Google Colab: TypeError in bsoid_assign() (Too many bodyparts?)

[JensBlack]

Great work with the Google Colab notebook. As a python user, I was looking forward to this! Finally have the time to try it out!
Sadly:
Running your Google Colab notebook with original DeepLabCut data yields the following error in line
f_10fps,tsne_feats,labels,tsne_fig = bsoid_assign(data,fps = 30,comp = 1,kclass = 50,it = 30)

As documented, i only changed fps (from 60 to 30).

I am using an animal model with 9 Points, single animal (top-down view) the points are called ['nose', 'neck', 'left_shoulder', 'right_shoulder', 'left_hip', 'centroid', 'right_shoulder', 'tail_root', 'tail_tip'].
I was not able to see if you are using hardcoded bodyparts for the feature extraction. If so, this might be the issue?

More Details
I run the code completely on google colab with google drive connected. The deeplabcut csv files are raw but quite big (> 50k rows each).

If you need any further info, I am happy to provide it.

[runninghsus]

Hi,

Thanks for your issue. You guessed correctly, it was because we hard-coded in 6 body-parts (we've done HCA based on all kinds of features using those 6 body parts so we are only confident in 6 at the moment), so 9 will cause an issue.

Implementation wise: For yours, if you could just use ['nose','left shoulder','right shoulder','left hip','right hip', 'tail root'], it should work. In other words, by removing the columns before b-soid_assign, it should run as expected. The number of rows and data file size should not be of concern.

Your data specifically: Obviously, 9 data points will provide more detailed information for b-soid to parse, and likely resulting in finer behavioral differences (though in my computation of the 6 points, I do estimate points to get around that). We can talk more about incorporating "neck" or "centroid" if you like. By having these additional 3 points, you can analyze a barrage of kinematics that 6 points are lacking, but does not necessarily add weight to extracting new behaviors.

Note that the python notebook is in beta and I have only tested it on my open-field data so far. I am currently still testing it on other experiments.

Please keep me posted on any new issues using Google Colab! I am eager to fully implement the python version of this algorithm (pip install bsoid will be great for python users as opposed to running it on the cloud!).

Alex

[JensBlack]

Hi Alex,

thanks for the quick reply. I will try out slicing the dataframe to 6 bodyparts!
Currently I am running open field data (from my DeepLabStream experiments) and would be happy to share any results/issues I will find on my way.

[JensBlack]

Changing the number of bodyparts to 6 did not solve the issue.

Screenshot

Followup question

is the number of csv files that are imported important?

I imported 2 csv files (6 bodyparts, no multiindex etc.; see the print(data[0].shape, data[1].shape) statement in the screenshot.

[runninghsus]

Hi,

I've fixed it. It was a panda version difference that they recently updated on Colab (1.03).
The new copy of the colab notebook is located here
What I changed (in the boxcar_center function):
.
.
moving_avg = np.array(a1.rolling(window = n,min_periods=1,center = True).mean())
.

Let me know if this solves this issue.

Alex

[JensBlack]

This solves the issue. It continues to run smoothly.
Thank you for the quick resolution!