def main_loop():
colors = []
num_episodes = args.num_ensembles
for i in range(num_episodes):
colors.append('#%06X' % randint(0, 0xFFFFFF))
improved_context_list_np = sample_initial_context_normal(
args.num_ensembles)
if initial_training:
train_on_initial(improved_context_list_np)
for i_iter in range(args.max_iter_num):
# generate multiple trajectories that reach the minimum batch_size
policy_np.training = False
if len(replay_memory) == 0 or not args.rm_as_context:
context_list_np = improved_context_list_np
else:
context_list_np = replay_memory.data
batch_np, log_np = agent_np.collect_episodes(context_list_np,
args.num_req_steps,
args.num_ensembles)
disc_rew_np = discounted_rewards(batch_np.memory, args.gamma)
iter_dataset_np = BaseDataset(batch_np.memory,
disc_rew_np,
args.device_np,
args.dtype,
max_len=max_episode_len)
print('np avg actions: ', log_np['action_mean'])
advantages_np = estimate_v_a(iter_dataset_np, disc_rew_np,
value_replay_memory, value_np, args)
improved_context_list_np = improvement_step_all(
iter_dataset_np, advantages_np, args.max_kl_np, args)
# training
value_replay_memory.add(iter_dataset_np)
train_value_np(value_replay_memory)
tn0 = time.time()
replay_memory.add(iter_dataset_np)
train_np(replay_memory)
tn1 = time.time()
tot_steps_np.append(tot_steps_np[-1] + log_np['num_steps'])
avg_rewards_np.append(log_np['avg_reward'])
if i_iter % args.log_interval == 0:
print(i_iter)
print('np: \tR_min {:.2f} \tR_max {:.2f} \tR_avg {:.2f}'.format(
log_np['min_reward'], log_np['max_reward'],
log_np['avg_reward']))
print('new sigma', args.fixed_sigma)
plot_rewards_history(tot_steps_np, avg_rewards_np)
store_avg_rewards(
tot_steps_np[-1], avg_rewards_np[-1],
np_file.replace(
str(args.seed) + '.csv', 'avg' + str(args.seed) + '.csv'))
if tot_steps_np[-1] > args.tot_steps:
break
def main_loop():
improved_context_list_np = sample_initial_context_normal(env, init_sigma=0.05)
train_on_initial(improved_context_list_np)
#policy_np.apply(InitFunc.init_zero)
for i_iter in range(args.max_iter_num):
# define context set
policy_np.training = False
if len(replay_memory) == 0 or not args.rm_as_context:
context_list_np = improved_context_list_np
else:
context_list_np = replay_memory.data
# collect samples
batch_np, log_np = agent_np.collect_episodes(context_list_np, args.num_req_steps, args.num_ensembles)
# compute discounted rewards
disc_rew_np = discounted_rewards(batch_np.memory, args.gamma)
iter_dataset_np = BaseDataset(batch_np.memory, disc_rew_np, args.device_np, args.dtype, max_len=max_episode_len)
# estimate advantages
if args.learn_baseline:
if args.value_net:
state_list = [ep['states'][:ep['real_len']] for ep in iter_dataset_np]
advantages_np, returns = critic_estimate(value_net, state_list, disc_rew_np, args)
update_critic(value_net, torch.cat(state_list, dim=0), torch.cat(returns, dim=0))
else:
advantages_np = estimate_v_a(iter_dataset_np, disc_rew_np, value_replay_memory, value_np, args)
value_replay_memory.add(iter_dataset_np)
train_value_np(value_replay_memory)
else:
advantages_np = disc_rew_np
# update step
improved_context_list_np = improvement_step_all(iter_dataset_np, advantages_np, args.max_kl_np, args)
# training
replay_memory.add(iter_dataset_np)
train_np(replay_memory)
# prints & plots
tot_steps_np.append(tot_steps_np[-1] + log_np['num_steps'])
avg_rewards_np.append(log_np['avg_reward'])
if i_iter % args.log_interval == 0:
print(i_iter)
print('np avg actions: ', log_np['action_mean'])
print('np: \tR_min {:.2f} \tR_max {:.2f} \tR_avg {:.2f}'.format(log_np['min_reward'], log_np['max_reward'], log_np['avg_reward']))
# plot_pca_proj(iter_dataset_np, advantages_np, policy_np)
print('new sigma', args.fixed_sigma)
plot_rewards_history(tot_steps_np, avg_rewards_np)
store_avg_rewards(tot_steps_np[-1], avg_rewards_np[-1], np_file.replace(str(args.seed)+'.csv', 'avg'+str(args.seed)+'.csv'))
if tot_steps_np[-1] > args.tot_steps or log_np['avg_reward'] < -5000:
break
"""clean up gpu memory"""
torch.cuda.empty_cache()
def main_loop(improved_context_list):
colors = []
num_episodes = args.num_ensembles
for i in range(num_episodes):
colors.append('#%06X' % randint(0, 0xFFFFFF))
for i_iter in range(args.max_iter_num):
print('sampling episodes') # (1)
# generate multiple trajectories that reach the minimum batch_size
batch, log = agent.collect_episodes(improved_context_list,
render=(i_iter % 10 == 0))
disc_rew = discounted_rewards(batch.memory, args.gamma)
complete_dataset = BaseDataset(batch.memory,
disc_rew,
args.device_np,
args.dtype,
max_len=max_episode_len)
advantages = estimate_v_a(complete_dataset, disc_rew)
t0 = time.time()
improved_context_list = improvement_step_all(complete_dataset,
advantages)
t1 = time.time()
# create training set
tn0 = time.time()
replay_memory.add(complete_dataset)
train_np(replay_memory)
tn1 = time.time()
tv0 = time.time()
if i_iter % args.plot_every == 0:
# plot_initial_context(improved_context_list, colors, env, args, i_iter)
# plot_training_set(i_iter, replay_memory, env, args)
plot_policy(model, improved_context_list, replay_memory, i_iter,
log['avg_reward'], env, args, colors)
plot_improvements(complete_dataset, disc_rew, env, i_iter, args,
colors)
tv1 = time.time()
tot_steps.append(tot_steps[-1] + log['num_steps'])
avg_rewards.append(log['avg_reward'])
if i_iter % args.log_interval == 0:
print(
'{}\tT_sample {:.4f} \tT_update {:.4f} \tR_min {:.2f} \tR_max {:.2f} \tR_avg {:.2f}'
.format(i_iter, log['sample_time'], t1 - t0, log['min_reward'],
log['max_reward'], log['avg_reward']))
print('Training: \tT_policy {:.2f} \nT_plots {:.2f}'.format(
tn1 - tn0, tv1 - tv0))
if i_iter % args.plot_every == 0:
plot_rewards_history(avg_rewards, tot_steps, args)
plot_rewards_history(avg_rewards, tot_steps, args)
"""clean up gpu memory"""
torch.cuda.empty_cache()
def main_loop():
improved_context = sample_initial_context()
avg_rewards = []
for i_iter in range(args.max_iter_num):
print('sampling episodes')
# (1)
# generate multiple trajectories that reach the minimum batch_size
# introduce param context=None when np is policy, these will be the context points used to predict
policy_np.training = False
batch, log, memory = agent.collect_samples(args.min_batch_size, context=improved_context) # batch of batch_size transitions from multiple
print(log['num_steps'], log['num_episodes']) # episodes (separated by mask=0). Stored in Memory
disc_rew = discounted_rewards(batch, args.gamma)
complete_dataset = BaseDataset(batch, disc_rew, args.device_np, args.dtype)
value_replay_memory.add(complete_dataset)
train_value_np(value_replay_memory)
estimated_disc_rew, values_stdevs = estimate_disc_rew(complete_dataset, i_iter)
memory.set_disc_rew(estimated_disc_rew)
t0 = time.time()
all_improved_context_0 = improvement_step(batch)
all_improved_context = imporove_mean_stdv(complete_dataset, estimated_disc_rew, values_stdevs)
t1 = time.time()
key = 'means' if args.improve_mean else 'actions'
improved_context = [all_improved_context['states'], all_improved_context[key]]
# plot improved context and actions' discounted rewards
plot_improvements(batch, improved_context, env, i_iter, args)
# create training set
training_set = all_improved_context['means']
frac_action_in_training = int(frac_replace_actions * training_set.shape[1])
training_set[:, :frac_action_in_training, :] = all_improved_context['actions'][:, :frac_action_in_training, :]
dataset = MemoryDatasetNP(batch, training_set, args.device_np, args.dtype, max_len=999)
replay_memory.add(dataset)
plot_training_set(i_iter, replay_memory, env, args)
print('replay memory size:', len(replay_memory))
train_np(replay_memory)
plot_NP_policy(policy_np, improved_context, i_iter, log['avg_reward'], env, args, num_samples=1)
avg_rewards.append(log['avg_reward'])
if i_iter % args.log_interval == 0:
print('{}\tT_sample {:.4f}\tT_update {:.4f}\tR_min {:.2f}\tR_max {:.2f}\tR_avg {:.2f}'.format(
i_iter, log['sample_time'], t1 - t0, log['min_reward'], log['max_reward'], log['avg_reward']))
plot_rewards_history(avg_rewards, args)
"""clean up gpu memory"""
torch.cuda.empty_cache()
def main_loop():
for i_iter in range(args.max_iter_num):
# collect samples
batch, log = agent.collect_episodes(args.num_req_steps)
# compute discounted rewards
disc_rew_mlp = discounted_rewards(batch.memory, args.gamma)
iter_dataset = BaseDataset(batch.memory,
disc_rew_mlp,
args.device_np,
args.dtype,
max_len=max_episode_len)
# estimate advantages
if args.value_net:
state_list = [ep['states'][:ep['real_len']] for ep in iter_dataset]
advantages, returns = critic_estimate(value_net, state_list,
disc_rew_mlp, args)
update_critic(value_net, torch.cat(state_list, dim=0),
torch.cat(returns, dim=0))
else:
advantages = estimate_v_a(iter_dataset, disc_rew_mlp,
value_replay_memory, value_net, args)
value_replay_memory.add(iter_dataset)
train_value(value_replay_memory)
# returned context not used but added to iter_dataset inside the function
improved_context_list = improvement_step_all(iter_dataset, advantages,
args.max_kl_mlp, args)
# training
replay_memory.add(iter_dataset)
train_policy(replay_memory)
# prints & plots
tot_steps.append(tot_steps[-1] + log['num_steps'])
avg_rewards.append(log['avg_reward'])
if i_iter % args.log_interval == 0:
print(i_iter)
print('mlp: \tR_min {:.2f} \tR_max {:.2f} \tR_avg {:.2f}'.format(
log['min_reward'], log['max_reward'], log['avg_reward']))
print('new sigma', args.fixed_sigma)
store_avg_rewards(
tot_steps[-1], avg_rewards[-1],
mlp_file.replace(
str(args.seed) + '.csv', 'avg' + str(args.seed) + '.csv'))
if i_iter % args.plot_every == 0:
plot_rewards_history(tot_steps, avg_rewards)
if tot_steps[-1] > args.tot_steps:
break
"""clean up gpu memory"""
torch.cuda.empty_cache()
def main_loop():
colors = []
num_episodes = args.num_ensembles
for i in range(num_episodes):
colors.append('#%06X' % randint(0, 0xFFFFFF))
#print('sampling initial context')
if args.init_normal:
improved_context_list = sample_initial_context_normal(args.num_ensembles)
else:
improved_context_list = sample_initial_context_uniform(args.num_ensembles)
plot_initial_context(improved_context_list, colors, env, args, '00')
if initial_training:
train_on_initial(improved_context_list)
for i_iter in range(args.max_iter_num):
print('sampling episodes')
# (1)
# generate multiple trajectories that reach the minimum batch_size
# introduce param context=None when np is policy, these will be the context points used to predict
policy_np.training = False
batch, log = agent.collect_episodes(improved_context_list) # batch of batch_size transitions from multiple
#print(log['num_steps'], log['num_episodes']) # episodes (separated by mask=0). Stored in Memory
disc_rew = discounted_rewards(batch.memory, args.gamma)
complete_dataset = BaseDataset(batch.memory, disc_rew, args.device_np, args.dtype, max_len=max_episode_len)
value_replay_memory.add(complete_dataset)
train_value_np(value_replay_memory)
estimated_disc_rew, values_stdevs = estimate_disc_rew(complete_dataset, i_iter, episode_specific_value=args.episode_specific_value)
t0 = time.time()
improved_context_list = improvement_step(complete_dataset, estimated_disc_rew, values_stdevs)
t1 = time.time()
#plot_initial_context(improved_context_list, colors, env, args, i_iter)
# plot improved context and actions' discounted rewards
if i_iter % args.plot_every == 0:
plot_improvements(complete_dataset, estimated_disc_rew, env, i_iter, args, colors)
# create training set
replay_memory.add(complete_dataset)
train_np(replay_memory)
#plot_training_set(i_iter, replay_memory, env, args)
if i_iter % args.plot_every == 0:
plot_NP_policy(policy_np, improved_context_list, i_iter, log['avg_reward'], env, args, colors)
avg_rewards.append(log['avg_reward'])
if i_iter % args.log_interval == 0 and False:
print('{}\tT_sample {:.4f}\tT_update {:.4f}\tR_min {:.2f}\tR_max {:.2f}\tR_avg {:.2f}'.format(
i_iter, log['sample_time'], t1 - t0, log['min_reward'], log['max_reward'], log['avg_reward']))
plot_rewards_history(avg_rewards, args)
"""clean up gpu memory"""
torch.cuda.empty_cache()
def main_loop():
for i_iter in range(args.max_iter_num):
# (1)
# generate multiple trajectories that reach the minimum batch_size
# introduce param context=None when np is policy, these will be the context points used to predict
batch, log, memory = agent.collect_samples(
args.min_batch_size
) # batch of batch_size transitions from multiple
print(log['num_steps'], log['num_episodes']
) # episodes (separated by mask=0). Stored in Memory
disc_rew = discounted_rewards(batch, args.gamma)
memory.set_disc_rew(disc_rew)
complete_dataset = BaseDataset(batch,
disc_rew,
args.device,
dtype,
max_len=max_episode_len)
t0 = time.time()
update_params_trpo(
batch, i_iter
) # estimate advantages from samples and update policy by TRPO step
t1 = time.time()
plot_policy(policy_net, (i_iter, log['avg_reward'], 'policies'))
if i_iter % args.log_interval == 0:
print(
'{}\tT_sample {:.4f}\tT_update {:.4f}\tR_min {:.2f}\tR_max {:.2f}\tR_avg {:.2f}'
.format(i_iter, log['sample_time'], t1 - t0, log['min_reward'],
log['max_reward'], log['avg_reward']))
if not args.episode_specific_value:
iter_dataset = {}
iter_states, iter_q = merge_padded_lists(
[episode['states'] for episode in complete_dataset], [
episode['discounted_rewards']
for episode in complete_dataset
],
max_lens=[episode['real_len'] for episode in complete_dataset])
iter_dataset['states'] = iter_states
iter_dataset['discounted_rewards'] = iter_q
iter_dataset['real_len'] = iter_states.shape[-2]
complete_dataset = [iter_dataset]
value_replay_memory.add(complete_dataset)
train_value_np(value_replay_memory)
def main_loop():
colors = []
num_episodes = args.num_ensembles
for i in range(num_episodes):
colors.append('#%06X' % randint(0, 0xFFFFFF))
improved_context_list = sample_initial_context_normal(args.num_ensembles)
if initial_training:
train_on_initial(improved_context_list)
for i_iter in range(args.max_iter_num):
if tot_steps_trpo[-1] - tot_steps_np[-1] < 1000:
batch_trpo, log_trpo, memory_trpo = agent_trpo.collect_samples(args.min_batch_size) # batch of batch_size transitions from multiple
update_params_trpo(batch_trpo) # generate multiple trajectories that reach the minimum batch_size
tot_steps_trpo.append(tot_steps_trpo[-1] + log_trpo['num_steps'])
avg_rewards_trpo.append(log_trpo['avg_reward'])
# generate multiple trajectories that reach the minimum batch_size
policy_np.training = False
batch, log = agent.collect_episodes(improved_context_list) # batch of batch_size transitions from multiple
#print(log['num_steps'], log['num_episodes']) # episodes (separated by mask=0). Stored in Memory
disc_rew = discounted_rewards(batch.memory, args.gamma)
complete_dataset = BaseDataset(batch.memory, disc_rew, args.device_np, args.dtype, max_len=max_episode_len)
value_replay_memory.add(complete_dataset)
advantages = estimate_v_a(complete_dataset, disc_rew)
tv0 = time.time()
train_value_np(value_replay_memory)
tv1 = time.time()
t0 = time.time()
improved_context_list = improvement_step_all(complete_dataset, advantages)
t1 = time.time()
#plot_initial_context(improved_context_list, colors, env, args, i_iter)
# create training set
tn0 = time.time()
replay_memory.add(complete_dataset)
train_np(replay_memory)
tn1 = time.time()
tot_steps_np.append(tot_steps_np[-1] + log['num_steps'])
#plot_training_set(i_iter, replay_memory, env, args)
if i_iter % args.plot_every == 0 and False:
plot_NP_policy(policy_np, improved_context_list, replay_memory, i_iter, log['avg_reward'], env, args, colors)
plot_improvements(complete_dataset, advantages, env, i_iter, args, colors)
avg_rewards_np.append(log['avg_reward'])
if i_iter % args.log_interval == 0:
print('{}\tT_sample {:.4f} \tT_update {:.4f} \tR_min {:.2f} \tR_max {:.2f} \tR_avg {:.2f}'.format(
i_iter, log['sample_time'], t1 - t0, log['min_reward'], log['max_reward'], log['avg_reward']))
print('Training: \tT_policy {:.2f} \tT_value {:.2f}'.format(tn1-tn0, tv1-tv0))
if i_iter % args.plot_every == 0:
plot_rewards_history(trpo=[tot_steps_trpo, avg_rewards_trpo], mi=[tot_steps_np, avg_rewards_np],
args=args)
plot_rewards_history(trpo=[tot_steps_trpo, avg_rewards_trpo], mi=[tot_steps_np, avg_rewards_np], args=args)
args.max_kl = args.max_kl*0.99
args.fixed_sigma = args.fixed_sigma*0.96
"""clean up gpu memory"""
torch.cuda.empty_cache()
def main_loop():
improved_context_list_mi = sample_initial_context_normal(env)
for i_iter in range(args.max_iter_num):
# define context set
if len(replay_memory_mi) == 0 or not args.rm_as_context:
context_list_np = improved_context_list_mi
else:
context_list_np = replay_memory_mi.data
# collect samples
batch_mi, log_mi = agent_mi.collect_episodes(context_list_np,
args.num_req_steps,
args.num_ensembles)
# compute discounted rewards
disc_rew_mi = discounted_rewards(batch_mi.memory, args.gamma)
iter_dataset_mi = BaseDataset(batch_mi.memory,
disc_rew_mi,
args.device_np,
args.dtype,
max_len=max_episode_len)
# estimate advantages
if args.value_net:
state_list = [
ep['states'][:ep['real_len']] for ep in iter_dataset_mi
]
advantages_mi, returns = critic_estimate(value_net, state_list,
disc_rew_mi, args)
update_critic(value_net, torch.cat(state_list, dim=0),
torch.cat(returns, dim=0))
else:
advantages_mi = estimate_v_a(iter_dataset_mi, disc_rew_mi,
value_replay_memory, model, args)
value_replay_memory.add(iter_dataset_mi)
# update step
improved_context_list_mi = improvement_step_all(
iter_dataset_mi, advantages_mi, args.max_kl_mi, args)
# training
replay_memory_mi.add(iter_dataset_mi)
train_mi(replay_memory_mi)
# prints & plots
tot_steps_mi.append(tot_steps_mi[-1] + log_mi['num_steps'])
avg_rewards_mi.append(log_mi['avg_reward'])
if i_iter % 1 == 0:
plot_pca_proj(iter_dataset_mi, advantages_mi, model)
if i_iter % args.log_interval == 0:
print(i_iter)
print('mi: \tR_min {:.2f} \tR_max {:.2f} \tR_avg {:.2f}'.format(
log_mi['min_reward'], log_mi['max_reward'],
log_mi['avg_reward']))
print('new sigma', args.fixed_sigma)
store_avg_rewards(
tot_steps_mi[-1], log_mi['avg_reward'],
mi_file.replace(
str(args.seed) + '.csv', 'avg' + str(args.seed) + '.csv'))
if i_iter % args.plot_every == 0:
plot_rewards_history(tot_steps_mi, avg_rewards_mi)
if tot_steps_mi[-1] > args.tot_steps:
break
"""clean up gpu memory"""
torch.cuda.empty_cache()
def main_loop(improved_context_list):
colors = []
num_episodes = args.num_ensembles
for i in range(num_episodes):
colors.append('#%06X' % randint(0, 0xFFFFFF))
for i_iter in range(args.max_iter_num):
if tot_steps_trpo[-1] - tot_steps_mi[-1] < 1000:
batch_trpo, log_trpo, memory_trpo = agent_trpo.collect_samples(
args.min_batch_size
) # batch of batch_size transitions from multiple
update_params_trpo(
batch_trpo
) # generate multiple trajectories that reach the minimum batch_size
tot_steps_trpo.append(tot_steps_trpo[-1] + log_trpo['num_steps'])
avg_rewards_trpo.append(log_trpo['avg_reward'])
batch_mi, log_mi = agent_mi.collect_episodes(improved_context_list,
render=(i_iter % 10 == 0))
disc_rew = discounted_rewards(batch_mi.memory, args.gamma)
complete_dataset = BaseDataset(batch_mi.memory,
disc_rew,
args.device_np,
args.dtype,
max_len=max_episode_len)
advantages = estimate_v_a(complete_dataset, disc_rew)
t0 = time.time()
improved_context_list = improvement_step_all(complete_dataset,
advantages)
t1 = time.time()
# create training set
tn0 = time.time()
replay_memory.add(complete_dataset)
train_np(replay_memory)
tn1 = time.time()
tv0 = time.time()
if False and i_iter % args.plot_every == 0:
# plot_initial_context(improved_context_list, colors, env, args, i_iter)
# plot_training_set(i_iter, replay_memory, env, args)
# plot_policy(model, improved_context_list, replay_memory, i_iter, log['avg_reward'], env, args, colors)
plot_improvements(complete_dataset, disc_rew, env, i_iter, args,
colors)
tv1 = time.time()
tot_steps_mi.append(tot_steps_mi[-1] + log_mi['num_steps'])
avg_rewards_mi.append(log_mi['avg_reward'])
if i_iter % args.log_interval == 0:
print(
'{}\n R_min_trpo {:.2f} \tR_max_trpo {:.2f} \tR_avg_trpo {:.2f}\nR_min_mi {:.2f} \tR_max_mi {:.2f} \tR_avg_mi {:.2f} '
.format(i_iter, log_trpo['min_reward'], log_trpo['max_reward'],
log_trpo['avg_reward'], log_mi['min_reward'],
log_mi['max_reward'], log_mi['avg_reward']))
if i_iter % args.plot_every == 0:
plot_rewards_history(trpo=[tot_steps_trpo, avg_rewards_trpo],
mi=[tot_steps_mi, avg_rewards_mi],
args=args)
plot_rewards_history(trpo=[tot_steps_trpo, avg_rewards_trpo],
mi=[tot_steps_mi, avg_rewards_mi],
args=args)
"""clean up gpu memory"""
torch.cuda.empty_cache()
def main_loop():
improved_context = sample_initial_context(args.min_batch_size, dtype=dtype)
avg_rewards = []
for i_iter in range(args.max_iter_num):
print('sampling episodes')
# (1)
# generate multiple trajectories that reach the minimum batch_size
# introduce param context=None when np is policy, these will be the context points used to predict
policy_np.training = False
batch, log, memory = agent.collect_samples(
args.min_batch_size, context=improved_context
) # batch of batch_size transitions from multiple
print(log['num_steps'], log['num_episodes']
) # episodes (separated by mask=0). Stored in Memory
disc_rew = discounted_rewards(batch, args.gamma)
memory.set_disc_rew(disc_rew)
t0 = time.time()
all_improved_context = improvement_step(batch)
t1 = time.time()
key = 'means' if improve_mean else 'actions'
improved_context = [
all_improved_context['states'], all_improved_context[key]
]
# plot improved context and actions' discounted rewards
plot_improvements(
batch,
[all_improved_context['states'], all_improved_context['means']],
i_iter)
# create training set
training_set = all_improved_context['means']
num_action_in_training = int(frac_replace_actions *
training_set.shape[1])
print('replacing {} means with actions'.format(num_action_in_training))
training_set[:, :num_action_in_training, :] = all_improved_context[
'actions'][:, :num_action_in_training, :]
dataset = MemoryDatasetNP(batch,
training_set,
device_np,
dtype,
max_len=999)
replay_memory.add(dataset)
plot_training_set(i_iter)
print('replay memory size:', len(replay_memory))
train_np(replay_memory)
plot_NP_policy(improved_context, i_iter, log['avg_reward'])
avg_rewards.append(log['avg_reward'])
if i_iter % args.log_interval == 0:
print(
'{}\tT_sample {:.4f}\tT_update {:.4f}\tR_min {:.2f}\tR_max {:.2f}\tR_avg {:.2f}'
.format(i_iter, log['sample_time'], t1 - t0, log['min_reward'],
log['max_reward'], log['avg_reward']))
fig_rew, ax_rew = plt.subplots(1, 1)
ax_rew.plot(np.arange(len(avg_rewards)), avg_rewards)
ax_rew.set_xlabel('iterations')
ax_rew.set_ylabel('average reward')
fig_rew.savefig(directory_path + '/average reward')
plt.close(fig_rew)
"""clean up gpu memory"""
torch.cuda.empty_cache()
def main_loop():
colors = []
num_episodes = args.num_ensembles
for i in range(num_episodes):
colors.append('#%06X' % randint(0, 0xFFFFFF))
improved_context_list = sample_initial_context_normal(args.num_ensembles)
if initial_training:
train_on_initial(improved_context_list)
for i_iter in range(args.max_iter_num):
print('sampling episodes')
# (1)
# generate multiple trajectories that reach the minimum batch_size
policy_np.training = False
batch, log = agent.collect_episodes(improved_context_list) # batch of batch_size transitions from multiple
#print(log['num_steps'], log['num_episodes']) # episodes (separated by mask=0). Stored in Memory
disc_rew = discounted_rewards(batch.memory, args.gamma)
complete_dataset = BaseDataset(batch.memory, disc_rew, args.device_np, args.dtype, max_len=max_episode_len)
if not args.episode_specific_value:
iter_dataset = {}
iter_states, iter_q = merge_padded_lists([episode['states'] for episode in complete_dataset],
[episode['discounted_rewards'] for episode in complete_dataset],
max_lens=[episode['real_len'] for episode in complete_dataset])
iter_dataset['states'] = iter_states
iter_dataset['discounted_rewards'] = iter_q
iter_dataset['real_len'] = iter_states.shape[-2]
value_replay_memory.add([iter_dataset])
else:
value_replay_memory.add(complete_dataset)
estimated_disc_rew, values_stdevs = estimate_disc_rew(complete_dataset, i_iter, episode_specific_value=args.episode_specific_value)
tv0 = time.time()
train_value_np(value_replay_memory)
tv1 = time.time()
t0 = time.time()
improved_context_list = improvement_step_all(complete_dataset, estimated_disc_rew)
t1 = time.time()
#plot_initial_context(improved_context_list, colors, env, args, i_iter)
# plot improved context and actions' discounted rewards
#if i_iter % args.plot_every == 0:
plot_improvements(complete_dataset, estimated_disc_rew, env, i_iter, args, colors)
# create training set
tn0 = time.time()
replay_memory.add(complete_dataset)
train_np(replay_memory)
tn1 = time.time()
#if i_iter % args.plot_every == 0:
# plot_NP_policy(policy_np, improved_context_list, replay_memory, i_iter, log['avg_reward'], env, args, colors)
avg_rewards.append(log['avg_reward'])
if i_iter % args.log_interval == 0:
print('{}\tT_sample {:.4f} \tT_update {:.4f} \tR_min {:.2f} \tR_max {:.2f} \tR_avg {:.2f}'.format(
i_iter, log['sample_time'], t1 - t0, log['min_reward'], log['max_reward'], log['avg_reward']))
print('Training: \tT_policy {:.2f} \tT_value {:.2f}'.format(tn1-tn0, tv1-tv0))
if log['avg_reward'] > 195:
print('converged')
plot_rewards_history(avg_rewards, args)
#if i_iter % args.plot_every == 0:
plot_rewards_history(avg_rewards, args)
#args.fixed_sigma = args.fixed_sigma * args.gamma
plot_rewards_history(avg_rewards, args)
"""clean up gpu memory"""
torch.cuda.empty_cache()
def main_loop():
colors = []
num_episodes = args.num_ensembles
for i in range(num_episodes):
colors.append('#%06X' % randint(0, 0xFFFFFF))
improved_context_list = sample_initial_context_normal(args.num_ensembles)
plot_initial_context(improved_context_list, colors, env, args, '00')
if initial_training:
train_on_initial(improved_context_list)
for i_iter in range(args.max_iter_num):
print('sampling episodes')
# (1)
# generate multiple trajectories that reach the minimum batch_size
policy_np.training = False
batch, log = agent.collect_episodes(
improved_context_list, args.num_req_steps, args.num_ensembles
) # batch of batch_size transitions from multiple
#print(log['num_steps'], log['num_episodes']) # episodes (separated by mask=0). Stored in Memory
estimated_disc_rew = discounted_rewards(batch.memory, args.gamma)
complete_dataset = BaseDataset(batch.memory,
estimated_disc_rew,
args.device_np,
args.dtype,
max_len=max_episode_len)
t0 = time.time()
improved_context_list = improvement_step_all(complete_dataset,
estimated_disc_rew)
t1 = time.time()
#plot_initial_context(improved_context_list, colors, env, args, i_iter)
# plot improved context and actions' discounted rewards
if i_iter % args.plot_every == 0:
plot_improvements(complete_dataset, estimated_disc_rew, env,
i_iter, args, colors)
# create training set
tn0 = time.time()
replay_memory.add(complete_dataset)
train_np(replay_memory)
tn1 = time.time()
#plot_training_set(i_iter, replay_memory, env, args)
if i_iter % args.plot_every == 0:
plot_NP_policy(policy_np, improved_context_list, replay_memory,
i_iter, log['avg_reward'], env, args, colors)
avg_rewards.append(log['avg_reward'])
tot_steps.append(tot_steps[-1] + log['num_steps'])
if i_iter % args.log_interval == 0:
print(
'{}\tT_sample {:.4f} \tT_update {:.4f} \tR_min {:.2f} \tR_max {:.2f} \tR_avg {:.2f}'
.format(i_iter, log['sample_time'], t1 - t0, log['min_reward'],
log['max_reward'], log['avg_reward']))
print('Training: \tT_policy {:.2f}'.format(tn1 - tn0))
if log['avg_reward'] > 95:
print('converged')
plot_rewards_history(avg_rewards, args)
if i_iter % args.plot_every == 0:
plot_rewards_history(avg_rewards, tot_steps, args)
#if args.pick_context:
plot_chosen_context(improved_context_list, args.num_context,
i_iter, args, env)
plot_all_training_set(i_iter, replay_memory, env, args)
if args.fixed_sigma is not None:
args.fixed_sigma = args.fixed_sigma * args.gamma
plot_rewards_history(avg_rewards, args)
"""clean up gpu memory"""
torch.cuda.empty_cache()
model = MultiLayerPerceptron(state_dim, action_dim,
512).to(device).double()
agent = AgentMLP(env, model, num_epi, device, fixed_sigma=args.fixed_sigma)
optimizer = torch.optim.Adam(model.parameters(), lr=3e-4)
model_trainer = MLPTrainer(device, model, optimizer, print_freq=50)
replay_memory = ReplayMemoryDataset(20)
tot_steps = [0]
avg_rewards = [0]
for i_iter in range(500):
batch, log = agent.collect_episodes(
) # batch of batch_size transitions from multiple
disc_rew = discounted_rewards(batch.memory, 0.999)
complete_dataset = BaseDataset(batch.memory,
disc_rew,
device,
torch.float64,
max_len=max_episode_len)
print('average reward at', i_iter, log['avg_reward'].item())
t0 = time.time()
improved_context_list_mi = improvement_step_all(
complete_dataset, disc_rew, 0.01, args)
t1 = time.time()
# create training set
tn0 = time.time()
replay_memory.add(complete_dataset)
data_loader = DataLoader(replay_memory, batch_size=1, shuffle=True)
def main_loop():
colors = []
num_episodes = args.num_ensembles
for i in range(num_episodes):
colors.append('#%06X' % randint(0, 0xFFFFFF))
improved_context_list_np = sample_initial_context_normal(
args.num_ensembles)
for i_iter in range(args.max_iter_num):
# generate multiple trajectories that reach the minimum batch_size
policy_np.training = False
if len(replay_memory) == 0 or not args.rm_as_context:
context_list_np = improved_context_list_np
else:
context_list_np = replay_memory.data
ts0 = time.time()
batch_np, log_np = agent_np.collect_episodes(context_list_np,
args.num_req_steps,
args.num_ensembles)
print('sampling:', time.time() - ts0)
disc_rew_np = discounted_rewards(batch_np.memory, args.gamma)
iter_dataset_np = BaseDataset(batch_np.memory,
disc_rew_np,
args.device_np,
args.dtype,
max_len=max_episode_len)
print('np avg actions: ', log_np['action_mean'])
if args.value_net:
state_list = [
ep['states'][:ep['real_len']] for ep in iter_dataset_np
]
advantages_np, returns = critic_estimate(state_list, disc_rew_np,
args)
update_critic(torch.cat(state_list, dim=0),
torch.cat(returns, dim=0))
else:
advantages_np = estimate_v_a(iter_dataset_np, disc_rew_np,
value_replay_memory, value_np, args)
value_replay_memory.add(iter_dataset_np)
train_value_np(value_replay_memory)
improved_context_list_np = improvement_step_all(
iter_dataset_np, advantages_np, args.max_kl_np, args)
# training
tn0 = time.time()
replay_memory.add(iter_dataset_np)
train_np(replay_memory)
tn1 = time.time()
tot_steps_np.append(tot_steps_np[-1] + log_np['num_steps'])
avg_rewards_np.append(log_np['avg_reward'])
if i_iter % args.plot_every in [0, 1]:
if 'CartPole' in args.env_name:
plot_NP_policy_CP(policy_np,
replay_memory,
i_iter,
env,
args,
use_np_sigma=args.plot_np_sigma)
plot_rm(replay_memory, i_iter, args)
plot_improvements_CP(iter_dataset_np, advantages_np, env,
i_iter, args, colors)
elif 'MountainCar' in args.env_name:
plot_NP_policy_MC(policy_np,
replay_memory,
i_iter,
env,
args,
use_np_sigma=args.plot_np_sigma)
plot_improvements_MC(iter_dataset_np, advantages_np, env,
i_iter, args, colors)
plot_improvements_MC_all(iter_dataset_np, advantages_np, env,
i_iter, args, colors)
if i_iter % args.log_interval == 0:
print(i_iter)
print('np: \tR_min {:.2f} \tR_max {:.2f} \tR_avg {:.2f}'.format(
log_np['min_reward'], log_np['max_reward'],
log_np['avg_reward']))
print('new sigma', args.fixed_sigma)
plot_rewards_history(tot_steps_np, avg_rewards_np)
store_avg_rewards(
tot_steps_np[-1], avg_rewards_np[-1],
np_file.replace(
str(args.seed) + '.csv', 'avg' + str(args.seed) + '.csv'))
if args.fixed_sigma is not None:
sigma_history.append(torch.tensor(args.fixed_sigma))
else:
sigma_history.append(
torch.cat([ep['stddevs']
for ep in iter_dataset_np.data]).mean(dim=0))
plot_sigma_history(sigma_history)
if tot_steps_np[-1] > args.tot_steps:
break
"""clean up gpu memory"""
torch.cuda.empty_cache()
def main_loop():
colors = []
num_episodes = args.num_ensembles
for i in range(num_episodes):
colors.append('#%06X' % randint(0, 0xFFFFFF))
improved_context_list_np = sample_initial_context_normal(
args.num_ensembles)
improved_context_list_mi = improved_context_list_np
if args.use_np:
if initial_training:
train_on_initial(improved_context_list_np)
for i_iter in range(args.max_iter_num):
if args.use_trpo and tot_steps_trpo[
-1] < args.tot_steps: # and tot_steps_trpo[-1] - max(tot_steps_mi[-1], tot_steps_np[-1]) < 1000:
batch_trpo, log, memory_trpo = agent_trpo.collect_samples(
args.min_batch_size
) # batch of batch_size transitions from multiple
store_rewards_trpo(memory_trpo.memory, trpo_file)
update_params_trpo(
batch_trpo
) # generate multiple trajectories that reach the minimum batch_size
tot_steps_trpo.append(tot_steps_trpo[-1] + log['num_steps'])
avg_rewards_trpo.append(log['avg_reward'])
print('trpo avg actions: ', log['action_mean'])
if args.use_np and tot_steps_np[-1] < args.tot_steps:
# generate multiple trajectories that reach the minimum batch_size
policy_np.training = False
batch_np, log_np = agent_np.collect_episodes(
improved_context_list_np
) # batch of batch_size transitions from multiple
store_rewards(batch_np.memory, np_file)
disc_rew_np = discounted_rewards(batch_np.memory, args.gamma)
complete_dataset_np = BaseDataset(batch_np.memory,
disc_rew_np,
args.device_np,
args.dtype,
max_len=max_episode_len)
print('np avg actions: ', log_np['action_mean'])
advantages_np = estimate_v_a(complete_dataset_np, disc_rew_np)
improved_context_list_np = improvement_step_all(
complete_dataset_np, advantages_np, args.max_kl_np)
# training
value_replay_memory.add(complete_dataset_np)
train_value_np(value_replay_memory)
tn0 = time.time()
replay_memory.add(complete_dataset_np)
train_np(replay_memory)
tn1 = time.time()
tot_steps_np.append(tot_steps_np[-1] + log_np['num_steps'])
avg_rewards_np.append(log_np['avg_reward'])
if args.use_mi and tot_steps_mi[-1] < args.tot_steps:
# generate multiple trajectories that reach the minimum batch_size
batch_mi, log_mi = agent_mi.collect_episodes(
improved_context_list_mi
) # batch of batch_size transitions from multiple
store_rewards(batch_mi.memory, mi_file)
#print(log['num_steps'], log['num_episodes']) # episodes (separated by mask=0). Stored in Memory
print('mi avg actions: ', log_mi['action_mean'])
disc_rew_mi = discounted_rewards(batch_mi.memory, args.gamma)
complete_dataset_mi = BaseDataset(batch_mi.memory,
disc_rew_mi,
args.device_np,
args.dtype,
max_len=max_episode_len)
advantages_mi = estimate_v_a_mi(complete_dataset_mi, disc_rew_mi)
t0 = time.time()
improved_context_list_mi = improvement_step_all(
complete_dataset_mi, advantages_mi, args.max_kl_mi)
t1 = time.time()
# create training set
tn0 = time.time()
replay_memory_mi.add(complete_dataset_mi)
train_mi(replay_memory_mi)
tn1 = time.time()
tot_steps_mi.append(tot_steps_mi[-1] + log_mi['num_steps'])
avg_rewards_mi.append(log_mi['avg_reward'].item())
if i_iter % args.log_interval == 0:
print(i_iter)
if args.use_trpo:
print('trpo: \tR_min {:.2f} \tR_max {:.2f} \tR_avg {:.2f}'.
format(log['min_reward'], log['max_reward'],
log['avg_reward']))
if args.use_np:
print(
'np: \tR_min {:.2f} \tR_max {:.2f} \tR_avg {:.2f}'.format(
log_np['min_reward'], log_np['max_reward'],
log_np['avg_reward']))
if args.use_mi:
print(
'mi: \tR_min {:.2f} \tR_max {:.2f} \tR_avg {:.2f}'.format(
log_mi['min_reward'], log_mi['max_reward'],
log_mi['avg_reward']))
print(args.fixed_sigma)
if i_iter % args.plot_every == 0:
plot_rewards_history(
[tot_steps_trpo, tot_steps_np, tot_steps_mi],
[avg_rewards_trpo, avg_rewards_np, avg_rewards_mi])
"""clean up gpu memory"""
torch.cuda.empty_cache()
