def breakout_a2c():
device = torch.device('cuda:0')
seed = 1001
# Set random seed in python std lib, numpy and pytorch
set_seed(seed)
# Create 16 environments evaluated in parallel in sub processess with all usual DeepMind wrappers
# These are just helper functions for that
vec_env = SubprocVecEnvWrapper(ClassicAtariEnv('BreakoutNoFrameskip-v4'),
frame_history=4).instantiate(
parallel_envs=16, seed=seed)
# Again, use a helper to create a model
# But because model is owned by the reinforcer, model should not be accessed using this variable
# but from reinforcer.model property
model = PolicyGradientModelFactory(backbone=NatureCnnFactory(
input_width=84, input_height=84, input_channels=4)).instantiate(
action_space=vec_env.action_space)
# Reinforcer - an object managing the learning process
reinforcer = OnPolicyIterationReinforcer(
device=device,
settings=OnPolicyIterationReinforcerSettings(
discount_factor=0.99,
batch_size=256,
),
model=model,
algo=A2CPolicyGradient(entropy_coefficient=0.01,
value_coefficient=0.5,
max_grad_norm=0.5),
env_roller=StepEnvRoller(
environment=vec_env,
device=device,
number_of_steps=5,
discount_factor=0.99,
))
# Model optimizer
optimizer = optim.RMSprop(reinforcer.model.parameters(),
lr=7.0e-4,
eps=1e-3)
# Overall information store for training information
training_info = TrainingInfo(
metrics=[
EpisodeRewardMetric(
'episode_rewards'), # Calculate average reward from episode
],
callbacks=[StdoutStreaming()
] # Print live metrics every epoch to standard output
)
# A bit of training initialization bookkeeping...
training_info.initialize()
reinforcer.initialize_training(training_info)
training_info.on_train_begin()
# Let's make 100 batches per epoch to average metrics nicely
num_epochs = int(1.1e7 / (5 * 16) / 100)
# Normal handrolled training loop
for i in range(1, num_epochs + 1):
epoch_info = EpochInfo(training_info=training_info,
global_epoch_idx=i,
batches_per_epoch=100,
optimizer=optimizer)
reinforcer.train_epoch(epoch_info)
training_info.on_train_end()
for i in range(10):
a = 0 #env.action_space.sample()
for _ in range(4):
(obnew, rr, done, _info) = env.step(a)
(obnew2, rr2, done2, _info2) = env2.step(a)
obnew2 = cv2.cvtColor(
cv2.resize(obnew2, (84, 84), interpolation=cv2.INTER_LINEAR),
cv2.COLOR_BGR2GRAY)
print(obnew[:, :, 0].shape, rr, done, _info)
print(obnew2.shape, rr2, done2, _info2)
print(abs(obnew[:, :, 0] - obnew2).mean())
print(obnew.repeat(3, 2).shape)
plt.imshow(
np.concatenate((obnew.repeat(3, 2), obnew2[..., None].repeat(3,
2))))
plt.show()
exit()
env = SubprocVecEnvWrapper(ClassicAtariEnv('BreakoutNoFrameskip-v4'),
frame_history=4).instantiate(parallel_envs=8,
seed=1)
print(env.observation_space, env.action_space, 'max_episode_steps',
env.venv.specs[0])
print(env)
ob = env.reset()
print('ob', ob.shape)
a = np.array([env.action_space.sample() for _ in range(8)])
print(a)
(obnew, rr, done, _info) = env.step(a)
print(obnew.shape, rr.shape, done.shape, _info)
def test_acer_breakout():
"""
1 iteration of ACER on breakout environment
"""
device = torch.device('cpu')
seed = 1001
# Set random seed in python std lib, numpy and pytorch
set_seed(seed)
# Create 16 environments evaluated in parallel in sub processess with all usual DeepMind wrappers
# These are just helper functions for that
vec_env = SubprocVecEnvWrapper(ClassicAtariEnv('BreakoutNoFrameskip-v4'),
frame_history=4).instantiate(
parallel_envs=16, seed=seed)
# Again, use a helper to create a model
# But because model is owned by the reinforcer, model should not be accessed using this variable
# but from reinforcer.model property
model_factory = QPolicyGradientModelFactory(backbone=NatureCnnFactory(
input_width=84, input_height=84, input_channels=4))
# Reinforcer - an object managing the learning process
reinforcer = BufferedMixedPolicyIterationReinforcer(
device=device,
settings=BufferedMixedPolicyIterationReinforcerSettings(
discount_factor=0.99,
experience_replay=2,
stochastic_experience_replay=False),
model=model_factory.instantiate(action_space=vec_env.action_space),
env=vec_env,
algo=AcerPolicyGradient(
model_factory=model_factory,
entropy_coefficient=0.01,
q_coefficient=0.5,
rho_cap=10.0,
retrace_rho_cap=1.0,
trust_region=True,
trust_region_delta=1.0,
max_grad_norm=10.0,
),
env_roller=ReplayQEnvRoller(environment=vec_env,
device=device,
number_of_steps=12,
discount_factor=0.99,
buffer_capacity=100,
buffer_initial_size=100,
frame_stack_compensation=4),
)
# Model optimizer
optimizer = optim.RMSprop(reinforcer.model.parameters(),
lr=7.0e-4,
eps=1e-3,
alpha=0.99)
# Overall information store for training information
training_info = TrainingInfo(
metrics=[
EpisodeRewardMetric(
'episode_rewards'), # Calculate average reward from episode
],
callbacks=[] # Print live metrics every epoch to standard output
)
# A bit of training initialization bookkeeping...
training_info.initialize()
reinforcer.initialize_training(training_info)
training_info.on_train_begin()
# Let's make 100 batches per epoch to average metrics nicely
num_epochs = 1
# Normal handrolled training loop
for i in range(1, num_epochs + 1):
epoch_info = EpochInfo(training_info=training_info,
global_epoch_idx=i,
batches_per_epoch=1,
optimizer=optimizer)
reinforcer.train_epoch(epoch_info, interactive=False)
training_info.on_train_end()
def qbert_ppo():
device = torch.device('cuda:0')
seed = 1001
# Set random seed in python std lib, numpy and pytorch
set_seed(seed)
# Create 16 environments evaluated in parallel in sub processess with all usual DeepMind wrappers
# These are just helper functions for that
vec_env = SubprocVecEnvWrapper(ClassicAtariEnv('QbertNoFrameskip-v4'),
frame_history=4).instantiate(
parallel_envs=8, seed=seed)
# Again, use a helper to create a model
# But because model is owned by the reinforcer, model should not be accessed using this variable
# but from reinforcer.model property
model = StochasticPolicyModelFactory(
input_block=ImageToTensorFactory(),
backbone=NatureCnnFactory(
input_width=84, input_height=84,
input_channels=4)).instantiate(action_space=vec_env.action_space)
# Set schedule for gradient clipping.
cliprange = LinearSchedule(initial_value=0.1, final_value=0.0)
# Reinforcer - an object managing the learning process
reinforcer = OnPolicyIterationReinforcer(
device=device,
settings=OnPolicyIterationReinforcerSettings(batch_size=256,
experience_replay=4,
number_of_steps=128),
model=model,
algo=PpoPolicyGradient(entropy_coefficient=0.01,
value_coefficient=0.5,
max_grad_norm=0.5,
discount_factor=0.99,
gae_lambda=0.95,
cliprange=cliprange),
env_roller=StepEnvRoller(
environment=vec_env,
device=device,
))
# Model optimizer
optimizer = optim.Adam(reinforcer.model.parameters(),
lr=2.5e-4,
eps=1.0e-5)
# Overall information store for training information
training_info = TrainingInfo(
metrics=[
EpisodeRewardMetric(
'episode_rewards'), # Calculate average reward from episode
],
callbacks=[
StdoutStreaming(
), # Print live metrics every epoch to standard output
FrameTracker(
1.1e7
) # We need frame tracker to track the progress of learning
])
# A bit of training initialization bookkeeping...
training_info.initialize()
reinforcer.initialize_training(training_info)
training_info.on_train_begin()
# Let's make 10 batches per epoch to average metrics nicely
# Rollout size is 8 environments times 128 steps
num_epochs = int(1.1e7 / (128 * 8) / 10)
# Normal handrolled training loop
for i in range(1, num_epochs + 1):
epoch_info = EpochInfo(training_info=training_info,
global_epoch_idx=i,
batches_per_epoch=10,
optimizer=optimizer)
reinforcer.train_epoch(epoch_info)
training_info.on_train_end()