Convert Inception21k to PyTorch

[rtaori]

Hi,

I am trying to convert the Inception21k model to PyTorch. I used the MMdnn tool to do the conversion; however, I am only seeing 65% top1 accuracy on the validation set (below the reported 68%).

What was the image preprocessing used to train the network? Currently I am doing:

1. Resize shortest size of image to 256
2. Center crop to 224
3. Multiply by 255 to scale image input range to [0, 255]
4. Subtract mean of 117 from the image

Is this correct?