Description

1. After using torch.jit, the performance of "predict" is improved by about 25%.
2. Changed parameter order for autoformer.xxxx_step.
3. Split AutoCorrelation.forward and trace the function using torch.jit.script.
4. torch.jit.optimized_execution can speed up model inference.

1. Why the change?

1. Speed up model inference.
2. Sync parameter order of TSDataset, because torch.jit will freeze the model.
3. AutoCorrelation.forward contains if, torch.jit.script to ensure that the model is tracked correctly and does not change the performance of the model.
4. We need to disable some optimizations, they may bring down the speed of prediction.
   Related link: https://discuss.pytorch.org/t/speed-of-custom-rnn-is-super-slow/63209/3

2. User API changes

yhat = autoformer.predict(data, use_jit=False) # use self.internal
# Inference speed(nyc_taxi): ~800ms

yhat = autoformer.predict(data, use_jit=True) # use torch.jit
# Inference speed(nyc_taxi): ~600ms

# tips: If we don't use `torch.jit.optimized_execution`, then jit_fp32 runs about 2s for the first and second time,

4. How to test?

• Unit test