search

euroargodev / BlueCloud

Working space for BlueCloud demontrator 3 - Lops Task
0 stars 1 forks source link

Differents results for the same fit #3

Open AndreaGarciaJuan opened 4 years ago

AndreaGarciaJuan commented 4 years ago

I fitted the mediterranean data set ds = argo_loader.region([-6, 35, 30, 46, 0, 1000, '2012', '2013']).to_xarray() to a model using k=8 and temperature as feature and I obteined the following spatial distribution of classes:

ArgoMed_map_labels_EX

Some minutes ago I fitted the model again with the same dataset and the same model inputs and I obtained a different result:

ArgoMed_map_labels_EX(2)

The spatial distribution of classes is different: in the fist example we have 2 classes in the western basin and in the second example there is only one class in the western basin.

Why using the same dataset and the same model the results of fitting are different? Is it related to the convengence of the model?

gmaze commented 4 years ago

yes, the classifier is starting from a random position so, depending on the cost function topography, the algorithm may not converge to the same (local) minimum. if you're seeing significant differences like this, this is a strong warning about the (poor) choice of the PCM parameters.