upload four experiments from v2 of paper

DiscreteSpectrumExampleExperiment.py

@@ -1,49 +1,101 @@
-import copy
-import random as r
-
-import training
-
-params = {}
-
-# settings related to dataset
-params['data_name'] = 'DiscreteSpectrumExample'
-params['len_time'] = 51
-n = 2  # dimension of system (and input layer)
-params['data_train_len'] = 10
-num_initial_conditions = 29400  # per training file (10 training data files)
-params['delta_t'] = 0.02
-
-# settings related to saving results
-params['folder_name'] = 'exp1'
-
-# settings related to network architecture
-w = 100  # width of hidden layers in encoder and decoder
-k = 2  # dimension of y-coordinates 
-params['widths'] = [n, w, w, w, k, k, w, w, w, n]
-params['widths_omega'] = [2, w, w, w, 1]
-
-# settings related to loss function
-params['num_shifts'] = 3
-params['Linf_lam'] = 10 ** (-6)
-
-# settings related to the training
-params['num_passes_per_file'] = 15 * 4
-params['num_steps_per_batch'] = 2
-params['learning_rate'] = 10 ** (-3)
-
-# settings related to the timing
-params['max_time'] = 4 * 60 * 60  # 4 hours
-params['min_halfway'] = 10 ** (-5)
-
-for count in range(200):  # loop to do random experiments
-    params['num_shifts_middle'] = r.randint(5, params['len_time'] - 1)
-    max_shifts = max(params['num_shifts'], params['num_shifts_middle'])
-    num_examples = num_initial_conditions * (params['len_time'] - max_shifts)
-    params['batch_size'] = int(2 ** (r.uniform(6, 9)))
-    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']
-
-    params['L2_lam'] = 10 ** (-r.uniform(12, 18))
-    params['L1_lam'] = 10 ** (-r.uniform(14, 18))
-
-    training.main_exp(copy.deepcopy(params))
+import copy
+import random as r
+
+import numpy as np
+
+import training
+
+params = {}
+
+# settings related to dataset
+params['data_name'] = 'DiscreteSpectrumExample'
+params['len_time'] = 51
+n = 2  # dimension of system (and input layer)
+num_initial_conditions = 5000  # per training file
+params['delta_t'] = 0.02
+
+# settings related to saving results
+params['folder_name'] = 'exp1'
+
+# settings related to network architecture
+params['num_real'] = 2
+params['num_complex_pairs'] = 0
+params['num_evals'] = 2
+k = params['num_evals']  # dimension of y-coordinates
+
+# settings related to loss function
+params['num_shifts'] = 30
+params['num_shifts_middle'] = params['len_time'] - 1
+max_shifts = max(params['num_shifts'], params['num_shifts_middle'])
+num_examples = num_initial_conditions * (params['len_time'] - max_shifts)
+params['recon_lam'] = .1
+params['L1_lam'] = 0.0
+
+# settings related to the training
+params['num_passes_per_file'] = 15 * 6 * 10
+params['num_steps_per_batch'] = 2
+params['learning_rate'] = 10 ** (-3)
+
+# settings related to the timing
+params['max_time'] = 4 * 60 * 60  # 4 hours
+params['min_5min'] = .5
+params['min_20min'] = .0004
+params['min_40min'] = .00008
+params['min_1hr'] = .00003
+params['min_2hr'] = .00001
+params['min_3hr'] = .000006
+params['min_halfway'] = .000006
+
+for count in range(200):  # loop to do random experiments
+    params['data_train_len'] = r.randint(1, 3)
+    params['batch_size'] = int(2 ** (r.randint(7, 9)))
+    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']
+    if r.random() < .5:
+        params['auto_first'] = 1
+    else:
+        params['auto_first'] = 0
+
+    params['L2_lam'] = 10 ** (-r.randint(13, 15))
+    if r.random() < .5:
+        params['Linf_lam'] = 0.0
+    else:
+        params['Linf_lam'] = 10 ** (-r.randint(6, 10))
+
+    d = r.randint(1, 4)
+    if d == 1:
+        wopts = np.arange(50, 160, 10)
+        w = wopts[r.randint(0, len(wopts) - 1)]
+        params['widths'] = [2, w, k, k, w, 2]
+    elif d == 2:
+        wopts = np.arange(15, 45, 5)
+        w = wopts[r.randint(0, len(wopts) - 1)]
+        params['widths'] = [2, w, w, k, k, w, w, 2]
+    elif d == 3:
+        wopts = np.arange(10, 25, 5)
+        w = wopts[r.randint(0, len(wopts) - 1)]
+        params['widths'] = [2, w, w, w, k, k, w, w, w, 2]
+    elif d == 4:
+        wopts = np.arange(10, 20, 5)
+        w = wopts[r.randint(0, len(wopts) - 1)]
+        params['widths'] = [2, w, w, w, w, k, k, w, w, w, w, 2]
+
+    do = r.randint(1, 4)
+    if do == 1:
+        wopts = np.arange(20, 110, 10)
+        wo = wopts[r.randint(0, len(wopts) - 1)]
+        params['hidden_widths_omega'] = [wo, ]
+    elif do == 2:
+        wopts = np.arange(10, 25, 5)
+        wo = wopts[r.randint(0, len(wopts) - 1)]
+        params['hidden_widths_omega'] = [wo, wo]
+    elif do == 3:
+        wopts = np.arange(5, 20, 5)
+        wo = wopts[r.randint(0, len(wopts) - 1)]
+        params['hidden_widths_omega'] = [wo, wo, wo]
+    elif do == 4:
+        wopts = np.arange(5, 15, 5)
+        wo = wopts[r.randint(0, len(wopts) - 1)]
+        params['hidden_widths_omega'] = [wo, wo, wo, wo]
+
+    training.main_exp(copy.deepcopy(params))

FluidFlowBoxExperiment.py

@@ -0,0 +1,78 @@
+import copy
+import random as r
+
+import numpy as np
+
+import training
+
+params = {}
+
+# settings related to dataset
+params['data_name'] = 'FluidFlowBox'
+params['len_time'] = 101
+n = 3  # dimension of system (and input layer)
+num_initial_conditions = 5000  # per training file
+params['delta_t'] = 0.01
+
+# settings related to saving results
+params['folder_name'] = 'exp4'
+
+# settings related to network architecture
+params['num_real'] = 1
+params['num_complex_pairs'] = 1
+params['num_evals'] = 3
+k = params['num_evals']  # dimension of y-coordinates
+
+# defaults related to initialization of parameters
+params['dist_weights'] = 'dl'
+params['dist_weights_omega'] = 'dl'
+
+# settings related to loss function
+params['num_shifts'] = 30
+params['num_shifts_middle'] = params['len_time'] - 1
+max_shifts = max(params['num_shifts'], params['num_shifts_middle'])
+num_examples = num_initial_conditions * (params['len_time'] - max_shifts)
+params['recon_lam'] = .1
+params['L1_lam'] = 0.0
+params['auto_first'] = 1
+
+# settings related to training
+params['num_passes_per_file'] = 15 * 6 * 10
+params['num_steps_per_batch'] = 2
+params['learning_rate'] = 10 ** (-3)
+
+# settings related to timing
+params['max_time'] = 6 * 60 * 60  # 6 hours
+params['min_5min'] = .45
+params['min_20min'] = .005
+params['min_40min'] = .0005
+params['min_1hr'] = .00025
+params['min_2hr'] = .00005
+params['min_3hr'] = .000007
+params['min_4hr'] = .000005
+params['min_halfway'] = 1
+
+for count in range(200):  # loop to do random experiments
+    params['data_train_len'] = r.randint(1, 4)
+    params['batch_size'] = int(2 ** (r.randint(7, 8)))
+    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']
+    params['L2_lam'] = 10 ** (-r.randint(13, 14))
+    params['Linf_lam'] = 10 ** (-r.randint(7, 9))
+
+    d = 1
+    wopts = np.arange(50, 200, 5)
+    w = wopts[r.randint(0, len(wopts) - 1)]
+    params['widths'] = [n, w, k, k, w, n]
+
+    do = r.randint(1, 2)
+    if do == 1:
+        wopts = np.arange(20, 135, 5)
+        wo = wopts[r.randint(0, len(wopts) - 1)]
+        params['hidden_widths_omega'] = [wo, ]
+    elif do == 2:
+        wopts = np.arange(10, 30, 5)
+        wo = wopts[r.randint(0, len(wopts) - 1)]
+        params['hidden_widths_omega'] = [wo, wo]
+
+    training.main_exp(copy.deepcopy(params))

FluidFlowOnAttractorExperiment.py

@@ -0,0 +1,78 @@
+import copy
+import random as r
+
+import numpy as np
+
+import training
+
+params = {}
+
+# settings related to dataset
+params['data_name'] = 'FluidFlowOnAttractor'
+params['len_time'] = 121
+n = 3  # dimension of system (and input layer)
+num_initial_conditions = 5000  # per training file
+params['delta_t'] = 0.05
+
+# settings related to saving results
+params['folder_name'] = 'exp3'
+
+# settings related to network architecture
+params['num_real'] = 0
+params['num_complex_pairs'] = 1
+params['num_evals'] = 2
+k = params['num_evals']  # dimension of y-coordinates
+
+# settings related to loss function
+params['num_shifts'] = 30
+params['num_shifts_middle'] = params['len_time'] - 1
+max_shifts = max(params['num_shifts'], params['num_shifts_middle'])
+num_examples = num_initial_conditions * (params['len_time'] - max_shifts)
+params['recon_lam'] = .1
+params['L1_lam'] = 0.0
+params['auto_first'] = 1
+
+# settings related to training
+params['num_passes_per_file'] = 15 * 6 * 10
+params['num_steps_per_batch'] = 2
+params['learning_rate'] = 10 ** (-3)
+
+# settings related to timing
+params['max_time'] = 6 * 60 * 60  # 6 hours
+params['min_5min'] = .45
+params['min_20min'] = .001
+params['min_40min'] = .0005
+params['min_1hr'] = .00025
+params['min_2hr'] = .00005
+params['min_3hr'] = .000005
+params['min_4hr'] = .0000007
+params['min_halfway'] = 1
+
+for count in range(200):  # loop to do random experiments
+    params['data_train_len'] = r.randint(1, 3)
+    params['batch_size'] = int(2 ** (r.randint(7, 8)))
+    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']
+    params['L2_lam'] = 10 ** (-r.randint(13, 14))
+    params['Linf_lam'] = 10 ** (-r.randint(7, 10))
+
+    d = r.randint(1, 2)
+    if d == 1:
+        wopts = np.arange(70, 135, 5)
+        w = wopts[r.randint(0, len(wopts) - 1)]
+        params['widths'] = [n, w, k, k, w, n]
+    elif d == 2:
+        wopts = np.arange(15, 30, 5)
+        w = wopts[r.randint(0, len(wopts) - 1)]
+        params['widths'] = [n, w, w, k, k, w, w, n]
+    do = r.randint(1, 2)
+    if do == 1:
+        wopts = np.arange(230, 450, 10)
+        wo = wopts[r.randint(0, len(wopts) - 1)]
+        params['hidden_widths_omega'] = [wo, ]
+    elif do == 2:
+        wopts = np.arange(25, 40, 5)
+        wo = wopts[r.randint(0, len(wopts) - 1)]
+        params['hidden_widths_omega'] = [wo, wo]
+
+    training.main_exp(copy.deepcopy(params))