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
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
# 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 _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
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
# (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 __init__(self,
env,
num_init_random_rollouts=10,
max_rollout_length=500,
num_onplicy_iters=10,
num_onpolicy_rollouts=10,
training_epochs=60,
training_batch_size=512,
render=False,
mpc_horizon=15,
num_random_action_selection=4096,
nn_layers=1):
self._env = env
self._max_rollout_length = max_rollout_length
self._num_onpolicy_iters = num_onplicy_iters
self._num_onpolicy_rollouts = num_onpolicy_rollouts
self._training_epochs = training_epochs
self._training_batch_size = training_batch_size
self._render = render
logger.info('Gathering random dataset')
self._random_dataset = self._gather_rollouts(utils.RandomPolicy(env),
num_init_random_rollouts)
logger.info('Creating policy')
self._policy = ModelBasedPolicy(
env,
self._random_dataset,
horizon=mpc_horizon,
num_random_action_selection=num_random_action_selection)
timeit.reset()
timeit.start('total')
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 = []
### 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
"""
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 _log(self, dataset):
timeit.stop('total')
dataset.log()
logger.dump_tabular(print_func=logger.info)
logger.debug('')
for line in str(timeit).split('\n'):
logger.debug(line)
timeit.reset()
timeit.start('total')
def _log(self, dataset):
# stop timing
timeit.stop('total')
# print logging information
dataset.log()
logger.dump_tabular(print_func=logger.info)
logger.debug('')
for line in str(timeit).split('\n'):
logger.debug(line)
# reset timing
timeit.reset()
timeit.start('total')
def _debug_rollout_and_record(self, policy, num_rollouts):
dataset = utils.Dataset()
for r_num in range(num_rollouts):
state = self._env.reset()
done = False
t = 0
states = [state]
pred_states = [state]
while not done:
if self._render:
timeit.start('render')
self._env.render()
timeit.stop('render')
timeit.start('get action')
action, next_state_pred = policy.get_action(state, True)
timeit.stop('get action')
timeit.start('env step')
next_state, reward, done, _ = self._env.step(action)
timeit.stop('env step')
done = done or (t >= self._max_rollout_length)
dataset.add(state, action, next_state, reward, done)
state = next_state
t += 1
pred_states.append(next_state_pred)
states.append(next_state)
states = np.array(states)
pred_states = np.array(pred_states)
self._debug_plot_states(states, pred_states, r_num)
return dataset
def _gather_rollouts(self, policy, num_rollouts):
dataset = utils.Dataset()
for _ in range(num_rollouts):
state = self._env.reset()
done = False
t = 0
while not done:
if self._render:
timeit.start('render')
self._env.render()
timeit.stop('render')
timeit.start('get action')
action = policy.get_action(state)
timeit.stop('get action')
timeit.start('env step')
next_state, reward, done, _ = self._env.step(action)
timeit.stop('env step')
done = done or (t >= self._max_rollout_length)
dataset.add(state, action, next_state, reward, done)
state = next_state
t += 1
return dataset
def _gather_rollouts(self, policy, num_rollouts):
# Initialize to have an empty dataset
dataset = utils.Dataset()
for i in range(num_rollouts):
state = self._env.reset()
done = False
t = 0
while not done:
if self._render:
timeit.start('render')
self._env.render()
timeit.stop('render')
timeit.start('get action')
action = policy.get_action(state)
timeit.stop('get action')
timeit.start('env step')
# Therefore, here env has the real dynamics
next_state, reward, done, _ = self._env.step(action)
timeit.stop('env step')
# max_rollout_length = 500, therefore there are typically 501 data points
done = done or (t >= self._max_rollout_length)
dataset.add(state, action, next_state, reward, done)
state = next_state
t += 1
if t % 100 == 0:
print('time step', t)
print('rollout', i)
return 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 _gather_rollouts(self, policy, num_rollouts):
# initialize dataset class
dataset = utils.Dataset()
# loop for num_rollouts
for _ in range(num_rollouts):
# reset gym env
t = 0
done = False
state = self._env.reset()
# generate gym rollout
while not done:
# perform rendering
if self._render:
timeit.start('render')
self._env.render()
timeit.stop('render')
# get action using MPC
timeit.start('get action')
action = policy.get_action(state)
timeit.stop('get action')
# step through environment
timeit.start('env step')
next_state, reward, done, _ = self._env.step(action)
timeit.stop('env step')
# add experience to dataset
done = done or (t >= self._max_rollout_length)
dataset.add(state, action, next_state, reward, done)
# update state variable
t += 1
state = next_state
return dataset