PyTorch consistent worse results than Torch with same model and data

1

Hi, I know it’s probably not a valid question to ask, but it really bothers me. I’m wondering if any of you have the same experience.

I’ve used PyTorch for a while and while it’s more elegant than Torch, its performance in term of accuracy are always worse than same models in Torch with same data and same model architecture.

I’ve used VGG-M and ResNet34 for my image data, which is of variable width so I add a global average pool layer to the standard VGG-M and ResNet34. No other adaptations were made except to change input to single channel.

Now, whether it’s VGG-M or ResNet, Pytorch results are always considerably worse than Torch. With ResNet, PyTorch accuracy is 89% versus torch’s 96%; with VGG, PyTorch 71% versus Torch 80%.

I use ResNet34 from torchvision.models for PyTorch, and Torch ResNet implementation from facebook. I use shortcut type “B” in fb.resnet.torch, which I think it’s the same as the ResNet34 model from torchvision.models.
For VGG-M, I built it using both Torch and PyTorch.

Optimization parameters are the same, as far as I can tell. No difference in terms of data loading.

Thanks for your suggestions.

2

Did you train from scratch?
From what I have heard from people, since pytorch is relatively new it may have some design issues such has initialization. People may have hard time reproducing results from a different environment such as tf or caffe.