def train():
env = gym.make("CartPole-v1")
agent = CartPoleDRQNAgent(DRQN, model=DRQN(), env=env,
exploration=LinearSchedule(10000, initial_p=1.0, final_p=0.02),
batch_size=4, memory_size=100000, min_mem=100)
agent.saving_dir = '/home/ur5/thesis/rdd_rl/gym_test/cartpole/data/drqn'
agent.train(10000, 10000, 100, False)
def train():
env = gym.make("CartPole-v1")
agent = CartPoleDRQNAgent(DRQN, model=DRQN(), env=env,
exploration=LinearSchedule(100000, initial_p=1.0, final_p=0.1),
batch_size=32, target_update_frequency=20, memory_size=1000)
agent.saving_dir = '/home/ur5/thesis/rdd_rl/gym_test/data/drqn_cartpole_partial'
agent.train(100000, 200, 100, False)
def plot(checkpoint):
env = None
agent = CartPoleDRQNAgent(DRQN, model=DRQN(), env=env,
exploration=LinearSchedule(100000, initial_p=1.0, final_p=0.1),
batch_size=32, target_update_frequency=20)
agent.saving_dir = '/home/ur5/thesis/rdd_rl/gym_test/data/drqn_cartpole_partial'
agent.loadCheckpoint(checkpoint)
plotLearningCurve(agent.episode_rewards, window=20)
plt.show()
def train():
env = gym.make("CartPole-v1")
agent = CartPoleDRQNAgent(DQN,
model=DQN(),
env=env,
exploration=LinearSchedule(10000,
initial_p=1.0,
final_p=0.02),
batch_size=32)
agent.saving_dir = '/home/ur5/thesis/rdd_rl/gym_test/data/dqn_cartpole'
agent.train(10000, 500, 10000, False)
def train():
env = gym.make("CartPole-v1")
agent = CartPoleDRQNSliceAgent(DRQN,
model=DRQN(),
env=env,
exploration=LinearSchedule(100000,
initial_p=1.0,
final_p=0.02),
batch_size=32,
memory_size=100000,
min_mem=10000,
sequence_len=32)
agent.saving_dir = '/home/ur5/thesis/rdd_rl/gym_test/cartpole/data/partial_drqn_slice'
agent.loadCheckpoint('20190212204507')
agent.train(100000, 500, 200, False)
final_mask_batch.append(
torch.tensor(list(episode_transition.final_mask),
dtype=torch.uint8))
pad_mask_batch.append(
torch.tensor(list(episode_transition.pad_mask),
dtype=torch.uint8))
state = (torch.stack(state_1_batch).to(self.device),
torch.stack(state_2_batch).to(self.device))
action = torch.stack(action_batch).to(self.device)
next_state = (torch.stack(next_state_1_batch).to(self.device),
torch.stack(next_state_2_batch).to(self.device))
reward = torch.stack(reward_batch).to(self.device)
final_mask = torch.stack(final_mask_batch).to(self.device)
pad_mask = torch.stack(pad_mask_batch)
non_pad_mask = 1 - pad_mask
return state, action, next_state, reward, final_mask, non_pad_mask
if __name__ == '__main__':
agent = ConvDRQNAgent(DRQN,
model=DRQN(),
env=ScoopEnv(),
exploration=LinearSchedule(10000,
initial_p=1.0,
final_p=0.1),
min_mem=1000)
agent.loadCheckpoint('20190212192630')
agent.train(100000, max_episode_steps=200)
def forward(self, x, hidden=None):
x = x.float() / 256
shape = x.shape
x = x.view(shape[0] * shape[1], shape[2], shape[3], shape[4])
conv_out = self.conv(x)
x = conv_out.view(shape[0], shape[1], -1)
if hidden is None:
x, hidden = self.lstm(x)
else:
x, hidden = self.lstm(x, hidden)
x = self.fc(x)
return x, hidden
if __name__ == '__main__':
env = gym.make('PongNoFrameskip-v4')
env = wrap_drqn(env)
agent = DRQNAgent(DRQN,
model=DRQN(env.observation_space.shape,
env.action_space.n),
env=env,
exploration=LinearSchedule(100000, 0.02),
batch_size=1,
target_update_frequency=1000,
memory_size=100000,
min_mem=10000)
agent.saving_dir = '/home/ur5/thesis/rdd_rl/gym_test/pong/data/drqn'
agent.train(10000, 10000, save_freq=50)
self.episode_rewards = checkpoint['episode_rewards']
self.episode_lengths = checkpoint['episode_lengths']
self.policy_net = HistoryDQN()
self.policy_net.load_state_dict(checkpoint['policy_state_dict'])
self.policy_net = self.policy_net.to(DEVICE)
self.policy_net.train()
self.target_net = HistoryDQN()
self.target_net.load_state_dict(checkpoint['policy_state_dict'])
self.target_net = self.target_net.to(DEVICE)
self.target_net.eval()
self.optimizer = optim.Adam(self.policy_net.parameters())
self.optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
if __name__ == '__main__':
agent = SynDQNAgent(
20005,
4,
LinearSchedule(10000, 0.1),
batch_size=256,
saving_dir='/home/ur5/thesis/simple_task/scoop_grasp_2d/data/sync_dqn')
agent.load_checkpoint('20181205143445')
# agent.train(10000)
plotLearningCurve(agent.episode_rewards)
plt.show()
plotLearningCurve(agent.episode_lengths, label='length', color='r')
plt.show()
action_batch.append(torch.cat(episode_transition.action))
reward_batch.append(torch.cat(episode_transition.reward))
final_mask_batch.append(torch.tensor(list(episode_transition.final_mask), dtype=torch.uint8))
pad_mask_batch.append(torch.tensor(list(episode_transition.pad_mask), dtype=torch.uint8))
state = (torch.stack(state_0_batch).to(self.device),
torch.stack(state_1_batch).to(self.device))
action = torch.stack(action_batch).to(self.device)
next_state = (torch.stack(next_state_0_batch).to(self.device),
torch.stack(next_state_1_batch).to(self.device))
reward = torch.stack(reward_batch).to(self.device)
final_mask = torch.stack(final_mask_batch).to(self.device)
pad_mask = torch.stack(pad_mask_batch)
non_pad_mask = 1 - pad_mask
return state, action, next_state, reward, final_mask, non_pad_mask
if __name__ == '__main__':
envs = []
for i in range(1):
env = ScoopEnv(19997 + i)
envs.append(env)
agent = Agent(DRQN(envs[0].observation_space[0].shape, envs[0].observation_space[1].shape, 4),
envs, LinearSchedule(10000, 0.1), batch_size=128, min_mem=1000)
agent.saving_dir = '/home/ur5/thesis/rdd_rl/scoop_vision/data/syn_drqn_dense'
agent.loadCheckpoint('20190220221354')
agent.train(100000, 200, save_freq=500)