Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Pass Loggers + Avoid Multiple Calls. #32

Open
wants to merge 3 commits into
base: gpu_upgrade_v1
Choose a base branch
from
Open

Pass Loggers + Avoid Multiple Calls. #32

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
65f7cbb
REFACTOR: Adding substeps to the logger and passing the logger tp eva…
vikashplus Oct 19, 2020
56e8bfd
FIX: Avoiding multiple calls to get_obs. Multiple calls results in du…
vikashplus Oct 19, 2020
11f8393
FIX: Removing duplicage log entry
vikashplus Oct 19, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
76 changes: 51 additions & 25 deletions mjrl/algos/model_accel/run_experiments/run_model_accel_npg.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -110,16 +110,16 @@ def buffer_size(paths_list):
termination_function = getattr(e.env.env, "truncate_paths", None)
if 'obs_mask' in globals(): e.obs_mask = obs_mask

models = [WorldModel(state_dim=e.observation_dim, act_dim=e.action_dim, seed=SEED+i,
models = [WorldModel(state_dim=e.observation_dim, act_dim=e.action_dim, seed=SEED+i,
**job_data) for i in range(job_data['num_models'])]
policy = MLP(e.spec, seed=SEED, hidden_sizes=job_data['policy_size'],
policy = MLP(e.spec, seed=SEED, hidden_sizes=job_data['policy_size'],
init_log_std=job_data['init_log_std'], min_log_std=job_data['min_log_std'])
if 'init_policy' in job_data.keys():
if job_data['init_policy'] != None: policy = pickle.load(open(job_data['init_policy'], 'rb'))
baseline = MLPBaseline(e.spec, reg_coef=1e-3, batch_size=256, epochs=1, learn_rate=1e-3,
use_gpu=(True if job_data['device'] == 'cuda' else False))
use_gpu=(True if job_data['device'] == 'cuda' else False))
agent = ModelAccelNPG(learned_model=models, env=e, policy=policy, baseline=baseline, seed=SEED,
normalized_step_size=job_data['step_size'], save_logs=True,
normalized_step_size=job_data['step_size'], save_logs=True,
reward_function=reward_function, termination_function=termination_function,
**job_data['npg_hp'])

Expand All @@ -130,13 +130,16 @@ def buffer_size(paths_list):

for outer_iter in range(job_data['num_iter']):

ts = timer.time()
# =================================
# Sample data from true dynamics
# =================================
t_start = timer.time()
print("================> ITERATION : %i " % outer_iter)
print("Getting interaction data from real dynamics ...")
# set the policy device back to CPU for env sampling
agent.policy.to('cpu')
samples_to_collect = job_data['init_samples'] if outer_iter == 0 else job_data['iter_samples']
iter_paths = sampler.sample_data_batch(samples_to_collect, agent.env,
iter_paths = sampler.sample_data_batch(samples_to_collect, agent.env,
agent.policy, eval_mode=False, base_seed=SEED + outer_iter, num_cpu=job_data['num_cpu'])
for p in iter_paths:
paths.append(p)
Expand All @@ -161,41 +164,60 @@ def buffer_size(paths_list):
except:
pass

t1 = timer.time()
logger.log_kv('data_collect_time', t1-ts)
t_sample = timer.time()
logger.log_kv('data_collect_time', t_sample-t_start)
print("Data gathered, fitting model ...")

# =================================
# Model updates
# =================================
if job_data['refresh_fit']:
models = [WorldModel(state_dim=e.observation_dim, act_dim=e.action_dim, seed=SEED+123*outer_iter,
**job_data) for i in range(job_data['num_models'])]

for i, model in enumerate(models):
loss_general = model.compute_loss(s[-samples_to_collect:],
loss_general = model.compute_loss(s[-samples_to_collect:],
a[-samples_to_collect:], sp[-samples_to_collect:]) # generalization error
dynamics_loss = model.fit_dynamics(s, a, sp, **job_data)
n_substep = len(dynamics_loss)
for i_substep in range(job_data['fit_epochs']):
logger.log_kv('dyn_loss_{}.{}'.format(i, i_substep), dynamics_loss[i_substep] if i_substep<n_substep else np.nan)
logger.log_kv('dyn_loss_' + str(i), dynamics_loss[-1])
logger.log_kv('dyn_loss_gen_' + str(i), loss_general)
if job_data['learn_reward']:

t_model = timer.time()
logger.log_kv('model_update_time', t_model-t_sample)

# =================================
# Rewards updates
# =================================

if job_data['learn_reward']:
for i, model in enumerate(models):
reward_loss = model.fit_reward(s, a, r.reshape(-1, 1), **job_data)
n_substep = len(reward_loss)
for i_substep in range(job_data['fit_epochs']):
logger.log_kv('rew_loss_{}.{}'.format(i, i_substep), reward_loss[i_substep] if i_substep<n_substep else np.nan)
logger.log_kv('rew_loss_' + str(i), reward_loss[-1])
t2 = timer.time()
logger.log_kv('model_update_time', t2-t1)
t_rewards = timer.time()
logger.log_kv('reward_update_time', t_rewards-t_model)


# =================================
# Refresh policy if necessary
# Policy updates
# =================================

# Refresh policy if necessary -----
if 'refresh_policy' in job_data.keys():
# start policy optimization from scratch (note that data has already been collected with an improved policy)
if job_data['refresh_policy']:
policy = MLP(e.spec, seed=SEED, hidden_sizes=job_data['policy_size'],
policy = MLP(e.spec, seed=SEED, hidden_sizes=job_data['policy_size'],
init_log_std=job_data['init_log_std'], min_log_std=job_data['min_log_std'])
agent.policy = policy
else:
pass

# =================================
# NPG updates
# =================================
# NPG udpates ---------------------
agent.learned_model = models
for inner_step in range(job_data['inner_steps']):
if job_data['start_state'] == 'init':
Expand All @@ -216,13 +238,17 @@ def buffer_size(paths_list):
init_states_2 = list(s[buffer_rand_idx])
init_states = init_states_1 + init_states_2

agent.train_step(N=len(init_states), init_states=init_states, horizon=job_data['horizon'])
statistics = agent.train_step(N=len(init_states), init_states=init_states, horizon=job_data['horizon'])
print_data = sorted(filter(lambda v: np.asarray(v[1]).size == 1,
agent.logger.get_current_log().items()))
print(tabulate(print_data))
logger.log_kv("rollout_score_mean_0.{}".format(inner_step), statistics[0])
logger.log_kv("rollout_score_std_0.{}".format(inner_step), statistics[1])
logger.log_kv("rollout_score_min_0.{}".format(inner_step), statistics[2])
logger.log_kv("rollout_score_max_0.{}".format(inner_step), statistics[3])

t3 = timer.time()
logger.log_kv('policy_update_time', t3-t2)
t_policy = timer.time()
logger.log_kv('policy_update_time', t_policy-t_rewards)

if job_data['eval_rollouts'] > 0:
print("Performing validation rollouts ... ")
Expand All @@ -233,7 +259,7 @@ def buffer_size(paths_list):
eval_score = np.mean([np.sum(p['rewards']) for p in eval_paths])
logger.log_kv('eval_score', eval_score)
try:
eval_metric = e.env.env.evaluate_success(eval_paths)
eval_metric = e.env.env.evaluate_success(eval_paths, logger)
logger.log_kv('eval_metric', eval_metric)
except:
pass
Expand All @@ -255,9 +281,9 @@ def buffer_size(paths_list):
pickle.dump(best_policy, open(OUT_DIR + '/iterations/best_policy.pickle', 'wb'))
agent.to(job_data['device'])

tf = timer.time()
logger.log_kv('eval_log_time', tf-t3)
logger.log_kv('iter_time', tf-ts)
t_final = timer.time()
logger.log_kv('eval_log_time', t_final-t_policy)
logger.log_kv('iter_time', t_final-t_start)
print_data = sorted(filter(lambda v: np.asarray(v[1]).size == 1,
logger.get_current_log().items()))
print(tabulate(print_data))
Expand Down
17 changes: 9 additions & 8 deletions mjrl/algos/model_accel/sampling.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,7 +29,7 @@ def policy_rollout(
a_max=None,
large_value=float(1e2),
):

# Only CPU rollouts are currently supported.
# TODO(Aravind) : Extend GPU support

Expand Down Expand Up @@ -245,18 +245,18 @@ def evaluate_policy(e, policy, learned_model, noise_level=0.0,
env_infos = []
t = 0
done = False
o = e.get_obs()
while t < e.horizon and done is False:
o = e.get_obs()
ifo = e.get_env_infos()
a = policy.get_action(o)
if type(a) == list:
a = a[1]['evaluation']
if noise_level > 0.0:
a = a + e.env.env.np_random.uniform(low=-noise_level, high=noise_level, size=a.shape[0])
if real_step is False:
next_s = learned_model.predict(o, a)
next_o = learned_model.predict(o, a)
r = 0.0 # temporarily
e.env.env.set_fitted_state(next_s)
e.env.env.set_fitted_state(next_o)
else:
next_o, r, done, ifo2 = e.step(a)
ifo = ifo2 if ifo == {} else ifo
Expand All @@ -268,6 +268,7 @@ def evaluate_policy(e, policy, learned_model, noise_level=0.0,
actions.append(a)
rewards.append(r)
env_infos.append(ifo)
o = next_o

path = dict(observations=np.array(observations), actions=np.array(actions),
rewards=np.array(rewards),
Expand All @@ -284,7 +285,7 @@ def evaluate_policy(e, policy, learned_model, noise_level=0.0,
return paths


def enforce_tensor_bounds(torch_tensor, min_val=None, max_val=None,
def enforce_tensor_bounds(torch_tensor, min_val=None, max_val=None,
large_value=float(1e4), device=None):
"""
Clamp the torch_tensor to Box[min_val, max_val]
Expand All @@ -299,17 +300,17 @@ def enforce_tensor_bounds(torch_tensor, min_val=None, max_val=None,

assert type(min_val) == float or type(min_val) == torch.Tensor
assert type(max_val) == float or type(max_val) == torch.Tensor

if type(min_val) == torch.Tensor:
if len(min_val.shape) > 0: assert min_val.shape[-1] == torch_tensor.shape[-1]
else:
min_val = torch.tensor(min_val)

if type(max_val) == torch.Tensor:
if len(max_val.shape) > 0: assert max_val.shape[-1] == torch_tensor.shape[-1]
else:
max_val = torch.tensor(max_val)

min_val = min_val.to(device)
max_val = max_val.to(device)

Expand Down