agent = agent_model.AgentDDPG(act_net, device=device)
exp_source = Experience.ExperienceSourceFirstLast(env, agent, gamma=gamma, steps_count=1)
buffer = Experience.ExperienceReplayBuffer(exp_source, buffer_size=replay_size)
if args.optimizer and args.optimizer == "RMSprop":
act_opt = optim.RMSprop(act_net.parameters(), lr=lr_actor)
crt_opt = optim.RMSprop(crt_net.parameters(), lr=lr_critic)
else:
act_opt = optim.Adam(act_net.parameters(), lr=lr_actor)
crt_opt = optim.Adam(crt_net.parameters(), lr=lr_critic)
utils.load_agent_state(act_net, crt_net, [act_opt, crt_opt], path=ckpt_save_path)
frame_idx = 0
drl_updates = 0
best_reward = None
with utils.RewardTracker(writer) as tracker:
with utils.TBMeanTracker(writer, batch_size=10) as tb_tracker:
while True:
frame_idx += 1
buffer.populate(1)
rewards_steps = exp_source.pop_rewards_steps()
if rewards_steps:
rewards, steps = zip(*rewards_steps)
tb_tracker.track("episode_steps", steps[0], frame_idx)
mean_reward = tracker.reward(rewards[0], frame_idx)
if mean_reward is not None and mean_reward > REWARD_TO_SOLVE:
print("environment solved in % steps" % frame_idx,
" (% episodes)" % len(tracker.total_rewards))
utils.save_agent_state(act_net, crt_net, [act_opt, crt_opt], frame_idx,
len(tracker.total_rewards), path=ckpt_save_path)
break
exp_source, buffer_size=params['replay_size'])
optimizer = optim.Adam(
net.parameters(),
lr=params['learning_rate']) # TODO: change to RMSprop
utils.load_agent_state(net,
optimizer,
selector,
load_optimizer=False,
env_name='boxing',
path='./agent_ckpt/agent_ls_dqn_-boxing.pth')
frame_idx = 0
drl_updates = 0
with utils.RewardTracker(writer, params['stop_reward']) as reward_tracker:
while True:
frame_idx += 1
buffer.populate(1)
epsilon_tracker.frame(frame_idx)
new_rewards = exp_source.pop_total_rewards()
if new_rewards:
if reward_tracker.reward(new_rewards[0], frame_idx,
selector.epsilon):
if save_for_analysis:
temp_model_name = model_name + "_" + str(frame_idx)
utils.save_agent_state(
net,
optimizer,
frame_idx,