TadGAN¶

path: orion.primitives.tadgan.TadGAN

description: this is a reconstruction model, namely Generative Adversarial Networks (GAN), containing multiple neural networks and cycle consistency loss. the proposed model is described in the related paper.

see json.

argument	type	description
parameters
`X`	`numpy.ndarray`	n-dimensional array containing the input sequences for the model
`y`	`numpy.ndarray`	n-dimensional array containing the target sequences we want to reconstruct. Typically `y` is a signal from a selected set of channels from `X`.
hyperparameters
`epochs`	`int`	number of epochs to train the model. An epoch is an iteration over the entire X data provided
`input_shape`	`tuple`	tuple denoting the shape of an input sample
`target_shape`	`tuple`	tuple denoting the shape of an reconstructed sample
`optimizer`	`str`	string (name of optimizer) or optimizer instance. Default is `keras.optimizers.Adam`
`learning_rate`	`float`	float denoting the learning rate of the optimizer. Default is 0.005
`latent_dim`	`int`	integer denoting dimension of latent space. Default is 20
`batch_size`	`int`	number of samples per gradient update. Default is 64
`iterations_critic`	`int`	number of critic training steps per generator/encoder training steps. Default is 5
`layers_encoder`	`list`	list containing layers of encoder
`layers_generator`	`list`	list containing layers of generator
`layers_critic_x`	`list`	list containing layers of `critic_x`
`layers_critic_z`	`list`	list containing layers of `critic_z`
output
`y`	`numpy.ndarray`	n-dimensional array containing the reconstructions for each input sequence
`critic`	`numpy.ndarray`	n-dimensional array containing the critic score for each input sequence

In [1]: import numpy as np

In [2]: from mlstars import load_primitive

In [3]: X = np.array([1] * 100).reshape(1, -1, 1)

In [4]: y = X[:,:, [0]] # signal to reconstruct from X (channel 0)

In [5]: primitive = load_primitive('orion.primitives.tadgan.TadGAN',
   ...:     arguments={"X": X, "y":X, "epochs": 5, "batch_size": 1,
   ...:                "iterations_critic": 1})
   ...: 

In [6]: primitive.fit()
Epoch: 1/5, Losses: {'cx_loss': 8.6444, 'cz_loss': 0.6155, 'eg_loss': 9.9115}
Epoch: 2/5, Losses: {'cx_loss': 8.0782, 'cz_loss': 2.2671, 'eg_loss': 9.0332}
Epoch: 3/5, Losses: {'cx_loss': 7.7063, 'cz_loss': 3.6521, 'eg_loss': 6.2166}
Epoch: 4/5, Losses: {'cx_loss': 7.1359, 'cz_loss': 4.2949, 'eg_loss': 5.3597}
Epoch: 5/5, Losses: {'cx_loss': 6.5839, 'cz_loss': 6.6349, 'eg_loss': 5.2059}

In [7]: y, critic = primitive.produce(X=X, y=y)

1/1 [==============================] - ETA: 0s
1/1 [==============================] - 1s 507ms/step

1/1 [==============================] - ETA: 0s
1/1 [==============================] - 1s 605ms/step

1/1 [==============================] - ETA: 0s
1/1 [==============================] - 0s 66ms/step

In [8]: print("average reconstructed value: {:.2f}, critic score {:.2f}".format(
   ...:     y.mean(), critic[0][0]))
   ...: 
average reconstructed value: 0.12, critic score 0.35

Dense AE AER