def generate_test():
mode = 1
T = 500
new_env = physic_env(train_cond, mass_list, force_list, init_mouse, T,
mode, prior, reward_stop)
# s_next, r, is_done, r_others = new_env.step(a)
control_vec = {
'obj': np.repeat(0, T),
'x': np.repeat(0, T),
'y': np.repeat(0, T)
}
current_cond = new_env.update_condition(new_env.cond['mass'],
new_env.cond['lf'])
new_env.update_bodies(current_cond)
local_data = new_env.initial_data()
for t in range(T):
local_trajectory = new_env.update_simulate_bodies(
current_cond, control_vec, t, local_data)
trajectory = []
for f in range(T):
trajectory.append([
0, 0, 0, 0, 0, 0, local_trajectory['o1']['x'][f],
local_trajectory['o1']['y'][f], local_trajectory['o1']['vx'][f],
local_trajectory['o1']['vy'][f], local_trajectory['o2']['x'][f],
local_trajectory['o2']['y'][f], local_trajectory['o2']['vx'][f],
local_trajectory['o2']['vy'][f], local_trajectory['o3']['x'][f],
local_trajectory['o3']['y'][f], local_trajectory['o3']['vx'][f],
local_trajectory['o3']['vy'][f], local_trajectory['o4']['x'][f],
local_trajectory['o4']['y'][f], local_trajectory['o4']['vx'][f],
local_trajectory['o4']['vy'][f]
])
return trajectory
from environment import physic_env
import numpy as np
from config import *
import math
import gizeh as gz
import moviepy.editor as mpy
new_env = physic_env(cond,mass_list,force_list,init_mouse,T,ig_mode, prior)
#control_vec = {'obj': np.append(np.repeat(0, 60), np.repeat(1, 180)), 'x':np.repeat(3, 240), 'y':np.repeat(3, 240)}
# test 10 time frame:
#print("*********",self.bodies[0].position[0])
# print("start")
#state, reward, is_done = new_env.step(0)
# state = new_env.reset()
# print(len(state))
#print(reward)
import time
# test the whole process.
#start = time.time()
#print("start",time.strftime("%H:%M:%S", time.localtime()))
for i in range(cond['timeout']/T):
idx = np.random.randint(0,645)
states, reward, is_done = new_env.step(idx)
print(is_done)
print(reward)
data = new_env.step_data()
fig = plt.gcf()
fig.savefig('Qlearning_{}_cum_reward.png'.format(name))
if args.save_model:
model_json = nn.to_json()
with open('Qlearning_{}.json'.format(name), 'w') as json_file:
json_file.write(model_json)
nn.save_weights('Qlearning_{}.h5'.format(name))
print("Model saved!")
np.savetxt('Qlearning_{}.txt'.format(name), (rewards, loss, np.cumsum(rewards).tolist()))
print("Training details saved!")
plt.show()
return
if __name__ == "__main__":
parser = argparse.ArgumentParser(description='training a q-function approximator')
parser.add_argument('--epochs', type=int, action='store', help='number of epoches to train', default=10)
parser.add_argument('--mode', type=int, action='store', help='type of intrinsic reward, 1 for mass, 2 for force', default=1)
parser.add_argument('--save_model', type=bool, action='store', help='save trained model or not', default=True)
parser.add_argument('--sessions', type=int, action='store', help='number of sessions to train per epoch', default=10)
args = parser.parse_args()
print(args)
# initialize the environment
new_env = physic_env(train_cond, mass_list, force_list, init_mouse, T, args.mode, prior, reward_stop)
# train
train_loop(args)