Where do the VGG16 weights come from?


Does anyone know if the VGG16 weights from PyTorch are loaded from the Caffe model proposed by the authors (Visual Geometry Group - University of Oxford)?

I was comparing the results from PyTorch and TensorFlow and despite the prediction classes being the same, the probability is different. Since the VGG16 preprocessing from PyTorch and TensorFlow are different, I am wondering if they trained the VGG16 instead of loading the weights from somewhere else?

Thank you.

TorchVision models are trained the models using the scripts in the references directory of the source tree.

Best regards


Thank you, Thomas.