Extremely long time-series data leads causes evaluation failure #462
Labels
bug
Something isn't working
Comments
|
Hi, |
|
Hi! The full log file and run history json file are as below: |
|
The problem is a bit complicated; this dataset only contains integral values. Therefore, it is considered a 'multiclass' dataset and all its 0 values are transformed into a very small value. Thereby, the target values are not properly scaled. |
10 tasks
|
Hi! a snapshot of one of the training sessions in the log file above. |
|
if you want to reproduce the results from the paper, you could use the code from this branch: https://github.com/dengdifan/Auto-PyTorch/tree/ecml22_apt_ts (I did some modifications to make it compatible with the current workflow of AutoPyTorch). The metric 'MASE' is only applied to evaluate the trained model as a loss that will be returned to SMAC. Depending on the head type, networks could be trained with different losses (for instance, this implementation is trained with negative log-likelihood loss (DistributionLoss)). Normally I would recommend not evaluating validation metrics at each iteration, as it might be too expensive for some auto-regressive networks to do inference at every iteration (for the same reason, the |
|
Thanks for your reply. I will try to run the code you provided. |
|
I am closing this issue due to inactivity. Feel free to reopen if the issue persists. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
NOTE: ISSUES ARE NOT FOR CODE HELP - Ask for Help at https://stackoverflow.com
Your issue may already be reported!
Also, please search on the issue tracker before creating one.
Issue Description
Expected Behavior
Evaluate the performance normally of the initial designs on Covid Death Daily Dataset.
Current Behavior
An error "ValueError: Expected parameter df (Tensor of shape (32, 3, 1)) of distribution Chi2() to satisfy the constraint GreaterThan(lower_bound=0.0)" occurs.
Possible Solution
Maybe there's a default maximum length of time-series sequence I didn't notice.
Your Code
Error message
Further information are recorded in following log file:
nan-error.log
Your Local environment
Ubuntu 20.04
Python 3.8
pip freezeorconda listpytorch=1.12+cu116
Supplementary Information
Covid Death dataset is uni-variate dataset, to give more evidence about what's really happening behind this error, the following
are the print-out dataframe of
y_trainandy_test. (In order to form a complete control group following the format of example code, I further modify the data frequency from '1D' to '1Y' to be identical with Longley dataset just for debug purpose.)nan-error-y-info.txt
The text was updated successfully, but these errors were encountered: