def main(env_id, render, num_process, lr_p, lr_v, gamma, tau, epsilon,
batch_size, ppo_mini_batch_size, ppo_epochs, max_iter, eval_iter,
save_iter, model_path, log_path, seed):
base_dir = log_path + env_id + "/PPO_exp{}".format(seed)
writer = SummaryWriter(base_dir)
ppo = PPO(env_id=env_id,
render=render,
num_process=1,
min_batch_size=batch_size,
lr_p=lr_p,
lr_v=lr_v,
gamma=gamma,
tau=tau,
clip_epsilon=epsilon,
ppo_epochs=ppo_epochs,
ppo_mini_batch_size=ppo_mini_batch_size,
seed=seed,
model_path='trained_models')
for i_iter in range(1, 6):
ppo.eval(i_iter, render=True)
torch.cuda.empty_cache()
def main(env_id, dim_latent, render, num_process, lr_p, lr_v, gamma, tau,
epsilon, batch_size, ppo_mini_batch_size, ppo_epochs, max_iter,
eval_iter, save_iter, model_path, log_path, seed):
base_dir = log_path + env_id + "/PPO_encoder_exp{}".format(seed)
writer = SummaryWriter(base_dir)
ppo = PPO(
env_id=env_id,
dim_latent=dim_latent,
render=render,
num_process=20, #cpu_count(),
min_batch_size=batch_size,
lr_p=lr_p,
lr_v=lr_v,
gamma=gamma,
tau=tau,
clip_epsilon=epsilon,
ppo_epochs=ppo_epochs,
ppo_mini_batch_size=ppo_mini_batch_size,
seed=seed)
for i_iter in range(1, max_iter + 1):
ppo.learn(writer, i_iter)
if i_iter % eval_iter == 0:
ppo.eval(i_iter, render=render)
if i_iter % save_iter == 0:
ppo.save(model_path)
pickle.dump(
ppo,
open('{}/{}_ppo_encoder.p'.format(model_path, env_id), 'wb'))
torch.cuda.empty_cache()
v_lr=args.v_lr,
dyn_lr=args.dyn_lr,
clip_param=args.clip_param,
value_coef=args.value_coef,
entropy_coef=args.entropy_coef,
grad_norm_max=args.grad_norm_max,
use_clipped_value_loss=True,
use_tensorboard=args.use_tensorboard,
add_intrinsic_reward=args.add_intrinsic_reward,
predict_delta_obs=args.predict_delta_obs,
device=device,
share_optim=args.share_optim,
debug=None)
ob_rms = agent.load_checkpoint(checkpoint_dir)
agent.eval()
# set same statistics for normalization as in training
if ob_rms is not None and isinstance(envs.venv, VecNormalize):
envs.venv.ob_rms = ob_rms
# start testing
start = time.time()
for trial in range(args.num_evals):
print('Trial ', trial, 'of', args.num_evals)
obs = envs.reset()
agent.rollouts.obs[0].copy_(obs[1])
agent.rollouts.to(device)
