Burgers_Experiment30e.py

# An experiment to try things out
import copy
import random as r

import numpy as np
import training
import tensorflow as tf

params = {}
params['data_name'] = 'Burgers_Eqn_exp30'  ## FILL IN HERE (from file name)
params['folder_name'] = 'Burgers_exp30e'  # UPDATE so goes in own folder
params['restore'] = 0 # Restore a previous model

n = 128  # number of inputs (spatial discretization)
params['delta_t'] = 0.002  # Time step
params['len_time'] = 51  # Number of time steps in each trajectory
params['mu'] = 1.0  # Strength of viscous (diffusion) term in Burgers' - only needed if seeding middle layer
numICs = 6000  # Number of initial conditions
params['data_train_len'] = 20  # Number of training data sets

params['num_evals'] = 128  # how many eigenvalues / how many frequencies / what's the low dimension
l = params['num_evals']

params['network_arch'] = 'fully_connected' # Choose 'convnet' or 'fully_connected'

# If fully-connected layers:
params['act_type'] = tf.nn.relu
params['widths'] = [n, n, n, n, n, n, l, l, n, n, n, n, n, n]
params['linear_encoder_layers'] = [4, 5]  # 0 is relu, 1&2 are linear
params['linear_decoder_layers'] = [0, 5]  # 0 linear, 1 relu, 2 linear
params['log_space'] = 0  # 1 if you want to take a logarithm at beginning of encoder/decoder

# If convolutional neural net
params['n_inputs'] = 128
params['conv1_filters'] = 32 # Number of filters in encoder convolutional layer
params['n_middle'] = l
params['conv2_filters'] = 16 # Number of filters in decoder convolutional layer
params['n_outputs'] = 128

# For either type of architecture
params['seed_middle'] = 0  # Seed middle three layers with heat equation (Fourier transform, diagonal matrix, inverse FT)
params['fix_middle'] = 0   # Can only fix middle layers if you also seed middle
params['add_identity'] = 1
# He initialization = tf.contrib.layers.variance_scaling_initializer(), Identity initialization = identity_initializer()
params['initialization'] = 'He' # Choose 'He' or 'identity'

params['relative_loss'] = 1
params['auto_first'] = 1 # Train autoencoder only for first 5 minutes
params['diag_L'] = 1  # Middle Layer (dynamics) forced to be diagonal 

params['num_shifts'] = params['len_time']-1  # For prediction loss
params['num_shifts_middle'] = params['len_time']-1  # For linearity loss

params['max_time'] =  1 * 20 * 60  # hours * minutes * seconds
params['num_passes_per_file'] = 15 * 6 * 10 * 50  # may be limiting factor
params['num_steps_per_batch'] = 2

# Stop if loss is greater than given number
params['min_5min'] = 10**3 
params['min_20min'] = 10**2
params['min_40min'] = 10
params['min_1hr'] = 10
params['min_2hr'] = 10
params['min_3hr'] = 10
params['min_4hr'] = 10
params['min_halfway'] = 10  

# Strength of loss terms
params['autoencoder_loss_lam'] = 1.0
params['prediction_loss_lam'] = 1.0
params['linearity_loss_lam'] = 1.0
params['inner_autoencoder_loss_lam'] = 1.0
params['outer_autoencoder_loss_lam'] = 1.0

# Regularization
params['L1_lam'] = 0.0
params['L2_lam'] = 10**(-8)

max_shifts = max(params['num_shifts'], params['num_shifts_middle'])
num_examples = numICs * (params['len_time'] - max_shifts)
params['batch_size'] = min(num_examples, 64)
steps_to_see_all = num_examples / params['batch_size']
params['num_steps_per_file_pass'] = (int(steps_to_see_all) + 1) * params['num_steps_per_batch']

for count in range(20):
    # only randomized part is learning_rate
    params['rand_seed'] = r.randint(0,10**(10))
    r.seed(params['rand_seed'])
    params['learning_rate'] = 10**(-r.uniform(3,6))
    print(params['learning_rate'])
    
    training.main_exp(copy.deepcopy(params))