def main():
env = gym.make('CartPole-v1')
# env = gym.make('CarRacing-v0')
# print("\nstate")
# state = env.reset()
# print(state.shape)
# print("\action")
# print(env.action_space)
# print(env.action_space.high)
# print(env.action_space.low)
# sys.exit()
save_ith_epoch = 1000
dir_name = "./model_saves/reinforce/try10k_expl/"
def create_video_callable(save_ith_episode):
def video_callable(episode):
return True if ((episode + 1) % save_ith_episode == 0) else False
return video_callable
kwargs = {
"directory": os.path.join(dir_name, "monitor"),
"resume": True,
"force": True,
"video_callable": create_video_callable(save_ith_epoch)
}
with gym.wrappers.Monitor(env, **kwargs) as env_monitor:
model = CartPoleModel()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
agent = DiscreteReinforce(env=env_monitor,
model=model,
optimizer=optimizer,
gamma=0.98,
exploration='proportional')
rewards = train_sessions(
num_epochs=10000,
agent=agent,
dir_name=dir_name,
save_ith_epoch=save_ith_epoch,
monitor=False,
)
env.close()
print(rewards)
plt.figure()
plt.plot(rewards)
plt.plot(np.convolve(rewards, np.ones((100, )) / 100, mode='same'))
plt.show()
def main():
env = gym.make('MountainCarContinuous-v0')
# env = gym.make('CarRacing-v0')
# print("\nstate")
# state = env.reset()
# print(state.shape)
# print("\action")
# print(env.action_space)
# print(env.action_space.high)
# print(env.action_space.low)
# sys.exit()
save_ith_epoch = 100
dir_name = "./model_saves/reinforce/try5/"
kwargs = {
"directory": os.path.join(dir_name, "monitor"),
"resume": False,
"force": True,
"video_callable": create_video_callable(save_ith_epoch)
}
with gym.wrappers.Monitor(env, **kwargs) as env_monitor:
model = MouintainCarContinousModel()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
agent = ContinousReinforce(
env=env_monitor,
model=model,
optimizer=optimizer,
gamma=0.98,
distribution="Normal",
decoder=None,
)
rewards = train_sessions(
num_epochs=10000,
agent=agent,
dir_name=dir_name,
save_ith_epoch=save_ith_epoch,
monitor=False,
)
env.close()
print(rewards)
plt.figure()
plt.plot(rewards)
plt.plot(np.convolve(rewards, np.ones((50, )) / 50, mode='same'))
plt.show()
def main():
env = gym.make('Pendulum-v0')
# print("\nstate")
# state = env.reset()
# print(state.shape)
# print("\action")
# print(env.action_space)
# print(env.action_space.high)
# print(env.action_space.low)
# sys.exit()
save_ith_epoch = 500
dir_name = "./model_saves/actorcritic/try5k/"
model = torch.load(os.path.join(dir_name, "best_model"))
# model = PendulumModel()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
kwargs = {"directory": os.path.join(dir_name, "monitor"), "resume": True, "force": True, "video_callable": create_video_callable(save_ith_epoch)}
with gym.wrappers.Monitor(env, **kwargs) as env_monitor:
agent = ContinuousActorCritic(
env=env_monitor,
model=model,
optimizer=optimizer,
gamma=0.98,
distribution="Normal",
decoder=None,
)
rewards = train_sessions(
num_epochs=5000,
agent=agent,
dir_name=dir_name,
save_ith_epoch=save_ith_epoch,
monitor=False,
init_expl=100,
start_epoch=10000,
)
env.close()
print(rewards)
plt.figure()
plt.plot(rewards)
plt.plot(np.convolve(rewards, np.ones((50,))/50, mode='same'))
plt.show()
def main():
num_epochs = 20
save_ith_epoch = 2
batch_size = 20
num_workers = None
representation_dim = 16
module_dir_vae = "./model_saves/simpleCNNVAE_startGas_nroll100_lroll100"
# module_vae = instaniate_SimpleCNNVAE(representation_dim, image_channels=3)
# criterion = vae_loss_fn
# optimizer_vae = torch.optim.Adam(module_vae.parameters(), lr=1e-3)
# train_observations = load_observations("./observations/startGas/observations_startGas_nroll100_lroll100_000000.npy")[:, ::10, :, :, :]
# train_observations = transform_observations_torch(train_observations)
# val_observations = load_observations("./observations/startGas/observations_startGas_nroll100_lroll100_000001.npy")[:, ::20, :, :, :]
# val_observations = transform_observations_torch(val_observations)
# train_vae(
# num_epochs = num_epochs,
# module = module_vae,
# module_dir = module_dir_vae,
# train_observations = train_observations,
# val_observations = val_observations,
# criterion = criterion,
# optimizer = optimizer_vae,
# save_ith_epoch = save_ith_epoch,
# batch_size = batch_size,
# num_workers = num_workers,
# )
module_vae = torch.load(os.path.join(module_dir_vae, "best_module"))
env = gym.make('CarRacing-v0')
# env.reset()
# env.reset()
# print(env.action_space.high)
# print(env.action_space.low)
# sys.exit()
encoder = module_vae.vae_encoder
decoder = module_vae.decoder
model = set_up_repr_dualhead_from_pixels(encoder, encoder_out_dim=representation_dim, out_dim=3)
optimizer = torch.optim.Adam([{"params": model.parameters()}, {"params": decoder.parameters()}], lr=1e-3)
save_ith_epoch = 2
dir_name = "./model_saves/reinforce/try0/"
kwargs = {"directory": os.path.join(dir_name, "monitor"), "resume": True, "force": True, "video_callable": create_video_callable(save_ith_epoch)}
with gym.wrappers.Monitor(env, **kwargs) as env_monitor:
agent = ContinousReinforce(
env=env_monitor,
model=model,
optimizer=optimizer,
gamma=0.98,
distribution="Normal",
decoder=decoder,
from_pixel=True,
convert_to_action_space=convert_to_action_space,
convert_from_action_space=convert_from_action_space,
)
rewards = train_sessions(
num_epochs=20,
agent=agent,
dir_name=dir_name,
save_ith_epoch=1,
monitor=False,
)
env.close()
print(rewards)
plt.figure()
plt.plot(rewards)
plt.plot(np.convolve(rewards, np.ones((50,))/50, mode='same'))
plt.show()