Extract new features from all the layer-wise activations of a trained autoencoder

extract_autoencoder_activations.py

import lightgbm as lgb

# compute list of activations for each hidden dense layer
layer_outputs = [layer.output for layer in autoencoder.layers if 'dense' in layer.name]
activation_model = tf.keras.models.Model(inputs=autoencoder.input, outputs=layer_outputs)

# compute training and validation activations
train_activations = activation_model.predict(X_train_standard)
val_activations = activation_model.predict(X_val_standard)

# concatenate activations into one big Dataframe
X_train_autoencoder_activations = pd.DataFrame(np.concatenate(train_activations, axis=1))
X_val_autoencoder_activations = pd.DataFrame(np.concatenate(val_activations, axis=1))

# standardize
scaler = StandardScaler()
X_train_autoencoder_activations = scaler.fit_transform(X_train_autoencoder_activations)
X_val_autoencoder_activations = scaler.transform((X_val_autoencoder_activations))

# convert to DMatrix for XGBoost
dtrain = lgb.Dataset(X_train_autoencoder_activations, y_train, free_raw_data=False).construct()
dval = lgb.Dataset(X_val_autoencoder_activations, y_val, free_raw_data=False).construct()