def prepare_dataset_(self):
with open(self.load_path, "rb") as file:
transitions = pickle.load(file)
if self.data_limit is not None:
transitions = transitions[:self.data_limit]
self.dataset = self.transitions_to_dataset_(transitions)
vis_utils.plot_many_histograms(
[self.dataset[constants.REWARDS], self.dataset[constants.STATE_LABELS],
self.dataset[constants.NEXT_STATE_LABELS]],
["rewards_hist", "state_labels_hist", "next_state_labels_hist"], xlabel="items",
num_bins=10, saver=self.saver
)
if self.oversample:
self.dataset.oversample((self.dataset[constants.REWARDS] > 0))
self.dataset.shuffle()
self.num_samples = self.dataset.size
self.valid_samples = int(self.num_samples * self.validation_fraction)
self.valid_dataset = self.dataset.split(self.valid_samples)
self.epoch_size = self.dataset[constants.STATES].shape[0] // self.batch_size
self.log_dataset_stats_()
def prepare_dataset_(self):
with open(self.load_path, "rb") as file:
transitions = pickle.load(file)
self.dataset = prepare_minatar_dataset(transitions)
vis_utils.plot_many_histograms([
self.dataset[constants.Q_VALUES].reshape(-1),
self.dataset[constants.NEXT_Q_VALUES].reshape(-1)
], ["q_values_hist", "next_q_values_hist"],
xlabel="q-values",
num_bins=50,
saver=self.saver)
self.dataset.shuffle()
num_samples = self.dataset.size
valid_samples = int(num_samples * self.validation_fraction)
self.valid_dataset = self.dataset.split(valid_samples)
self.epoch_size = self.dataset[
constants.STATES].shape[0] // self.batch_size
self.train_not_dones = np.logical_not(self.dataset[constants.DONES])
self.valid_not_dones = np.logical_not(
self.valid_dataset[constants.DONES])
self.log_dataset_stats_()
def plot_cluster_failures_(self):
vis_utils.plot_many_histograms([self.cluster_failures.reshape(-1)],
["cluster_failure_probs"],
xlabel="bins",
num_bins=50,
saver=self.saver)
def prepare_dataset_(self):
with open(self.load_path, "rb") as file:
transitions = pickle.load(file)
self.dataset = prepare_bisim_dataset_with_q_values(
transitions, self.grid_size, self.num_pucks, gt_q_values=self.gt_q_values,
only_one_q_value=self.only_one_q_value, include_goal_state=self.include_goal_states,
shift_qs_by=self.discount if self.shift_q_values else None
)
self.dataset[constants.NEW_DONES] = self.dataset[constants.REWARDS][:, self.active_task_indices] > 0.0
vis_utils.plot_many_histograms(
[self.dataset[constants.Q_VALUES].reshape(-1), self.dataset[constants.NEXT_Q_VALUES].reshape(-1)],
["q_values_hist", "next_q_values_hist"], xlabel="q-values",
num_bins=50, saver=self.saver
)
vis_utils.plot_many_histograms(
[self.dataset[constants.STATE_LABELS], self.dataset[constants.NEXT_STATE_LABELS]],
["state_labels_hist", "next_state_labels_hist"], xlabel="label index",
num_bins=10, saver=self.saver
)
original_size = self.dataset[constants.STATES][0].shape[0]
assert len(self.dataset[constants.STATES][0].shape) == 2
assert original_size == self.dataset[constants.STATES][0].shape[1]
if original_size == self.RESIZE_SIZE:
self.disable_resize = True
self.depth_size = self.RESIZE_SIZE
if not self.disable_resize:
self.images_pl, self.resized_t = \
self.dataset.resize_tf(original_size, self.RESIZE_SIZE, [constants.STATES, constants.NEXT_STATES])
self.logger.info("original depth size: {:d}, new size: {:d}".format(original_size, self.depth_size))
if self.oversample:
self.dataset.oversample((self.dataset[constants.REWARDS] > 0))
self.means, self.stds = self.dataset.normalize_together(
[constants.STATES, constants.NEXT_STATES], std_threshold=0.001
)
self.dataset.shuffle()
num_samples = self.dataset.size
valid_samples = int(num_samples * self.validation_fraction)
self.valid_dataset = self.dataset.split(valid_samples)
self.epoch_size = self.dataset[constants.STATES].shape[0] // self.batch_size
self.train_not_dones = np.logical_not(self.dataset[constants.NEW_DONES])
self.valid_not_dones = np.logical_not(self.valid_dataset[constants.NEW_DONES])
self.log_dataset_stats_()
def plot_cluster_q_values_(self):
print(self.cluster_q_values)
vis_utils.plot_many_histograms(
[self.cluster_q_values.reshape(-1)],
["cluster_q_values"], xlabel="bins",
num_bins=50, saver=self.saver
)
def plot_perplexities_(self):
train_perplexities = evaluate.get_perplexities(self.train_cluster_log_probs)
valid_perplexities = evaluate.get_perplexities(self.valid_cluster_log_probs)
vis_utils.plot_many_histograms(
[train_perplexities, valid_perplexities],
["train_state_perplexities", "valid_state_perplexities"],
xlabel="perplexity", num_bins=50, saver=self.saver
)
def plot_cluster_rewards_(self, cluster_rewards, name):
vis_utils.plot_many_histograms([cluster_rewards.reshape(-1)], [name],
xlabel="bins",
num_bins=50,
saver=self.saver)