TSFEL custom features

[mkatgithub123]

I started using tsfel. My main goal is custom feature creation. I could not make this work. I want to use both the built-in features AND custom features in addition / jointly.

I tried:

1) "pip install tsfel"

Creating a custom feature json works with

features_path = os.path.join(folderpath, features_filename)
json_path     = os.path.join(folderpath, json_filename)
tsfel.add_feature_json(features_path, json_path)

and creates the desired, non-empty json:

{
 "temporal": {
  "p2p": {
   "description": "Description of your feature.",
   "parameters": {
    "param": null
   },
   "function": "p2p",
   "use": "yes"
  }
 }
}

This resulted in a custom feature json as intended based on the following custom feature custom_features.py file I provide the feature_path to:

import numpy as np
from tsfel.feature_extraction import set_domain


@set_domain("domain", "temporal")
def p2p(signal, fs):
    """Description of your feature.

    Parameters
    ----------
    signal:
        The time series to calculate the feature of.
    param:
        Parameters of your feature (optional)
    Returns
    -------
    float
        Calculated feature

    """
    # Feature implementation
    p2p = np.max(signal) - np.min(signal)
    return p2p

Calling tsfel by

cfg_file = tsfel.get_features_by_domain() 
X_train  = tsfel.time_series_features_extractor(cfg_file, X_train_sig, fs=50, window_size=250, features_path = features_path)

runs through and then yields all the standard tsfel features, but not my custom p2p feature - I cannot retrieve my custom feature. There is no error message or warning.

2) Downloading the tsfel repository and run it via "pip install --editable ."

I tried this for debugging/to see where things get lost. There were several issues with this approach:

• I use the same function calls as described in 1). To install the package, I had to first add package_dir={"":"tsfel"} to the setup.py file. Otherwise it would not install.

• Calling tsfel.add_feature_json(features_path, json_path) then surprisingly failed upon the complexity computation compute_complexity call in add_personal_features.py, line 102 . Tracing down the issue in compute_complexity, I can see that my custom feature is accessed and computed with the wave function test signal in calculate_complexity.py. There, compute_complexityis executed and in line 124:

calc_window_features(one_feat_dict, wave[: int(ti)], fs, features_path=features_path)

is called. This produces an error in `calc_features.py', line 470 that says "invalid index to scalar variable".

I can confirm that eval_result = locals()[func_total](signal_window, **parameters_total) in line 465 of calc_features.py yields the proper output for my custom feature, a scalar with value "1" (i.e. the peak to peak value for a wave signal with unit amplitude). In line 470 that scalar then is indexed via len(eval_result[ax].shape) and naturally, that scalar value of "1" cannot be indexed. If, instead of just returning the p2p value in my custom feature definition, I return np.array([p2p]), I can circumvent that issue, but it seems unnatural to manipulate the return from my custom feature definition into this. Also, for the built-in features, this does not seem necessary either.

However, from 1) I do have a valid json file that I can use as a workaround for the above error. Still, this behaviour is surprising, at least to me.

• Proceeding with debugging, I find that in calc_features.py, lines 405ff. the custom-feature-related files are executed. However, browsing the domain_feats in line 430,

domain_feats = dict_features[_type].keys()

does not show my custom feature "p2p". In order to have it computed in the following feature computation loop, it should be included in domain_feats, if I understand the code correctly. Between lines 405 and 430, there is no code for integrating any custom features into "domain_feats" which I take as that custom features cannot be computed as of today.

Can you confirm this? If so, why is this not apparent from the docs?