test_env = common.make_env(test=True)
writer = SummaryWriter(comment="-01_a2c_" + args.name + suffix)
net = common.AtariA2C(envs[0].observation_space.shape, envs[0].action_space.n).to(device)
print(net)
optimizer = optim.RMSprop(net.parameters(), lr=LEARNING_RATE, eps=1e-5)
step_idx = 0
total_steps = 0
best_reward = None
ts_start = time.time()
best_test_reward = None
with ptan.common.utils.TBMeanTracker(writer, batch_size=100) as tb_tracker:
for mb_obs, mb_rewards, mb_actions, mb_values, _, done_rewards, done_steps in \
common.iterate_batches(envs, net, device=device):
if len(done_rewards) > 0:
total_steps += sum(done_steps)
speed = total_steps / (time.time() - ts_start)
if best_reward is None:
best_reward = done_rewards.max()
elif best_reward < done_rewards.max():
best_reward = done_rewards.max()
tb_tracker.track("total_reward_max", best_reward, step_idx)
tb_tracker.track("total_reward", done_rewards, step_idx)
tb_tracker.track("total_steps", done_steps, step_idx)
print("%d: done %d episodes, mean_reward=%.2f, best_reward=%.2f, speed=%.2f" % (
step_idx, len(done_rewards), done_rewards.mean(), best_reward, speed))
common.train_a2c(net, mb_obs, mb_rewards, mb_actions, mb_values,
optimizer, tb_tracker, step_idx, device=device)
obs = envs[0].reset()
obs_v = ptan.agent.default_states_preprocessor([obs]).to(device)
res = net_i2a(obs_v)
optimizer = optim.RMSprop(net_i2a.parameters(), lr=LEARNING_RATE, eps=1e-5)
policy_opt = optim.Adam(net_policy.parameters(), lr=POLICY_LR)
step_idx = 0
total_steps = 0
ts_start = time.time()
best_reward = None
best_test_reward = None
with ptan.common.utils.TBMeanTracker(writer, batch_size=10) as tb_tracker:
for mb_obs, mb_rewards, mb_actions, mb_values, mb_probs, done_rewards, done_steps in \
common.iterate_batches(envs, net_i2a, device):
if len(done_rewards) > 0:
total_steps += sum(done_steps)
speed = total_steps / (time.time() - ts_start)
if best_reward is None:
best_reward = done_rewards.max()
elif best_reward < done_rewards.max():
best_reward = done_rewards.max()
tb_tracker.track("total_reward_max", best_reward, step_idx)
tb_tracker.track("total_reward", done_rewards, step_idx)
tb_tracker.track("total_steps", done_steps, step_idx)
print("%d: done %d episodes, mean_reward=%.2f, best_reward=%.2f, speed=%.2f f/s" % (
step_idx, len(done_rewards), done_rewards.mean(), best_reward, speed))
obs_v = common.train_a2c(net_i2a, mb_obs, mb_rewards, mb_actions, mb_values,
optimizer, tb_tracker, step_idx, device=device)
obs = envs[0].reset()
obs_v = ptan.agent.default_states_preprocessor([obs]).to(device)
res = net_i2a(obs_v)
optimizer = optim.RMSprop(net_i2a.parameters(), lr=LEARNING_RATE, eps=1e-5)
policy_opt = optim.Adam(net_policy.parameters(), lr=POLICY_LR)
step_idx = 0
total_steps = 0
ts_start = time.time()
best_reward = None
best_test_reward = None
with ptan.common.utils.TBMeanTracker(writer, batch_size=10) as tb_tracker:
for mb_obs, mb_rewards, mb_actions, mb_values, mb_probs, done_rewards, done_steps in \
common.iterate_batches(envs, net_i2a, device):
if len(done_rewards) > 0:
total_steps += sum(done_steps)
speed = total_steps / (time.time() - ts_start)
if best_reward is None:
best_reward = done_rewards.max()
elif best_reward < done_rewards.max():
best_reward = done_rewards.max()
tb_tracker.track("total_reward_max", best_reward, step_idx)
tb_tracker.track("total_reward", done_rewards, step_idx)
tb_tracker.track("total_steps", done_steps, step_idx)
print("%d: done %d episodes, mean_reward=%.2f, best_reward=%.2f, speed=%.2f f/s" % (
step_idx, len(done_rewards), done_rewards.mean(), best_reward, speed))
obs_v = common.train_a2c(net_i2a, mb_obs, mb_rewards, mb_actions, mb_values,
optimizer, tb_tracker, step_idx, device=device)
test_env = common.make_env(test=True)
writer = SummaryWriter(comment="-01_a2c_" + args.name + suffix)
net = common.AtariA2C(envs[0].observation_space.shape, envs[0].action_space.n).to(device)
print(net)
optimizer = optim.RMSprop(net.parameters(), lr=LEARNING_RATE, eps=1e-5)
step_idx = 0
total_steps = 0
best_reward = None
ts_start = time.time()
best_test_reward = None
with ptan.common.utils.TBMeanTracker(writer, batch_size=100) as tb_tracker:
for mb_obs, mb_rewards, mb_actions, mb_values, _, done_rewards, done_steps in \
common.iterate_batches(envs, net, device=device):
if len(done_rewards) > 0:
total_steps += sum(done_steps)
speed = total_steps / (time.time() - ts_start)
if best_reward is None:
best_reward = done_rewards.max()
elif best_reward < done_rewards.max():
best_reward = done_rewards.max()
tb_tracker.track("total_reward_max", best_reward, step_idx)
tb_tracker.track("total_reward", done_rewards, step_idx)
tb_tracker.track("total_steps", done_steps, step_idx)
print("%d: done %d episodes, mean_reward=%.2f, best_reward=%.2f, speed=%.2f" % (
step_idx, len(done_rewards), done_rewards.mean(), best_reward, speed))
common.train_a2c(net, mb_obs, mb_rewards, mb_actions, mb_values,
optimizer, tb_tracker, step_idx, device=device)
suffix = ""
test_env = common.make_env(test=True)
writer = SummaryWriter(comment="-01_a2c_" + args.name + suffix)
net = common.AtariA2C(envs[0].observation_space.shape, envs[0].action_space.n)
if args.cuda:
net.cuda()
print(net)
optimizer = optim.RMSprop(net.parameters(), lr=LEARNING_RATE, eps=1e-5)
step_idx = 0
best_reward = None
best_test_reward = None
with ptan.common.utils.TBMeanTracker(writer, batch_size=100) as tb_tracker:
for mb_obs, mb_rewards, mb_actions, mb_values, done_rewards, done_steps in common.iterate_batches(envs, net, cuda=args.cuda):
if len(done_rewards) > 0:
if best_reward is None:
best_reward = done_rewards.max()
elif best_reward < done_rewards.max():
best_reward = done_rewards.max()
tb_tracker.track("total_reward_max", best_reward, step_idx)
tb_tracker.track("total_reward", done_rewards, step_idx)
tb_tracker.track("total_steps", done_steps, step_idx)
print("%d: done %d episodes, mean_reward=%.2f, best_reward=%.2f" % (
step_idx, len(done_rewards), done_rewards.mean(), best_reward))
common.train_a2c(net, mb_obs, mb_rewards, mb_actions, mb_values,
optimizer, tb_tracker, step_idx, cuda=args.cuda)
step_idx += 1
if args.steps is not None and args.steps < step_idx:
obs = envs[0].reset()
obs_v = ptan.agent.default_states_preprocessor([obs], cuda=args.cuda)
res = net_i2a(obs_v)
optimizer = optim.RMSprop(net_i2a.parameters(), lr=LEARNING_RATE, eps=1e-5)
policy_opt = optim.Adam(net_policy.parameters(), lr=POLICY_LR)
step_idx = 0
total_steps = 0
ts_start = time.time()
best_reward = None
best_test_reward = None
with ptan.common.utils.TBMeanTracker(writer, batch_size=10) as tb_tracker:
for mb_obs, mb_rewards, mb_actions, mb_values, mb_probs, done_rewards, done_steps in \
common.iterate_batches(envs, net_i2a, cuda=args.cuda):
if len(done_rewards) > 0:
total_steps += sum(done_steps)
speed = total_steps / (time.time() - ts_start)
if best_reward is None:
best_reward = done_rewards.max()
elif best_reward < done_rewards.max():
best_reward = done_rewards.max()
tb_tracker.track("total_reward_max", best_reward, step_idx)
tb_tracker.track("total_reward", done_rewards, step_idx)
tb_tracker.track("total_steps", done_steps, step_idx)
print(
"%d: done %d episodes, mean_reward=%.2f, best_reward=%.2f, speed=%.2f f/s"
% (step_idx, len(done_rewards), done_rewards.mean(),
best_reward, speed))