env_params.register(params)
model = Net(params.n_actions,
input_shape=(1 if grayscale else 3, im_height, im_width))
if params.cuda_enabled:
model.cuda()
loss_fn = nn.MSELoss(size_average=False)
optimizer = optim.Adam(model.parameters(),
lr=run.getfloat("learning", "lr"))
action_selector = ActionSelectorEpsilonGreedy(epsilon=run.getfloat(
"defaults", "epsilon"),
params=params)
target_net = agent.TargetNet(model)
dqn_agent = agent.DQNAgent(dqn_model=model,
action_selector=action_selector)
exp_source = experience.ExperienceSource(env=env_pool,
agent=dqn_agent,
steps_count=run.getint(
"defaults", "n_steps"))
exp_replay = experience.ExperienceReplayBuffer(exp_source,
buffer_size=run.getint(
"exp_buffer", "size"))
use_target_dqn = run.getboolean("dqn", "target_dqn", fallback=False)
use_double_dqn = run.getboolean("dqn", "double_dqn", fallback=False)
if use_target_dqn:
target_model = target_net.target_model
else:
target_model = model
model = Net(env_pool[0].action_space.n,
input_shape=(frames_count if grayscale else 3 * frames_count,
im_height, im_width),
dueling=run.getboolean("dqn", "dueling"))
if cuda_enabled:
model.cuda()
loss_fn = utils.WeightedMSELoss(size_average=True)
optimizer = optim.Adam(model.parameters(),
lr=run.getfloat("learning", "lr"))
action_selector = ptan.actions.EpsilonGreedyActionSelector(
epsilon=run.getfloat("defaults", "epsilon"))
target_net = agent.TargetNet(model)
dqn_agent = agent.DQNAgent(dqn_model=model,
action_selector=action_selector,
cuda=cuda_enabled)
exp_source = experience.ExperienceSource(env=env_pool,
agent=dqn_agent,
steps_count=run.getint(
"defaults", "n_steps"))
exp_replay = experience.ExperienceReplayBuffer(exp_source,
buffer_size=run.getint(
"exp_buffer", "size"))
# exp_replay = experience.PrioritizedReplayBuffer(exp_source, buffer_size=run.getint("exp_buffer", "size"),
# prob_alpha=run.getfloat("exp_buffer", "prio_alpha"),
# weight_beta=run.getfloat("exp_buffer", "prio_beta"))
use_target_dqn = run.getboolean("dqn", "target_dqn", fallback=False)
use_double_dqn = run.getboolean("dqn", "double_dqn", fallback=False)