rknn toolkit 1.6.1 量化 ELU 不支持？

[复制链接] · 发表于 2021-9-20 10:51:16

onnx模型转RNKK模型，进行精度分析，生成entire_qnt_error_analysis.txt，好像ELU算子不支持，如何解决？

Conv_Conv_48_80_out0_nhwc_1_32_57_128.tensor                  eculidean_norm=0.765876 cosine_norm=0.706717 eculidean=1.237127 cosine=0.706717
Elu_Elu_49_79_out0_nhwc_1_32_57_128.tensor                eculidean_norm=0.765806 cosine_norm=0.706771 eculidean=1.231775 cosine=0.706771
Conv_Conv_50_78_out0_nhwc_1_32_57_128.tensor                  eculidean_norm=0.000000 cosine_norm=nan eculidean=0.000000 cosine=nan
Elu_Elu_51_77_out0_nhwc_1_32_57_128.tensor                eculidean_norm=0.000000 cosine_norm=nan eculidean=0.000000 cosine=nan
Conv_Conv_52_76_out0_nhwc_1_32_57_128.tensor                  eculidean_norm=0.000000 cosine_norm=nan eculidean=0.000000 cosine=nan
Elu_Elu_53_72_out0_nhwc_1_32_57_128.tensor                eculidean_norm=0.000000 cosine_norm=nan eculidean=0.000000 cosine=nan
Conv_Conv_54_68_out0_nhwc_1_32_57_128.tensor                  eculidean_norm=0.000000 cosine_norm=nan eculidean=0.000000 cosine=nan
Elu_Elu_55_63_out0_nhwc_1_32_57_128.tensor                eculidean_norm=0.000000 cosine_norm=nan eculidean=0.000000 cosine=nan
Conv_Conv_56_59_out0_nhwc_1_32_57_128.tensor                  eculidean_norm=0.000000 cosine_norm=nan eculidean=0.000000 cosine=nan
Elu_Elu_57_55_out0_nhwc_1_32_57_128.tensor                eculidean_norm=0.000000 cosine_norm=nan eculidean=0.000000 cosine=nan
Conv_Conv_58_50_out0_nhwc_1_32_57_128.tensor                  eculidean_norm=0.000000 cosine_norm=nan eculidean=0.000000 cosine=nan
Elu_Elu_59_46_out0_nhwc_1_32_57_128.tensor                eculidean_norm=0.000000 cosine_norm=nan eculidean=0.000000 cosine=nan

网络里用到的ELU
def conv_dw_no_bn(in_channels, out_channels, kernel_size=3, padding=1, stride=1, dilation=1):
return nn.Sequential(
      nn.Conv2d(in_channels, in_channels, kernel_size, stride, padding, dilation=dilation, groups=in_channels, bias=False),
      nn.ELU(inplace=True),

      nn.Conv2d(in_channels, out_channels, 1, 1, 0, bias=False),
      nn.ELU(inplace=True),
)