Has SUSI quantization or topographic error? about susi HOT 13 CLOSED

Just want to compare with other SOM results

from susi.

Hi @akol67 ,

Are we talking about unsupervised learning or supervised? If supervised, then classification?

After quick research, it seems that a quantization error can be implemented very easily. Please give me more information on how you define it, for what you would like to use et etc.

from susi.

from susi.

Thanks, I will implement it and will keep you updated!

from susi.

@akol67 Please update your SuSi package:

pip uninstall susi
git clone https://github.com/felixriese/susi.git
cd susi/
git checkout quanti-error
python setup.py install

And then please test the new function som.som.get_quantization_error(). If it works fine, I will merge it to main.

from susi.

from susi.

from susi.

About your error: Thanks for testing. I don't see the image, though. Please paste your error message here. What can be missing is som.fit(X) before you calculate any error.

About the topographic error: Feel free to fork the repository, implement the topographic error, and create a pull request. That would be the easiest solution.

About how to train without error metric: There are several aspects to consider.

1. The way the learning rate etc. is set up, the number of iterations does not need to be optimized that much. Therefore, I had no need for error metrics so far.
2. The package is mostly used in(semi-) supervised learning where you get classification/regression quality metrics. They are also very helpful for the training.

from susi.

Hi Felix,

Fit was already done before. Seems to me there is something wrong when using semisupervised and supervised.
It works fine if we use unlabeled_size = 0.
Below is with unlabeled_size not equal zero.

unlabeled_size = 0.8
rng = np.random.RandomState(42)
random_unlabeled_points = rng.rand(len(y_train)) < unlabeled_size
y_train[random_unlabeled_points] = -1

print("Datapoints for training with label (supervised):\t", y_train.shape[0])
print("Datapoints for training with label (semi-supervised):\t", np.sum(y_train != -1))
print("Datapoints for testing:\t\t\t\t\t", y_test.shape[0])

>Datapoints for training with label (supervised):	 6012
>Datapoints for training with label (semi-supervised):	 1203
>Datapoints for testing:					 2577


som_superv = susi.SOMClassifier(
    n_rows= siz1,
    n_columns= siz2,
    n_iter_unsupervised= n_epocas,
    n_iter_supervised= n_epocas,
    missing_label_placeholder=-1,
    random_state=42,
    verbose=1,
    n_jobs=1)

som_superv.fit(data_train, y_train)
print("Training:"+str(som_superv.score(data_train,y_train)))
print("Test:"+str(som_superv.score(data_test,y_test)))
y_pred = som_superv.predict(data_test)
y_pred_proba = som_superv.predict_proba(data_test)

>Training:0.18196939454424485
>Test:0.9064804035700427

som_superv.get_quantization_error()

>---------------------------------------------------------------------------
>ValueError                                Traceback (most recent call last)
<ipython-input-36-58ee03585184> in <module>
>----> 1 som_superv.get_quantization_error()

>~/anaconda3/lib/python3.8/site-packages/susi-1.2-py3.8.egg/susi/SOMClustering.py in get_quantization_error(self)
>    954 
>    955         quantization_errors = np.linalg.norm(
--> 956             np.subtract(weights_per_datapoint, self.X_)
>    957         )
>    958 

>ValueError: operands could not be broadcast together with shapes (1203,4) (6012,4) 

from susi.

Thanks!

I fixed a bug, see #26. Please do the following before you run your code again within the susi folder:

pip uninstall susi
git pull
python setup.py install

from susi.

Hi @akol67, Did you check the latest update? Does it work for you?

from susi.

from susi.

Thanks, I added it to the list!

from susi.