Results python vs opencl

[Carolina710]

Good morning, it is possible to obtain different breaks and means values for the same data (and dates) inputs, when I run the algorithm in python and openCL?

The input data is a simple one with 124 series 17x17. In this case, I executed the algorithm in python, and next, I executed in openCL and obtained different values for breaks and means. I used in both cases the default values for k = 3, freq = 365, trend = False, hfrac = 0.25, level = 0.05 and I just changed the backend value. I must say that for some start_monitor dates the results for the breaks are equal (which is good), but not always. It is possible to obtain different breaks and means values when I change only the backend value, or I'm doing something wrong?

Thank you all, in advance!
Best regards

[mortvest]

Yes, it is currently possible. We are working on fixing the discrepancies between the two backends. Could you provide us the dataset that you used to get the error? A single time series would suffice. Thank you

[Carolina710]

Hello Dmitry,

Thank you for your response.

I share data "subset_1": https://drive.google.com/drive/folders/1MkvA-MSvZCu_8WHkQ4p3R8Z0Qv_Vq-JC?usp=sharing
And data "Tile_1": https://drive.google.com/drive/folders/15DK5T9hZUaaIBNsfzR3tTp5pFb9XCWt5?usp=sharing

The data images are named with the dates, and you can simply collect in this way:

Those are the parameters that I used (I tried also freq=365). I didn't need to crop the data or dates since I used all the subset (124 Images). Nan_value = 0.

I'm obtaining results for the bfast main function not detailed. The one that just returns breaks and means.

Another thing that I would like to inform you about

For the "Tile_1" data, I did the same test above, and in this case, I detected 2 errors in the python version.

The first one is in file "base.py" here:

In line 200 I had to add that line of code, initializing the value for "first_break", because if you don't, and return, the algorithm assumes there is no value for the first_break parameter.

Second, which I don't have a solution:

After updating the first solution, an error occurs while processing row 120, in Python version. This error does not always happen, in this case, it happens when I use start_monitor = datetime(2018, 6, 18), for example.

But running in OpenCL I can obtain results.

Last questions:

I read this article https://arxiv.org/pdf/1807.01751.pdf

But I don't have entirely sure what are the values returned in the means array. Corresponds to the magnitude? My means values are always negative, and in python version is like Max:0 Min:-3.... and OpenCL Max:0 Min:-44....

If you prefer I can open other "issues" for all my other questions.

Obrigada/Best regards,
Carolina Oliveira

[mortvest]

Hello Carolina,

Thank you for the information.

1. Could you please create a gist with the code that you have used?

2. For the "first_break" issue, I am familiar with it and the solution would be added to the develop branch soon

3. What did you mean by "first solution"? Could you also create a gist with the code?

4. For the magnitude issue, there is a couple of problems and I am also working on this.

Could you split your question into multiple issues? That would make it easier for us to keep track of all problems.

Thanks,
Dmitry

[Carolina710]

Hello,

1. Could you please create a gist with the code that you have used?

Here is the gist (more complete): https://gist.github.com/Carolina710/330b473478dd59188edaaa6a3fd76bb6

2. For the "first_break" issue, I am familiar with it and the solution would be added to the develop branch soon

Thank you!

3. What did you mean by "first solution"? Could you also create a gist with the code?

I mean after I add the line self.first_break = -1 I can run the algorithm in Python for Tile_1 and detect the 2nd error that I mention (in Fig 3).

4. For the magnitude issue, there is a couple of problems and I am also working on this.

Thank you!

I will split these issues later as soon as possible :)

[mortvest]

Should be fixed in develop