I was going through the tensorflow keras tutorials, and tried repro their numbers in pytorch. It looks like the keras
.fit gets better accuracy out of the box compared to the pytorch flow. I tried making sure that the optimizer parameters are the same, and even tried initializing the pytorch weights to the ones in keras – still same.
This is the tutorial I was going through: Basic classification: Classify images of clothing | TensorFlow Core
Here is the notebook gist that shows the training and inference results for the same number of epochs: TF_vs_PyTorch.ipynb · GitHub
Is there anything I can do with the pytorch flow to roughly match the keras numbers?