def write_tboard(self, data):
with lock:
loss, train_step, writer, summary_ops, summary_vars, num_agents = data
# Tensorboard
episode_b_rewards = []
for j in range(self.num_env):
if self.args.env_type == "mpe":
episode_b_rewards.append(
np.mean(self.ep_rewards[j][self.print_step:]))
else:
episode_b_rewards.append(
np.mean(self.ep_success[j][self.print_step:]))
episode_b_rewards = np.mean(np.array(episode_b_rewards))
num_steps = train_step * self.num_env
# Add to tensorboard only when actor agent is updated
if loss[0][1] is not None:
fd = {}
for i, key in enumerate(summary_vars):
if i == 0:
fd[key] = episode_b_rewards
else:
agnt_idx = int((i - 1) / 5)
if agnt_idx == num_agents: agnt_idx -= 1
if loss[agnt_idx] is not None:
fd[key] = loss[agnt_idx][int((i - 1) % 5)]
summary_str = U.get_session().run(summary_ops, feed_dict=fd)
writer.add_summary(summary_str, num_steps)
writer.flush()
def create_tensorboard(self):
""" Setup tensorboard """
if not (self.args.benchmark or self.args.display):
self.writer = tf.compat.v1.summary.FileWriter(
self.tboard_dir,
U.get_session().graph)
self.writer.flush()
self.summary_ops, self.summary_vars = build_summaries(
self.num_agents, self.args)