pnnx.py and ncnn.py for resnet18 has a slight different final outputs

[Bing1002]

error log | 日志或报错信息 | ログ

pnnx.py output:
$ python resnet18_torchscript_pnnx.py
tensor([[-1.5752e+00, 6.8381e-01, 1.4599e+00, 1.1986e+00, 1.0503e+00,
-5.0585e-01, 9.1963e-01, 1.4127e-01, -1.2256e+00, -3.8200e-01,
1.1840e+00, 2.5817e+00, 1.3319e+00, 2.8250e+00, 2.6328e+00,
1.1827e+00, 1.6862e+00, 2.9742e-01, 1.5851e+00, 1.9562e+00,
6.5654e-01, 2.2326e+00, 1.8743e+00, 1.8355e+00, 5.0387e-01,
1.7909e-01, 4.9693e-01, 6.8537e-01, -1.9073e-01, -1.3404e+00,
-5.2751e-01, 1.2330e+00, -5.4223e-01, 9.8114e-01, 2.2979e+00,
-5.1950e-01, -4.2796e-01, -1.1896e-01, 1.3294e+00, 1.9918e-01,
2.3761e+00, 1.0134e+00, 1.8879e+00, 3.5642e-01, 1.6943e+00,
4.4768e-01, 2.1218e+00, -4.0468e-01, -9.0379e-02, -7.7178e-02,

ncnn.py output:
$ python resnet18_torchscript_ncnn.py
tensor([[-1.5879e+00, 6.8066e-01, 1.4541e+00, 1.1836e+00, 1.0615e+00,
-5.0098e-01, 9.2188e-01, 1.5088e-01, -1.2207e+00, -3.6450e-01,
1.1777e+00, 2.5645e+00, 1.3398e+00, 2.8066e+00, 2.6289e+00,
1.1924e+00, 1.6914e+00, 2.9736e-01, 1.5830e+00, 1.9443e+00,
6.4844e-01, 2.2285e+00, 1.8730e+00, 1.8477e+00, 5.0488e-01,
1.8408e-01, 5.0098e-01, 6.8896e-01, -1.8799e-01, -1.3389e+00,
-5.3809e-01, 1.2305e+00, -5.4541e-01, 9.7656e-01, 2.2754e+00,
-5.2344e-01, -4.3628e-01, -1.1523e-01, 1.3418e+00, 2.1558e-01,
2.3828e+00, 1.0273e+00, 1.8945e+00, 3.7061e-01, 1.7021e+00,
4.5581e-01, 2.1328e+00, -3.9160e-01, -8.2886e-02, -5.1147e-02,

Even i changed tensor.contiguous(), the difference is still there. I want to know if this minor difference is normal.

context | 编译/运行环境 | バックグラウンド

Raspberry pi 5

how to reproduce | 复现步骤 | 再現方法

more | 其他 | その他