-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathrun_exp4.py
125 lines (100 loc) · 6.13 KB
/
run_exp4.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
import os
import argparse
import numpy as np
import pickle as pkl
from model import ODENetwork
from data_duffing import get_dataset
from utils import leapfrog_solver, runge_kutta_solver, reshape_data
import matplotlib.pyplot as plt
this_dir = os.path.abspath(os.getcwd())
def get_args():
parser = argparse.ArgumentParser(description=None)
parser.add_argument('--save_folder', default='experiment-duffing-4', type=str)
# Data variables
parser.add_argument('--osc_params', default=(1, 0, 0.1, 0), type=tuple)
parser.add_argument('--train_nb_sample', default=50, type=int)
parser.add_argument('--train_nb_timestep', default=30, type=int)
parser.add_argument('--train_t_span', default=(0, 3), type=tuple)
parser.add_argument('--train_seed', default=0, type=int)
parser.add_argument('--train_noise_level', default=0.1, type=float)
parser.add_argument('--test_nb_sample', default=50, type=int)
parser.add_argument('--test_nb_timestep', default=200, type=int)
parser.add_argument('--test_t_span', default=(0, 20), type=tuple)
parser.add_argument('--test_seed', default=999, type=int)
parser.add_argument('--test_noise_level', default=0.0, type=float)
# Model variables
parser.add_argument('--nb_object', default=1, type=int)
parser.add_argument('--nb_coords', default=1, type=int)
parser.add_argument('--time_horizon', default=10, type=int)
parser.add_argument('--time_augment', action='store_true')
parser.add_argument('--nb_units', default=1000, type=int)
parser.add_argument('--nb_layers', default=1, type=int)
parser.add_argument('--activation', default='tanh', type=str)
parser.add_argument('--ls_cond', default=[['ode', 0.0, False], ['hamiltonian', 0.0, False], ['ode', 0.5, True]])
parser.add_argument('--ls_color', default=['red', 'dodgerblue', 'seagreen'])
parser.add_argument('--learning_rate', default=2e-4, type=float)
parser.add_argument('--epochs', default=5000, type=int)
parser.add_argument('--verbose', default=0, type=int)
return parser.parse_args()
def run(args):
plt.rc('font', size=10)
plt.rc('axes', labelsize=12)
plt.rcParams['lines.linewidth'] = 3
plt.rcParams['figure.figsize'] = 4.5, 4
save_dir = os.path.join(this_dir, args.save_folder)
os.makedirs(save_dir, exist_ok=True)
print('save directory:', save_dir)
for i in range(len(args.ls_cond)):
cond, col = args.ls_cond[i], args.ls_color[i]
function_type, lambda_trs, use_time_dep_lambda = cond[0], cond[1], cond[2]
split_name = str(function_type) + '_' + str(lambda_trs) + '_' + str(use_time_dep_lambda)
print('experimental condition:', split_name)
ts, xs = get_dataset(nb_samples=args.train_nb_sample, nb_timestep=args.train_nb_timestep,
t_span=args.train_t_span, seed=args.train_seed, noise_level=args.train_noise_level,
params=args.osc_params)
t_train, x_train, y_train = reshape_data(ts=ts, xs=xs, substep=args.time_horizon)
oden = ODENetwork(nb_object=args.nb_object, nb_coords=args.nb_coords, function_type=function_type,
time_horizon=args.time_horizon, time_augment=args.time_augment,
nb_units=args.nb_units, nb_layers=args.nb_layers,
activation=args.activation, lambda_trs=lambda_trs,
use_time_dep_lambda=use_time_dep_lambda, learning_rate=args.learning_rate)
oden.solver().fit(x=[t_train, x_train], y=y_train, epochs=args.epochs, batch_size=len(x_train), verbose=args.verbose)
oden.func.save_weights(os.path.join(save_dir, split_name + '_' + 'weight.h5'))
ts, xs = get_dataset(nb_samples=args.test_nb_sample, nb_timestep=args.test_nb_timestep,
t_span=args.test_t_span, seed=args.test_seed, noise_level=args.test_noise_level,
params=args.osc_params)
t_test, x_test, y_test = ts, xs[:, 0, :], xs[:, 1:, :]
if args.time_augment:
y_pred = runge_kutta_solver(ts=t_test, x0=x_test, func=oden.func)
else:
y_pred = leapfrog_solver(ts=t_test, x0=x_test, func=oden.func, dim=int(args.nb_object * args.nb_coords))
q_true, p_true, q_pred, p_pred = y_test[:, :, 0], y_test[:, :, 1], y_pred[:, :, 0], y_pred[:, :, 1]
e_true = args.osc_params[0] * np.square(q_true) + (args.osc_params[1] * np.square(np.square(q_true)) / 2) + np.square(p_true)
e_pred = args.osc_params[0] * np.square(q_pred) + (args.osc_params[1] * np.square(np.square(q_pred)) / 2) + np.square(p_pred)
mse_t_per_sample = np.mean(np.mean(np.square(y_test - y_pred), axis=-1), axis=-1)
mse_t, std_t = 1e2 * np.mean(mse_t_per_sample), 1e2 * np.std(mse_t_per_sample)
mse_e_per_sample = np.mean(np.square(e_true - e_pred), axis=-1)
mse_e, std_e = 1e2 * np.mean(mse_e_per_sample), 1e2 * np.std(mse_e_per_sample)
print('trajectory MSE: {:.2f} pm {:.2f}'.format(mse_t, std_t))
print('energy MSE: {:.2f} pm {:.2f}'.format(mse_e, std_e))
result = {'ground_truth': y_test, 'predicted': y_pred, 'mse_t': mse_t, 'std_t': std_t, 'mse_e': mse_e, std_t: std_e}
with open(os.path.join(save_dir, split_name + '_' + 'result.pkl'), 'wb') as f:
pkl.dump(result, f)
plt.plot(q_true[1], p_true[1], c='k', label='Ground truth')
plt.plot(q_pred[1], p_pred[1], c=col, label=split_name)
plt.grid(), plt.legend()
plt.xlabel('Position q'), plt.ylabel('Momentum p')
plt.xlim([-1.2, 1.2]), plt.ylim([-1.2, 1.2])
plt.tight_layout()
plt.savefig(os.path.join(save_dir, split_name + '_' + 'trajectory.png'))
plt.close()
plt.plot(t_test[0, 1:].squeeze(), e_true[1], c='k', label='Ground truth')
plt.plot(t_test[0, 1:].squeeze(), e_pred[1], c=col, label=split_name)
plt.grid(), plt.legend()
plt.xlabel('Time'), plt.ylabel('Total energy')
plt.xlim([0, args.test_t_span[1]]), plt.ylim([0, 1.2])
plt.tight_layout()
plt.savefig(os.path.join(save_dir, split_name + '_' + 'energy.png'))
plt.close()
if __name__ == '__main__':
run(get_args())