"""
Stefania Fresca, MOX Laboratory, Politecnico di Milano
April 2019
"""
import tensorflow as tf
import numpy as np
from Net import Net
class ROMNet(Net):
def __init__(self, config):
Net.__init__(self, config)
self.n = config['n']
self.n_params = config['n_params']
self.size = 5
self.n_layers = 10
self.n_neurons = 50
self.n_h = config['n_h']
def inference(self):
# encoder function providing the low-dimensional representation of the FOM solution
conv1 = tf.layers.conv2d(inputs = self.input,
filters = 8,
kernel_size = [self.size, self.size],
padding = 'SAME',
strides = 1,
kernel_initializer = tf.keras.initializers.he_uniform(),
activation = tf.nn.elu,
name = 'conv1')
conv2 = tf.layers.conv2d(inputs = conv1,
filters = 16,
kernel_size = [self.size, self.size],
padding = 'SAME',
strides = 2,
kernel_initializer = tf.keras.initializers.he_uniform(),
activation = tf.nn.elu,
name = 'conv2')
conv3 = tf.layers.conv2d(inputs = conv2,
filters = 32,
kernel_size = [self.size, self.size],
padding = 'SAME',
strides = 2,
kernel_initializer = tf.keras.initializers.he_uniform(),
activation = tf.nn.elu,
name = 'conv3')
conv4 = tf.layers.conv2d(inputs = conv3,
filters = 64,
kernel_size = [self.size, self.size],
padding = 'SAME',
strides = 2,
kernel_initializer = tf.keras.initializers.he_uniform(),
activation = tf.nn.elu,
name = 'conv4')
feature_dim_enc = conv4.shape[1] * conv4.shape[2] * conv4.shape[3]
conv4 = tf.reshape(conv4, [-1, feature_dim_enc])
fc1 = tf.layers.dense(conv4, 256, activation = tf.nn.elu, kernel_initializer = tf.keras.initializers.he_uniform(), name = 'fc1')
self.enc = tf.layers.dense(fc1, self.n, activation = tf.nn.elu, kernel_initializer = tf.keras.initializers.he_uniform(), name = 'fc2')
# feed-forward neural network for reduced dynamics learning
fc_n = tf.layers.dense(self.params,
self.n_neurons,
activation = tf.nn.elu,
kernel_initializer = tf.keras.initializers.he_uniform())
for i in range(self.n_layers):
fc_n = tf.layers.dense(fc_n,
self.n_neurons,
activation = tf.nn.elu,
kernel_initializer = tf.keras.initializers.he_uniform())
self.u_n = tf.layers.dense(fc_n,
self.n,
activation = tf.nn.elu,
kernel_initializer = tf.keras.initializers.he_uniform())
# decoder function for reduced nonlinear trial manifold learning
fc1_t = tf.layers.dense(self.u_n, 256, activation = tf.nn.elu, kernel_initializer = tf.keras.initializers.he_uniform(), name = 'fc1_t')
fc2_t = tf.layers.dense(fc1_t, self.N_h, activation = tf.nn.elu, kernel_initializer = tf.keras.initializers.he_uniform(), name = 'fc2_t')
fc2_t = tf.reshape(fc2_t, [-1, self.n_h, self.n_h, 64])
conv1_t = tf.layers.conv2d_transpose(inputs = fc2_t,
filters = 64,
kernel_size = [self.size, self.size],
padding = 'SAME',
strides = 2,
kernel_initializer = tf.keras.initializers.he_uniform(),
activation = tf.nn.elu,
name = 'conv1_t')
conv2_t = tf.layers.conv2d_transpose(inputs = conv1_t,
filters = 32,
kernel_size = [self.size, self.size],
padding = 'SAME',
strides = 2,
kernel_initializer = tf.keras.initializers.he_uniform(),
activation = tf.nn.elu,
name = 'conv2_t')
conv3_t = tf.layers.conv2d_transpose(inputs = conv2_t,
filters = 16,
kernel_size = [self.size, self.size],
padding = 'SAME',
strides = 2,
kernel_initializer = tf.keras.initializers.he_uniform(),
activation = tf.nn.elu,
name = 'conv3_t')
conv4_t = tf.layers.conv2d_transpose(inputs = conv3_t,
filters = 1,
kernel_size = [self.size, self.size],
padding = 'SAME',
strides = 1,
kernel_initializer = tf.keras.initializers.he_uniform(),
name = 'conv4_t')
feature_dim_dec = conv4_t.shape[1] * conv4_t.shape[2] * conv4_t.shape[3]
self.u_h = tf.reshape(conv4_t, [-1, feature_dim_dec])
Datasets
Models
