search

philipperemy / keras-tcn

Keras Temporal Convolutional Network.
MIT License
1.89k stars 454 forks source link

Testing time_series_forecasting.py with version 3.3.1, the prediction results are very different from version 3.2.1 #193

Closed HalfMAI closed 3 years ago

HalfMAI commented 3 years ago

Describe the bug Testing time_series_forecasting.py with version 3.3.1, the prediction results are very different from version 3.2.1

Paste a snippet 3.3.1 notebook: image https://colab.research.google.com/drive/1xTjddAcNgNRpXzRXZbkq3HzmwTxsBVJg?usp=sharing

3.2.1 notbook: image https://colab.research.google.com/drive/1lgmAD2Em6rjxrOMj6wkcpabxvWbCEX84?usp=sharing

Dependencies tensorflow == 2.4.1

philipperemy commented 3 years ago

@HalfMAI we changed some of the default parameters in 3.3.1. In your first plot, it seems that the training got stuck at the beginning and the optimizer could not work properly. If you re-run a few times do you always see this behavior?

philipperemy commented 3 years ago

Can you try the 3.3.1 with the previous set of hyper parameters here: https://github.com/philipperemy/keras-tcn/commit/d3e01fd5a95146871487a0065959d87027c18ea7 (the one in red on the left are the previous ones).

HalfMAI commented 3 years ago

Can you try the 3.3.1 with the previous set of hyper parameters here: d3e01fd (the one in red on the left are the previous ones).

Thanks for the reply. After resetting the default parameters as before, it did work after a few tests. Here are the results after rebooting and testing several times. image image

philipperemy commented 3 years ago

@HalfMAI okay so might be worth reverting to the default parameters.