Inceptionv3 weight is not consistent with tensorflow

I am trying to reproduce the Frechet Inception Distance with Pytorch. The original code uses InceptionV3 trained with Tensorflow. But using the pretrained weight provided by Pytorch, the behavior is much different with Tensorflow version.

After some research I found that the weight is totally different. For example, the first convolution kernel data of Pytorch is:

> dic['Conv2d_1a_3x3.conv.weight'][0,0]
tensor([[-0.2103, -0.3441, -0.0344],
        [-0.1420, -0.2520, -0.0280],
        [ 0.0736,  0.0183,  0.0381]])

While tensorflow is

> x=f['weights'].value
> x.shape
Out[15]: (3, 3, 3, 32)
> x[:,:,0,0]
array([[ 0.01260555, -0.00162022,  0.09091024],
       [-0.10557341, -0.15358226, -0.04656353],
       [-0.16552085, -0.17688492, -0.10908931]], dtype=float32)

I wonder how the weight of Pytorch is obtained. And is there a way of perfectly copying tensorflow's weight to Pytorch?

I noticed there are repo for converting Inceptionv3 to Torch (https://github.com/Moodstocks/inception-v3.torch), but it does not work for Pytorch. Can anyone help me? Thank you very much.

Start by comparing the reported first convolution weights and tensor shapes between the TensorFlow and PyTorch InceptionV3 models. Determine whether the discrepancy comes from weight conversion or model behavior, and document a verified conversion path or the reason the weights cannot match exactly.