Can't reproduce the result for UCR PLAID dataset

[Wensi-Tang]

Could you please give me a hand?

I found that cannot reproduce the result for PLAID dataset.

To prevent the problem from the dataset, I re-download the PLAID dataset at here. But the acc result is about 0.16387337032754107.

[hfawaz]

I am not sure what's the problem here but based on the link you have sent, it says that the time series length is 0.

[Wensi-Tang]

Thanks for your reply!

The length is shown 0 is not a problem. This is because the time series in this dataset are of different lengths. UCR uses nan padding to make them of the same length.

I adding a print(x_train) to your code here the output result is like this

[[[ 1.7339e-01]
  [ 1.3045e-01]
  [ 1.3499e-01]
  ...
  [        nan]
  [        nan]
  [        nan]]

And the acc result is 0.16387337032754107.

If you need any further information just tell me.

[hfawaz]

Okay so in order to reproduce results on the unequal length datasets, you should use the normalized version (to have equal length). it can be found here: https://www.cs.ucr.edu/~eamonn/time_series_data_2018/

[Wensi-Tang]

Hello,

Does the normalized version refer to the clickable link in this sentence "Then you can download the entire archive (about 260 MB in zipped format)." from the website https://www.cs.ucr.edu/~eamonn/time_series_data_2018/ ?

I compared data with what I downloaded before. They are the same.

Buy now what I have done is:

1. change the root_dir in main.

2. change this line to UNIVARIATE_DATASET_NAMES = ['PLAID']

3. then run it with python3 main.py InceptionTime

[hfawaz]

There is a folder inside the archive that contains a normalized version of the unormalized datasets.

[Wensi-Tang]

Could you please just give me the download link or some detailed instructions? I cannot find that /(ㄒoㄒ)\

[hfawaz]

Everything should be on this page.

[hadifawaz1999]

@hfawaz