_STATUS_UNSUPPORT

===============================================================

如果关闭分通道量化, 则成功生成离线模型.

环境背景:

官网获取的docker镜像: cambricon_pytorch_u16.04_v1.7.0_pt0.15.0, 驱动版本v1.7.0