Customer Transformers Not Working

[habeeb3579]

TPOT does not work with custom transformers. When I ran to optimization below, I got the warning that "Warning: Normalization2 is not available and will not be used by TPOT."

class Normalization2(BaseEstimator, TransformerMixin):
def init(self, method='minmax', feature_range=(0, 1)):
self.method = method
self.feature_range = feature_range
self.min_ = None
self.max_ = None

def fit(self, spectrum):
    if self.method == 'minmax':
        self.min_ = np.min(spectrum)
        self.max_ = np.max(spectrum)
    elif self.method == 'zscore':
        self.mean = np.mean(spectrum)
        self.std = np.std(spectrum)
    else:
        raise ValueError("Unsupported normalization method. Choose 'minmax' or 'zscore'.")
    return self

def transform(self, spectrum):
    if self.method == 'minmax':
        if self.min_ is None or self.max_ is None:
            raise ValueError("Normalization instance is not fitted yet.")
        range_min, range_max = self.feature_range
        return ((spectrum - self.min_) / (self.max_ - self.min_)) * (range_max - range_min) + range_min
    elif self.method == 'zscore':
        if self.mean is None or self.std is None:
            raise ValueError("Normalization instance is not fitted yet.")
        return (spectrum - self.mean) / self.std
        
 config = {
'Normalization2': {  # Wrapper for Normalization
    'method': ['minmax'],
    'feature_range': [(0, 1)],
},
'sklearn.linear_model.Ridge': {  # Replace PLSRegression with Ridge
    'alpha': [0.1, 1.0, 10.0],  # Regularization strength
    'solver': ['auto', 'svd', 'cholesky', 'lsqr', 'sag', 'saga'],  # Solvers to optimize
},

}

spot = TPOTRegressor(
generations=100,
population_size=100,
cv=3
config_dict=config,
verbosity=2,
n_jobs=-1,
max_eval_time_mins=10
)

[perib]

TPOT can only import from modules. To get this to work you first have to put your custom class in a module. For example

mymodule (folder):
    __init__.py
    myclassfile.py
myscript.py

where myclassfile.py contains your custom class. __init__.py imports the file. Then you change the config dictionary to reference this class.

For example

myclassfile.py

from sklearn.base import BaseEstimator, TransformerMixin
from tpot import TPOTRegressor

class Normalization2(BaseEstimator, TransformerMixin):
    def init(self, method='minmax', feature_range=(0, 1)):
        self.method = method
        self.feature_range = feature_range
        self.min_ = None
        self.max_ = None

    def fit(self, spectrum):
        if self.method == 'minmax':
            self.min_ = np.min(spectrum)
            self.max_ = np.max(spectrum)
        elif self.method == 'zscore':
            self.mean = np.mean(spectrum)
            self.std = np.std(spectrum)
        else:
            raise ValueError("Unsupported normalization method. Choose 'minmax' or 'zscore'.")
        return self

    def transform(self, spectrum):
        if self.method == 'minmax':
            if self.min_ is None or self.max_ is None:
                raise ValueError("Normalization instance is not fitted yet.")
            range_min, range_max = self.feature_range
            return ((spectrum - self.min_) / (self.max_ - self.min_)) * (range_max - range_min) + range_min
        elif self.method == 'zscore':
            if self.mean is None or self.std is None:
                raise ValueError("Normalization instance is not fitted yet.")
            return (spectrum - self.mean) / self.std
        

init.py

from .myclassfile import Normalization2

myscript.py

...
config = {
'mymodule.Normalization2': {  # Wrapper for Normalization
    'method': ['minmax'],
    'feature_range': [(0, 1)],
},
'sklearn.linear_model.Ridge': {  # Replace PLSRegression with Ridge
    'alpha': [0.1, 1.0, 10.0],  # Regularization strength
    'solver': ['auto', 'svd', 'cholesky', 'lsqr', 'sag', 'saga'],  # Solvers to optimize
},

}
...

(This is not necessary in TPOT2 which uses a different API)