def test_CNN(preprocessed_images):
batch_size = preprocessed_images.shape[0]
history_length = preprocessed_images.shape[-1]
nb_actions = 10
cnn = CNN(history_length, nb_actions)
q_values = cnn(preprocessed_images)
assert type(q_values) == torch.Tensor
assert list(q_values.shape) == [batch_size, nb_actions]
def Q():
'''
Generate a Q function
'''
agent_history_length = pytest.agent_history_length
nb_actions = pytest.nb_actions
device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')
Q = CNN(agent_history_length, nb_actions).to(device)
return Q
def Q():
'''
Generate a Q function
'''
agent_history_length = pytest.agent_history_length
nb_actions = gym.make(pytest.env_name).action_space.n
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
Q = CNN(agent_history_length, nb_actions).to(device)
return Q
def main(env_name):
AGENT_HISTORY_LENGTH = 4
env = gym.make(env_name)
env = SkipFrames(env, AGENT_HISTORY_LENGTH - 1, preprocess)
Q_network = CNN(AGENT_HISTORY_LENGTH, env.action_space.n)
train_deepq(
env=env,
env_name=env_name,
Q_network=Q_network,
input_as_images=True,
)
def test_get_training_data(replay_memory):
nb_actions, nb_timesteps, replay_memory = replay_memory
Q_hat = CNN(agent_history_length=nb_timesteps, nb_actions=nb_actions)
batch_size = 32
phi_t_training, actions_training, y = get_training_data(
Q_hat, replay_memory, batch_size, 0.99)
assert type(phi_t_training) == torch.Tensor
assert phi_t_training.shape[0] == batch_size
assert type(actions_training) == list
for action in actions_training:
assert action < nb_actions
assert action >= 0
assert len(actions_training) == batch_size
assert type(y) == torch.Tensor
assert y.shape[0] == batch_size
import gym
from deepq.wrapper_gym import KFrames
from deepq.deepq import train_deepq
from deepq.neural_nets import CNN
from deepq.utils import preprocess
AGENT_HISTORY_LENGTH = 4
NB_ACTIONS = 6
env = gym.make("PongNoFrameskip-v4")
env = KFrames(env, AGENT_HISTORY_LENGTH)
Q_network = CNN(AGENT_HISTORY_LENGTH, NB_ACTIONS)
train_deepq(
env=env,
name='Pong',
nb_actions=NB_ACTIONS,
Q_network=Q_network,
preprocess_fn=preprocess,
tensorboard_freq=5,
demo_tensorboard=True,
)