nixtla / transfer-learning-time-series Goto Github PK

great job thanks
please explain your conclusion that TS deep NN transformers performance are bad
may you clarify what this your video
https://youtu.be/r43JnfHrmJk?t=913

from one hand you confirm your conclusion that TS deep NN transformers performance are bad

from another hand you say
you developed ML model which uses TS deep NN transformers and it is very good

Very minor, but the link in the readme is not working:

https://nixtla-public.s3.amazonaws.com/transfer/pretrained_models/nhits_m4_monthyl.ckpt

It should read instead:

https://nixtla-public.s3.amazonaws.com/transfer/pretrained_models/nhits_m4_monthly.ckpt

The notebook associated with this repository uses the model 'MQNHITS', which is nowhere to be found. The pretrained models may be based on vanilla NHITS and NBEATS but that isn't working. I don't know what 'h' and 'input_size' was used to train the models. Still, I tried with h=24 with was mentioned somewhere in the notebook, which resulted in another error.

model = NHITS.load_from_checkpoint(checkpoint_path='C://Users//raghu//Downloads//nhits_m4_hourly.ckpt',h=24,input_size=24*3)

AssertionError: max is not in ['MaxPool1d', 'AvgPool1d']

If I try to pass 'MaxPool1d' in another argument for pooling then too, there is an error

RuntimeError: Error(s) in loading state_dict for NBEATS:
Missing key(s) in state_dict: "blocks.0.layers.0.weight", "blocks.0.layers.0.bias", "blocks.0.layers.1.weight", "blocks.0.layers.1.bias",

And so on
Same kinda error with NBEATS as well.

So, just let me know if this project is updated and working