Layer-wise stacked auto-encoder

This is implementation of layer-wise training of a stacked auto-encoder.

Deep Neural Network Based Instrument Extraction From Music

Intro

Environment

Python3 pip install -r requirements.txt

How-to

Initiate the class with

SAE = Stacked_AEC(*save_dir*, *mode*)

mode - is currently either the 'Identity' initilization with 1s across the diagonal or 'LS' for least square initialization.

SAE.fit(train_X=tr_X, train_y=tr_y, test_X=test_X, test_y=test_y, depth=depth, epochs=epoch, lr_rate=rate, batch_size=100, opti='SGD', loss='mse')

Pass 'Adam' for Adam optimizer or pass the whole optimizer yourself:

opt = RMSprop(rate)
SAE.fit(train_X=tr_X, train_y=tr_y, test_X=test_X, test_y=test_y, depth=depth, epochs=epoch, lr_rate=rate, batch_size=100, opti=opt, loss='mse')

Mode

Eye initialization

Initializes the layer weights with 1 on the main diagonal and 0s everywhere else.

[1 0 0, 
0 1 0, 
0 0 1]

Least-squares initialization

Initializes the weights based on the normalized input values (first layer with the training set, following layers with the predictions of the network). Read more here!

kenkyusha / SAE