Implementation of dense VAE from paper.

[maffettone]

@prinaldi3 to take the first crack at this.
@maffettone To review and edit.
@lbanko to review final PR.

Potential approach

1. Branch from Black hooks #9
2. Encoder: build_dense_encoder_model(), following lead from

   xca/xca/ml/tf_models.py

   Lines 18 to 30 in dfdad35

   	def build_CNN_model(*,
   	data_shape,
   	filters,
   	kernel_sizes,
   	strides,
   	ReLU_alpha,
   	pool_sizes,
   	batchnorm,
   	n_classes,
   	dense_dims=(),
   	dense_dropout=0.,
   	**kwargs
   	):

   
   Returns: Model(input_x, [z_mean, z_log_var], name="encoder")
3. Decoder: build_dense_decoder_model()
   Returns: Model(z_in, x_dec, name="decoder")
4. VAE class: VAE(tf.keras.Model)
   with methods:

• __init__(encoder_model, decoder_model, kl_loss_weight)
• encode(x) -> mean, log_var,
• reparameterize(mean, log_var) -> z_sample,
• decode(z) -> x_reconstruction,
• kl_loss(z_mean, z_log_var),
• reconstruction_loss(x, x_reconstruction)

Citation

Deep learning for visualization and novelty detection in large X-ray diffraction datasets
Paper in press at npj Computational Materials.
https://arxiv.org/abs/2104.04392