Datasets
Models
Downloads: 1
Diff of
/ROMNet.py
[000000]
..
[6d7935]
Switch to unified view
| a | b/ROMNet.py | ||
|---|---|---|---|
| 1 | """ |
||
| 2 | |||
| 3 | Stefania Fresca, MOX Laboratory, Politecnico di Milano |
||
| 4 | April 2019 |
||
| 5 | |||
| 6 | """ |
||
| 7 | |||
| 8 | import tensorflow as tf |
||
| 9 | import numpy as np |
||
| 10 | |||
| 11 | from Net import Net |
||
| 12 | |||
| 13 | class ROMNet(Net): |
||
| 14 | def __init__(self, config): |
||
| 15 | Net.__init__(self, config) |
||
| 16 | |||
| 17 | self.n = config['n'] |
||
| 18 | self.n_params = config['n_params'] |
||
| 19 | |||
| 20 | self.size = 5 |
||
| 21 | self.n_layers = 10 |
||
| 22 | self.n_neurons = 50 |
||
| 23 | self.n_h = config['n_h'] |
||
| 24 | |||
| 25 | def inference(self): |
||
| 26 | # encoder function providing the low-dimensional representation of the FOM solution |
||
| 27 | conv1 = tf.layers.conv2d(inputs = self.input, |
||
| 28 | filters = 8, |
||
| 29 | kernel_size = [self.size, self.size], |
||
| 30 | padding = 'SAME', |
||
| 31 | strides = 1, |
||
| 32 | kernel_initializer = tf.keras.initializers.he_uniform(), |
||
| 33 | activation = tf.nn.elu, |
||
| 34 | name = 'conv1') |
||
| 35 | conv2 = tf.layers.conv2d(inputs = conv1, |
||
| 36 | filters = 16, |
||
| 37 | kernel_size = [self.size, self.size], |
||
| 38 | padding = 'SAME', |
||
| 39 | strides = 2, |
||
| 40 | kernel_initializer = tf.keras.initializers.he_uniform(), |
||
| 41 | activation = tf.nn.elu, |
||
| 42 | name = 'conv2') |
||
| 43 | conv3 = tf.layers.conv2d(inputs = conv2, |
||
| 44 | filters = 32, |
||
| 45 | kernel_size = [self.size, self.size], |
||
| 46 | padding = 'SAME', |
||
| 47 | strides = 2, |
||
| 48 | kernel_initializer = tf.keras.initializers.he_uniform(), |
||
| 49 | activation = tf.nn.elu, |
||
| 50 | name = 'conv3') |
||
| 51 | conv4 = tf.layers.conv2d(inputs = conv3, |
||
| 52 | filters = 64, |
||
| 53 | kernel_size = [self.size, self.size], |
||
| 54 | padding = 'SAME', |
||
| 55 | strides = 2, |
||
| 56 | kernel_initializer = tf.keras.initializers.he_uniform(), |
||
| 57 | activation = tf.nn.elu, |
||
| 58 | name = 'conv4') |
||
| 59 | feature_dim_enc = conv4.shape[1] * conv4.shape[2] * conv4.shape[3] |
||
| 60 | conv4 = tf.reshape(conv4, [-1, feature_dim_enc]) |
||
| 61 | fc1 = tf.layers.dense(conv4, 256, activation = tf.nn.elu, kernel_initializer = tf.keras.initializers.he_uniform(), name = 'fc1') |
||
| 62 | self.enc = tf.layers.dense(fc1, self.n, activation = tf.nn.elu, kernel_initializer = tf.keras.initializers.he_uniform(), name = 'fc2') |
||
| 63 | # feed-forward neural network for reduced dynamics learning |
||
| 64 | fc_n = tf.layers.dense(self.params, |
||
| 65 | self.n_neurons, |
||
| 66 | activation = tf.nn.elu, |
||
| 67 | kernel_initializer = tf.keras.initializers.he_uniform()) |
||
| 68 | for i in range(self.n_layers): |
||
| 69 | fc_n = tf.layers.dense(fc_n, |
||
| 70 | self.n_neurons, |
||
| 71 | activation = tf.nn.elu, |
||
| 72 | kernel_initializer = tf.keras.initializers.he_uniform()) |
||
| 73 | self.u_n = tf.layers.dense(fc_n, |
||
| 74 | self.n, |
||
| 75 | activation = tf.nn.elu, |
||
| 76 | kernel_initializer = tf.keras.initializers.he_uniform()) |
||
| 77 | # decoder function for reduced nonlinear trial manifold learning |
||
| 78 | fc1_t = tf.layers.dense(self.u_n, 256, activation = tf.nn.elu, kernel_initializer = tf.keras.initializers.he_uniform(), name = 'fc1_t') |
||
| 79 | fc2_t = tf.layers.dense(fc1_t, self.N_h, activation = tf.nn.elu, kernel_initializer = tf.keras.initializers.he_uniform(), name = 'fc2_t') |
||
| 80 | fc2_t = tf.reshape(fc2_t, [-1, self.n_h, self.n_h, 64]) |
||
| 81 | conv1_t = tf.layers.conv2d_transpose(inputs = fc2_t, |
||
| 82 | filters = 64, |
||
| 83 | kernel_size = [self.size, self.size], |
||
| 84 | padding = 'SAME', |
||
| 85 | strides = 2, |
||
| 86 | kernel_initializer = tf.keras.initializers.he_uniform(), |
||
| 87 | activation = tf.nn.elu, |
||
| 88 | name = 'conv1_t') |
||
| 89 | conv2_t = tf.layers.conv2d_transpose(inputs = conv1_t, |
||
| 90 | filters = 32, |
||
| 91 | kernel_size = [self.size, self.size], |
||
| 92 | padding = 'SAME', |
||
| 93 | strides = 2, |
||
| 94 | kernel_initializer = tf.keras.initializers.he_uniform(), |
||
| 95 | activation = tf.nn.elu, |
||
| 96 | name = 'conv2_t') |
||
| 97 | conv3_t = tf.layers.conv2d_transpose(inputs = conv2_t, |
||
| 98 | filters = 16, |
||
| 99 | kernel_size = [self.size, self.size], |
||
| 100 | padding = 'SAME', |
||
| 101 | strides = 2, |
||
| 102 | kernel_initializer = tf.keras.initializers.he_uniform(), |
||
| 103 | activation = tf.nn.elu, |
||
| 104 | name = 'conv3_t') |
||
| 105 | conv4_t = tf.layers.conv2d_transpose(inputs = conv3_t, |
||
| 106 | filters = 1, |
||
| 107 | kernel_size = [self.size, self.size], |
||
| 108 | padding = 'SAME', |
||
| 109 | strides = 1, |
||
| 110 | kernel_initializer = tf.keras.initializers.he_uniform(), |
||
| 111 | name = 'conv4_t') |
||
| 112 | feature_dim_dec = conv4_t.shape[1] * conv4_t.shape[2] * conv4_t.shape[3] |
||
| 113 | self.u_h = tf.reshape(conv4_t, [-1, feature_dim_dec]) |