def single_process(job):
job_start_time = timer.time()
# Allow process to parallelize things internally
curr_proc = mp.current_process()
curr_proc.daemon = False
os.chdir(cwd)
dirpath = os.path.join(job['save_dir'], job['job_name'])
os.makedirs(dirpath, exist_ok=True)
# start job
os.chdir(cwd)
job_start_time = timer.time()
print('Started New Job : ', job['job_name'], '=======================')
print('Job specifications : \n', job)
# Make Env
e = GymEnv(job['env_name'])
# Make baseline
baseline = MLPBaseline(e.spec)
# save job details
job['horizon'] = e.horizon
job['ctrl_timestep'] = e.env.env.dt
job['sim_timestep'] = e.env.env.model.opt.timestep
# job['sim_skip'] = e.env.env.skip
job_data_file = open(dirpath + '/job_data.txt', 'w')
pprint.pprint(job, stream=job_data_file)
job_data_file.close()
# Make policy (???vik: sizes are hard coded)
if 'init_policy' in job:
policy = MLP(e.spec,
init_log_std=job['init_std'],
hidden_sizes=(32, 32),
seed=job['seed'])
loaded_policy = pickle.load(open(job['init_policy'], 'rb'))
loaded_params = loaded_policy.get_param_values()
print('log std values in loaded policy = ')
print(params[-policy.m:])
# NOTE: if the log std is too small
# (say <-2.0, it is problem dependent and intuition should be used)
# then we need to bump it up so that it explores
# params[-policy.m:] += 1.0
policy.set_param_values(loaded_params)
del job['init_policy']
else:
policy = MLP(e.spec,
init_log_std=job['init_std'],
hidden_sizes=(32, 32),
seed=job['seed'])
# Agent
agent = NPG(e, policy, baseline, seed=job['seed'], \
normalized_step_size=job['normalized_step_size'], \
save_logs=job['save_logs'], FIM_invert_args=job['FIM_invert_args'])
# Train Agent
train_agent(
job_name=dirpath,
agent=agent,
seed=job['seed'],
niter=job['niter'],
gamma=job['gamma'],
gae_lambda=job['gae_lambda'],
num_cpu=job['num_cpu'],
sample_mode=job['sample_mode'],
num_traj=job['num_traj'],
evaluation_rollouts=job['evaluation_rollouts'],
save_freq=job['save_freq'],
plot_keys={'stoc_pol_mean', 'stoc_pol_std'},
)
total_job_time = timer.time() - job_start_time
print('Job', job['job_name'],
'took %f seconds ==============' % total_job_time)
return total_job_time
def single_process(job):
job_start_time = timer.time()
# Allow process to parallelize things internally
curr_proc = mp.current_process()
curr_proc.daemon = False
# Create a directory for the job results.
job_dir = os.path.join(job['output_dir'])
if not os.path.isdir(job_dir):
os.mkdir(job_dir)
# start job
job_start_time = timer.time()
print('Started New Job : ', job['job_name'], '=======================')
print('Job specifications : \n', job)
# Make Env
env_name = job['env_name']
# adept_envs.global_config.set_config(env_name, {
# 'robot_params': job['robot'],
# **job.get('env_params', {}),
# })
e = GymEnv(env_name)
# Make baseline
baseline = MLPBaseline(e.spec)
# save job details
job['horizon'] = e.horizon
job['ctrl_timestep'] = e.env.env.dt
job['sim_timestep'] = e.env.env.model.opt.timestep
# job['sim_skip'] = e.env.env.skip
with open(os.path.join(job_dir, 'job_data.txt'), 'w') as job_data_file:
pprint.pprint(job, stream=job_data_file)
if 'init_policy' in job:
policy = MLP(e.spec, init_log_std=job['init_std'], hidden_sizes=(32,32), seed=job['seed'])
loaded_policy = pickle.load(open(job['init_policy'], 'rb'))
loaded_params = loaded_policy.get_param_values()
print("log std values in loaded policy = ")
print(loaded_params[-policy.m:])
# NOTE: if the log std is too small
# (say <-2.0, it is problem dependent and intuition should be used)
# then we need to bump it up so that it explores
loaded_params[-policy.m:] += job['init_std']
policy.set_param_values(loaded_params)
del job['init_policy']
else:
policy = MLP(
e.spec,
init_log_std=job['init_std'],
hidden_sizes=job['hidden_sizes'],
# hidden_sizes=(32, 32),
seed=job['seed'])
# Agent
agent = NPG(
e,
policy,
baseline,
seed=job['seed'],
normalized_step_size=job['normalized_step_size'],
save_logs=job['save_logs'],
FIM_invert_args=job['FIM_invert_args'])
# Train Agent
train_agent(
job_name=job['job_name'],
agent=agent,
# save_dir=job_dir,
seed=job['seed'],
niter=job['niter'],
gamma=job['gamma'],
gae_lambda=job['gae_lambda'],
num_cpu=job['num_cpu'],
sample_mode=job['sample_mode'],
num_traj=job.get('num_traj'),
num_samples=job.get('num_samples'),
evaluation_rollouts=job['evaluation_rollouts'],
save_freq=job['save_freq'],
plot_keys={'stoc_pol_mean', 'stoc_pol_std'},
)
total_job_time = timer.time() - job_start_time
print('Job', job['job_name'],
'took %f seconds ==============' % total_job_time)
return total_job_time
