def _train_policy(self, dataset):
"""
Train the model-based policy
implementation details:
(a) Train for self._training_epochs number of epochs
(b) The dataset.random_iterator(...) method will iterate through the dataset once in a random order
(c) Use self._training_batch_size for iterating through the dataset
(d) Keep track of the loss values by appending them to the losses array
"""
timeit.start('train policy')
losses = []
### PROBLEM 1
### YOUR CODE HERE
for _ in range(self._training_epochs):
for states, actions, next_states, _, _ in dataset.random_iterator(self._training_batch_size):
# import sys; print(sys._getframe().f_code.co_name,sys._getframe().f_lineno)
# from IPython import embed; embed()
losses.append(self._policy.train_step(states, actions, next_states))
logger.record_tabular('TrainingLossStart', losses[0])
logger.record_tabular('TrainingLossFinal', losses[-1])
timeit.stop('train policy')
def _train_policy(self, dataset):
"""
Train the model-based policy
implementation details:
(a) Train for self._training_epochs number of epochs
(b) The dataset.random_iterator(...) method will iterate through the dataset once in a random order
(c) Use self._training_batch_size for iterating through the dataset
(d) Keep track of the loss values by appending them to the losses array
"""
timeit.start('train policy')
losses = []
### PROBLEM 1
### YOUR CODE HERE
# (a) Train for self._training_epochs number of epochs
for _ in range(self._training_epochs):
# (b) The dataset.random_iterator(...) method will iterate through the dataset once in a random order
# (c) Use self._training_batch_size for iterating through the dataset
epoch_losses = []
for _, (states, actions, next_states, _, _) in enumerate(
dataset.random_iterator(self._training_batch_size)):
loss = self._policy.train_step(states, actions, next_states)
epoch_losses.append(loss)
# (d) Keep track of the loss values by appending them to the losses array
losses.append(np.mean(epoch_losses))
logger.record_tabular('TrainingLossStart', losses[0])
logger.record_tabular('TrainingLossFinal', losses[-1])
timeit.stop('train policy')
def _train_policy(self, dataset):
"""
Train the model-based policy
implementation details:
(a) Train for self._training_epochs number of epochs
(b) The dataset.random_iterator(...) method will iterate through the dataset once in a random order
(c) Use self._training_batch_size for iterating through the dataset
(d) Keep track of the loss values by appending them to the losses array
"""
timeit.start('train policy')
losses = []
### PROBLEM 1
### YOUR CODE HERE
# raise NotImplementedError
for _ in range(self._training_epochs):
current_batches = dataset.random_iterator(self._training_batch_size)
while True:
state, action, next_state, _, _ = \
next(current_batches, [None] * 5)
if state is None:
break
loss = self._policy.train_step(state, action, next_state)
losses.append(loss)
logger.record_tabular('TrainingLossStart', losses[0])
logger.record_tabular('TrainingLossFinal', losses[-1])
timeit.stop('train policy')
plt.figure()
plt.plot(losses)
plt.savefig(os.path.join(logger.dir, 'training.png'))
def _train_policy(self, dataset):
"""
Train the model-based policy
implementation details:
(a) Train for self._training_epochs number of epochs
(b) The dataset.random_iterator(...) method will iterate through the dataset once in a random order
(c) Use self._training_batch_size for iterating through the dataset
(d) Keep track of the loss values by appending them to the losses array
"""
timeit.start('train policy')
losses = []
### PROBLEM 1
### YOUR CODE HERE
for epoch in range(self._training_epochs):
for state, action, next_state, _, _ in dataset.random_iterator(
self._training_batch_size):
loss = self._policy.train_step(states=state,
actions=action,
next_states=next_state)
losses.append(loss)
# raise NotImplementedError
logger.record_tabular('TrainingLossStart', losses[0])
logger.record_tabular('TrainingLossFinal', losses[-1])
timeit.stop('train policy')
def _train_policy(self, dataset):
"""
Train the model-based policy
implementation details:
(a) Train for self._training_epochs number of epochs
(b) The dataset.random_iterator(...) method will iterate through the dataset once in a random order
(c) Use self._training_batch_size for iterating through the dataset
(d) Keep track of the loss values by appending them to the losses array
"""
timeit.start('train policy')
losses = []
### PROBLEM 1
### YOUR CODE HERE
for epoch in range(self._training_epochs):
for states, actions, next_states, _, _ in dataset.random_iterator(
self._training_batch_size):
loss = self._policy.train_step(states, actions, next_states)
losses.append(loss)
# self._random_dataset_test = self._gather_rollouts(self._policy, 2)
# for states, actions, next_states, _, _ in self._random_dataset_test.random_iterator(len(self._random_dataset_test)):
# eval_loss = self._policy.eval_loss(states, actions, next_states)
# print("Test loss: " + str(eval_loss))
logger.record_tabular('TrainingLossStart', losses[0])
logger.record_tabular('TrainingLossFinal', losses[-1])
timeit.stop('train policy')
def _train_policy(self, dataset):
"""
Train the model-based policy
implementation details:
(a) Train for self._training_epochs number of epochs
(b) The dataset.random_iterator(...) method will iterate through the dataset once in a random order
(c) Use self._training_batch_size for iterating through the dataset
(d) Keep track of the loss values by appending them to the losses array
"""
timeit.start('train policy')
losses = []
# Added: Training policy iteration
for epoch_num in range(self._training_epochs):
logger.info('Epoch %i' % (epoch_num + 1))
for batch_num, (states, actions, next_states, _, _) in enumerate(
dataset.random_iterator(self._training_batch_size)):
loss = self._policy.train_step(states, actions, next_states)
losses.append(loss)
logger.info('\tLoss: {:.3f}'.format(losses[-1]))
logger.record_tabular('TrainingLossStart', losses[0])
logger.record_tabular('TrainingLossFinal', losses[-1])
timeit.stop('train policy')
def _train_policy(self, dataset):
"""
Train the model-based policy
"""
timeit.start('train policy')
losses = []
### PROBLEM 1
### YOUR CODE HERE
# (a) Train for self._training_epochs number of epochs
for ep in range(self._training_epochs):
# (b) The dataset.random_iterator(...) method will iterate through the dataset once in a random order,
# it will return a dataset with random sequence
# (c) Use self._training_batch_size for iterating through the dataset
_iter = dataset.random_iterator(self._training_batch_size)
for states, actions, next_states, _, _ in _iter:
loss = self._policy.train_step(states, actions, next_states)
# dataset_size = dones[0]
# batch_start_index = np.array(0,dataset_size, self._training_batch_size)
# batch_end_index = batch_start_index + self._training_batch_size
# for (start, end) in zip(batch_start_index,batch_end_index):
# loss = self._policy.train_step(states[start:end], actions[start:end], next_states[start:end]
losses.append(loss)
logger.record_tabular('TrainingLossStart', losses[0])
logger.record_tabular('TrainingLossFinal', losses[-1])
timeit.stop('train policy')
def train_policy(self, dataset):
"""
trains the model-based policy
"""
timeit.start('train policy')
losses = []
for _ in range(self.training_epochs):
loss_total = 0.0
num_data = 0
d = dataset.random_iterator(self.training_batch_size)
for states, actions, next_states, _, _ in d:
loss = self.policy.train_step(states, actions, next_states)
loss_total += loss
num_data += 1
losses.append(loss / num_data)
# plt.plot(losses)
# plt.show()
logger.record_tabular('TrainingLossStart', losses[0])
logger.record_tabular('TrainingLossFinal', losses[-1])
timeit.stop('train policy')
return
def _train_policy(self, dataset):
"""
Train the model-based policy
implementation details:
(a) Train for self._training_epochs number of epochs
(b) The dataset.random_iterator(...) method will iterate through the dataset once in a random order
(c) Use self._training_batch_size for iterating through the dataset
(d) Keep track of the loss values by appending them to the losses array
"""
timeit.start('train policy')
losses = []
### PROBLEM 1
### YOUR CODE HERE
# raise NotImplementedError
for ep in range(self._training_epochs):
data_generator = dataset.random_iterator(self._training_batch_size)
for i, batch_x in enumerate(data_generator):
states, actions, next_states = batch_x[:3]
loss = self._policy.train_step(states, actions, next_states)
losses.append(loss)
# logger.debug('%s/Epoch, it %s: Loss: %s' % (ep, i, loss))
logger.info('loss start to end: %s, %s' % (losses[0], losses[-1]))
logger.record_tabular('TrainingLossStart', losses[0])
logger.record_tabular('TrainingLossFinal', losses[-1])
timeit.stop('train policy')
def run_q3(self):
"""
start with random dataset, train policy on dataset, gather rollouts with policy, add to
dataset, repeat
"""
dataset = self.random_dataset
itr = -1
logger.info('Iteration {0}'.format(itr))
logger.record_tabular('Itr', itr)
self.log(dataset)
for itr in range(self.num_onpolicy_iters + 1):
logger.info('Iteration {0}'.format(itr))
logger.record_tabular('Itr', itr)
logger.info('Training policy...')
self.train_policy(self.random_dataset)
logger.info('Gathering rollouts...')
new_dataset = self.gather_rollouts(self.policy,
self.num_onpolicy_rollouts)
logger.info('Appending dataset...')
dataset.append(new_dataset)
self.log(new_dataset)
def run_q3(self):
"""
Starting with the random dataset, train the policy on the dataset, gather rollouts with the policy,
append the new rollouts to the existing dataset, and repeat
"""
dataset = self._random_dataset
itr = -1
logger.info('Iteration {0}'.format(itr))
logger.record_tabular('Itr', itr)
self._log(dataset)
for itr in range(self._num_onpolicy_iters + 1):
logger.info('Iteration {0}'.format(itr))
logger.record_tabular('Itr', itr)
### PROBLEM 3
### YOUR CODE HERE
logger.info('Training policy...')
self._train_policy(dataset)
### PROBLEM 3
### YOUR CODE HERE
logger.info('Gathering rollouts...')
new_dataset = self._gather_rollouts(self._policy,
self._num_onpolicy_rollouts)
### PROBLEM 3
### YOUR CODE HERE
logger.info('Appending dataset...')
dataset.append(new_dataset)
self._log(new_dataset)
def _train_policy(self, dataset):
# timing for policy training
timeit.start('train policy')
losses = []
# loop for self._training_epochs
for _ in range(self._training_epochs):
# iterate over dataset
for states, actions, next_states, _, _ in \
dataset.random_iterator(self._training_batch_size):
# compute loss
loss = self._policy.train_step(states, actions, next_states)
losses.append(loss)
# perform logging
logger.record_tabular('TrainingLossStart', losses[0])
logger.record_tabular('TrainingLossFinal', losses[-1])
timeit.stop('train policy')
def _train_policy(self, dataset):
"""
Train the model-based policy
implementation details:
(a) Train for self._training_epochs number of epochs
(b) The dataset.random_iterator(...) method will iterate through the dataset once in a random order
(c) Use self._training_batch_size for iterating through the dataset
(d) Keep track of the loss values by appending them to the losses array
"""
timeit.start('train policy')
### PROBLEM 1
### YOUR CODE HERE
# raise NotImplementedError
print('batch size is', self._training_batch_size)
print('epoch size is', self._training_epochs)
# Iterate dataset once in an epoch
losses = []
for epoch in range(self._training_epochs):
t_loss = 0
# Alternative way:
# random_data = dataset.random_iterator(self._training_batch_size)
# for states, actions, next_states, _, _ in random_data:
# Enumerate is to add index!
for r_num, (states, actions, next_states, _, _) in enumerate(
dataset.random_iterator(self._training_batch_size)):
loss = self._policy.train_step(states, actions, next_states)
t_loss += loss
t_loss = t_loss / (r_num + 1)
losses.append(t_loss)
# TO-DO: why not print out in q1?
logger.record_tabular('TrainingLossStart', losses[0])
logger.record_tabular('TrainingLossFinal', losses[-1])
timeit.stop('train policy')
def log(self):
end_idxs = np.nonzero(self._dones)[0] + 1
returns = []
start_idx = 0
for end_idx in end_idxs:
rewards = self._rewards[start_idx:end_idx]
returns.append(np.sum(rewards))
start_idx = end_idx
logger.record_tabular('ReturnAvg', np.mean(returns))
logger.record_tabular('ReturnStd', np.std(returns))
logger.record_tabular('ReturnMin', np.min(returns))
logger.record_tabular('ReturnMax', np.max(returns))
def train(self):
dataset = self._random_dataset
itr = -1
logger.info('Iteration {0}'.format(itr))
logger.record_tabular('Itr', itr)
self._log(dataset)
for itr in range(self._num_onpolicy_iters + 1):
logger.info('Iteration {0}'.format(itr))
logger.record_tabular('Itr', itr)
logger.info('Training policy...')
self._train_policy(dataset)
logger.info('Gathering rollouts...')
new_dataset = self._gather_rollouts(self._policy,
self._num_onpolicy_rollouts)
logger.info('Appending dataset...')
dataset.append(new_dataset)
self._log(new_dataset)
def log(self, write_table_header=False):
logger.log("Logging data in directory: %s" % logger.get_snapshot_dir())
logger.record_tabular("Episode", self.num_episodes)
logger.record_tabular("Accumulated Training Steps",
self.num_train_interactions)
logger.record_tabular("Policy Error", self.logging_policies_error)
logger.record_tabular("Q-Value Error", self.logging_qvalues_error)
logger.record_tabular("V-Value Error", self.logging_vvalues_error)
logger.record_tabular("Alpha", np_ify(self.log_alpha.exp()).item())
logger.record_tabular("Entropy",
np_ify(self.logging_entropy.mean(dim=(0, ))))
act_mean = np_ify(self.logging_mean.mean(dim=(0, )))
act_std = np_ify(self.logging_std.mean(dim=(0, )))
for aa in range(self.action_dim):
logger.record_tabular("Mean Action %02d" % aa, act_mean[aa])
logger.record_tabular("Std Action %02d" % aa, act_std[aa])
# Evaluation Stats to plot
logger.record_tabular("Test Rewards Mean",
np_ify(self.logging_eval_rewards.mean()))
logger.record_tabular("Test Rewards Std",
np_ify(self.logging_eval_rewards.std()))
logger.record_tabular("Test Returns Mean",
np_ify(self.logging_eval_returns.mean()))
logger.record_tabular("Test Returns Std",
np_ify(self.logging_eval_returns.std()))
# Add the previous times to the logger
times_itrs = gt.get_times().stamps.itrs
train_time = times_itrs.get('train', [0])[-1]
sample_time = times_itrs.get('sample', [0])[-1]
eval_time = times_itrs.get('eval', [0])[-1]
epoch_time = train_time + sample_time + eval_time
total_time = gt.get_times().total
logger.record_tabular('Train Time (s)', train_time)
logger.record_tabular('(Previous) Eval Time (s)', eval_time)
logger.record_tabular('Sample Time (s)', sample_time)
logger.record_tabular('Epoch Time (s)', epoch_time)
logger.record_tabular('Total Train Time (s)', total_time)
# Dump the logger data
logger.dump_tabular(with_prefix=False,
with_timestamp=False,
write_header=write_table_header)
# Save pytorch models
self.save_training_state()
logger.log("----")
# Load buffer
replay_buffer = utils.ReplayBuffer()
if args.env_name == 'Multigoal-v0':
replay_buffer.load_point_mass(buffer_name,
bootstrap_dim=4,
dist_cost_coeff=0.01)
else:
replay_buffer.load(buffer_name, bootstrap_dim=4)
evaluations = []
episode_num = 0
done = True
training_iters = 0
while training_iters < args.max_timesteps:
pol_vals = policy.train(replay_buffer, iterations=int(args.eval_freq))
ret_eval, var_ret, median_ret = evaluate_policy(policy)
evaluations.append(ret_eval)
np.save("./results/" + file_name, evaluations)
training_iters += args.eval_freq
print("Training iterations: " + str(training_iters))
logger.record_tabular('Training Epochs',
int(training_iters // int(args.eval_freq)))
logger.record_tabular('AverageReturn', ret_eval)
logger.record_tabular('VarianceReturn', var_ret)
logger.record_tabular('MedianReturn', median_ret)
logger.dump_tabular()
mmd_sigma=args.mmd_sigma,
lagrange_thresh=args.lagrange_thresh,
use_kl=(True if args.distance_type == "KL" else False),
use_ensemble=(False
if args.use_ensemble_variance == "False" else True),
kernel_type=args.kernel_type)
evaluations = []
episode_num = 0
done = True
training_iters = 0
while training_iters < args.max_timesteps:
pol_vals = policy.train(replay_buffer, iterations=int(args.eval_freq))
# NOTE : @dhruvramani - commeneted this for colabs
# ret_eval, var_ret, median_ret = evaluate_policy(policy)
# evaluations.append(ret_eval)
# np.save("./results/" + file_name, evaluations)
training_iters += args.eval_freq
print("Training iterations: " + str(training_iters))
logger.record_tabular('Training Epochs',
int(training_iters // int(args.eval_freq)))
# logger.record_tabular('AverageReturn', ret_eval)
# logger.record_tabular('VarianceReturn', var_ret)
# logger.record_tabular('MedianReturn', median_ret)
logger.dump_tabular()
print("Iter done")
def train(self, replay_buffer, iterations, batch_size=100):
for it in range(iterations):
# Sample replay buffer / batch
state, action, next_state, reward, not_done = replay_buffer.sample(
batch_size)
# Critic Training
with torch.no_grad():
_, _, next_action = self.actor_target(next_state,
self.vae.decode)
target_Q1, target_Q2 = self.critic_target(
next_state, next_action)
target_Q = self.lmbda * torch.min(target_Q1, target_Q2) + (
1 - self.lmbda) * torch.max(target_Q1, target_Q2)
target_Q = reward + not_done * self.discount * target_Q
current_Q1, current_Q2 = self.critic(state, action)
critic_loss = F.mse_loss(current_Q1, target_Q) + F.mse_loss(
current_Q2, target_Q)
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
# Actor Training
latent_actions, mid_actions, actions = self.actor(
state, self.vae.decode)
actor_loss = -self.critic.q1(state, actions).mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# Update Target Networks
for param, target_param in zip(self.critic.parameters(),
self.critic_target.parameters()):
target_param.data.copy_(self.tau * param.data +
(1 - self.tau) * target_param.data)
for param, target_param in zip(self.actor.parameters(),
self.actor_target.parameters()):
target_param.data.copy_(self.tau * param.data +
(1 - self.tau) * target_param.data)
# Logging
logger.record_dict(
create_stats_ordered_dict(
'Q_target',
target_Q.cpu().data.numpy(),
))
logger.record_tabular('Actor Loss', actor_loss.cpu().data.numpy())
logger.record_tabular('Critic Loss', critic_loss.cpu().data.numpy())
logger.record_dict(
create_stats_ordered_dict('Actions',
actions.cpu().data.numpy()))
logger.record_dict(
create_stats_ordered_dict('Mid Actions',
mid_actions.cpu().data.numpy()))
logger.record_dict(
create_stats_ordered_dict('Latent Actions',
latent_actions.cpu().data.numpy()))
logger.record_dict(
create_stats_ordered_dict(
'Latent Actions Norm',
torch.norm(latent_actions, dim=1).cpu().data.numpy()))
logger.record_dict(
create_stats_ordered_dict(
'Perturbation Norm',
torch.norm(actions - mid_actions, dim=1).cpu().data.numpy()))
logger.record_dict(
create_stats_ordered_dict('Current_Q',
current_Q1.cpu().data.numpy()))
assert (np.abs(np.mean(target_Q.cpu().data.numpy())) < 1e6)
def train(self,
replay_buffer,
iterations,
batch_size=100,
discount=0.99,
tau=0.005,
policy_noise=0.2,
noise_clip=0.5,
policy_freq=2):
for it in range(iterations):
# Sample replay buffer
x, y, u, r, d, mask = replay_buffer.sample(batch_size)
state = torch.FloatTensor(x).to(device)
action = torch.FloatTensor(u).to(device)
next_state = torch.FloatTensor(y).to(device)
done = torch.FloatTensor(1 - d).to(device)
reward = torch.FloatTensor(r).to(device)
recon, mean, std = self.vae(state, action)
recon_loss = F.mse_loss(recon, action)
KL_loss = -0.5 * (1 + torch.log(std.pow(2)) - mean.pow(2) -
std.pow(2)).mean()
vae_loss = recon_loss + 0.5 * KL_loss
self.vae_optimizer.zero_grad()
vae_loss.backward()
self.vae_optimizer.step()
# Select action according to policy and add clipped noise
noise = torch.FloatTensor(u).data.normal_(0,
policy_noise).to(device)
noise = noise.clamp(-noise_clip, noise_clip)
next_action = (self.actor_target(next_state) + noise).clamp(
-self.max_action, self.max_action)
# Compute the target Q value
target_Q1, target_Q2 = self.critic_target(next_state, next_action)
target_Q = torch.min(target_Q1, target_Q2)
target_Q = reward + (done * discount * target_Q).detach()
# Get current Q estimates
current_Q1, current_Q2 = self.critic(state, action)
std_loss = torch.std(
torch.cat([current_Q1.unsqueeze(0),
current_Q2.unsqueeze(0)], 0),
dim=0,
unbiased=False,
)
# Compute critic loss
critic_loss = F.mse_loss(current_Q1, target_Q) + F.mse_loss(
current_Q2, target_Q)
# Optimize the critic
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
# Delayed policy updates
if it % policy_freq == 0:
# Compute actor loss
actor_loss = -self.critic.Q1(state, self.actor(state)).mean()
sampled_actions = self.vae.decode(state)
actor_actions = self.actor(state)
action_divergence = ((sampled_actions -
actor_actions)**2).sum(-1)
# Optimize the actor
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# Update the frozen target models
for param, target_param in zip(
self.critic.parameters(),
self.critic_target.parameters()):
target_param.data.copy_(tau * param.data +
(1 - tau) * target_param.data)
for param, target_param in zip(self.actor.parameters(),
self.actor_target.parameters()):
target_param.data.copy_(tau * param.data +
(1 - tau) * target_param.data)
logger.record_dict(
create_stats_ordered_dict(
'Q_target',
target_Q.cpu().data.numpy(),
))
logger.record_tabular('Actor Loss', actor_loss.cpu().data.numpy())
logger.record_tabular('Critic Loss', critic_loss.cpu().data.numpy())
logger.record_tabular('Std Loss', std_loss.cpu().data.numpy().mean())
logger.record_dict(
create_stats_ordered_dict('Action_Divergence',
action_divergence.cpu().data.numpy()))
policy = algos.LatentPerturbation(vae_trainer.vae, state_dim,
action_dim, latent_dim, max_action,
**vars(args))
evaluations = []
episode_num = 0
done = True
training_iters = 0
while training_iters < args.max_timesteps:
# Train
pol_vals = policy.train(replay_buffer,
iterations=int(args.eval_freq),
batch_size=args.batch_size)
training_iters += args.eval_freq
print("Training iterations: " + str(training_iters))
logger.record_tabular('Training Epochs',
int(training_iters // int(args.eval_freq)))
# Save Model
if training_iters % args.save_freq == 0 and args.save_model:
policy.save('model_' + str(training_iters), folder_name)
# Eval
info = eval_policy(policy, env)
evaluations.append(info['AverageReturn'])
np.save(os.path.join(folder_name, 'eval'), evaluations)
eval_dict = eval_critic(policy.select_action, policy.critic.q1, env)
for k, v in eval_dict.items():
logger.record_tabular('Eval_critic/' + k, v)
for k, v in info.items():
logger.record_tabular(k, v)