Difference in training speed between basic PyTorch and PyTorch Lightning solutions

[philippejuhel]

Hi,

I write 2 toy regression programs to compare speed of training.

The first one uses only PyTorch basic functions to program a 2 inputs - 1 output NN (program regressionSimpleXY_Z.txt to be rename in .py)

regressionSimpleXY_Z.txt

The second one uses PyTorch Lightning with the same NN (program predictionWithLightning.txt)
predictionWithLightning.txt

When I run the first program, it takes 0.17s to perform the fit function but when I run the second program, it takes more than 30s with a NVIDIA 2080 GPU card. And for the second program, the number of epochs is 10 times (100 epochs) lower than for the first program (1000 epochs).

How can we explain this?

Regards,

Philippe