def learner2gif(lnr: DDPGLearner, s_format,
group_interp: GroupAgentInterpretation, name: str, extra: str):
meta = f'{lnr.memory.__class__.__name__}_{"FEED_TYPE_STATE" if s_format==FEED_TYPE_STATE else "FEED_TYPE_IMAGE"}'
interp = AgentInterpretation(lnr, ds_type=DatasetType.Train)
interp.plot_rewards(cumulative=True, per_episode=True, group_name=meta)
group_interp.add_interpretation(interp)
group_interp.to_pickle(
f'../docs_src/data/{name}_{lnr.model.name.lower()}/',
f'{lnr.model.name.lower()}_{meta}')
[g.write(f'../res/run_gifs/{name}_{extra}') for g in interp.generate_gif()]
def test_ddpg_models_ant(model_cls, s_format, experience):
group_interp = GroupAgentInterpretation()
extra_s = f'{experience.__name__}_{model_cls.__name__}_{s_format}'
for i in range(5):
print('\n')
# data=MDPDataBunch.from_env('AntPyBulletEnv-v0', render='human', bs=64, add_valid=False, keep_env_open=False,
# feed_type=s_format, memory_management_strategy='k_partitions_top', k=3)
# exploration_method=OrnsteinUhlenbeck(size=data.action.taken_action.shape, epsilon_start=1, epsilon_end=0.1,
# decay=0.00001)
# memory=experience(memory_size=1000000, reduce_ram=True)
# model=create_ddpg_model(data=data, base_arch=model_cls, lr=1e-3, actor_lr=1e-4)
# learner=ddpg_learner(data=data, model=model, memory=memory, exploration_method=exploration_method,
# opt_func=torch.optim.Adam, callback_fns=[RewardMetric, EpsilonMetric])
# learner.fit(4)
learner = trained_learner(model_cls,
'AntPyBulletEnv-v0',
s_format,
experience,
decay=0.0001,
render='rgb_array',
epochs=1000)
learner2gif(learner, s_format, group_interp, 'ant', extra_s)
# meta=f'{experience.__name__}_{"FEED_TYPE_STATE" if s_format==FEED_TYPE_STATE else "FEED_TYPE_IMAGE"}'
# interp=AgentInterpretation(learner, ds_type=DatasetType.Train)
# interp.plot_rewards(cumulative=True, per_episode=True, group_name=meta)
# group_interp.add_interpretation(interp)
# group_interp.to_pickle(f'../docs_src/data/ant_{model.name.lower()}/',
# f'{model.name.lower()}_{meta}')
# [g.write('../res/run_gifs/ant', frame_skip=5) for g in interp.generate_gif()]
del learner
def test_ddpg_models_acrobot(model_cls, s_format, experience):
group_interp = GroupAgentInterpretation()
extra_s = f'{experience.__name__}_{model_cls.__name__}_{s_format}'
for i in range(5):
print('\n')
learner = trained_learner(model_cls,
'Acrobot-v1',
s_format,
experience,
decay=0.0001,
render='rgb_array')
learner2gif(learner, s_format, group_interp, 'acrobot', extra_s)
del learner
def test_dqn_models_cartpole(model_cls, s_format, experience):
group_interp = GroupAgentInterpretation()
extra_s = f'{experience.__name__}_{model_cls.__name__}_{s_format}'
for i in range(5):
learn = trained_learner(model_cls,
'CartPole-v1',
s_format,
experience,
bs=32,
layers=[64, 64],
memory_size=1000000,
decay=0.001)
learner2gif(learn, s_format, group_interp, 'cartpole', extra_s)
def test_dqn_fit_maze_env(model_cls, s_format, mem):
group_interp = GroupAgentInterpretation()
extra_s = f'{mem.__name__}_{model_cls.__name__}_{s_format}'
for i in range(5):
learn = trained_learner(model_cls,
'maze-random-5x5-v0',
s_format,
mem,
bs=32,
layers=[32, 32],
memory_size=1000000,
decay=0.00001,
res_wrap=partial(ResolutionWrapper,
w_step=3,
h_step=3))
learner2gif(learn, s_format, group_interp, 'maze_5x5', extra_s)
def test_dqn_models_lunarlander(model_cls, s_format, experience):
group_interp = GroupAgentInterpretation()
extra_s = f'{experience.__name__}_{model_cls.__name__}_{s_format}'
for i in range(5):
learn = trained_learner(model_cls,
'LunarLander-v2',
s_format,
experience,
bs=32,
layers=[128, 64],
memory_size=1000000,
decay=0.00001,
copy_over_frequency=600,
lr=[0.001, 0.00025],
epochs=1000)
learner2gif(learn, s_format, group_interp, 'lunarlander', extra_s)
del learn
gc.collect()
def test_ddpg_models_mountain_car_continuous(model_cls, s_format, experience):
group_interp = GroupAgentInterpretation()
for i in range(5):
print('\n')
data = MDPDataBunch.from_env(
'MountainCarContinuous-v0',
render='rgb_array',
bs=40,
add_valid=False,
keep_env_open=False,
feed_type=s_format,
memory_management_strategy='k_partitions_top',
k=3,
res_wrap=partial(ResolutionWrapper, w_step=2, h_step=2))
exploration_method = OrnsteinUhlenbeck(
size=data.action.taken_action.shape,
epsilon_start=1,
epsilon_end=0.1,
decay=0.0001)
memory = experience(memory_size=1000000, reduce_ram=True)
model = create_ddpg_model(data=data, base_arch=model_cls)
learner = ddpg_learner(data=data,
model=model,
memory=memory,
exploration_method=exploration_method,
callback_fns=[RewardMetric, EpsilonMetric])
learner.fit(450)
meta = f'{experience.__name__}_{"FEED_TYPE_STATE" if s_format==FEED_TYPE_STATE else "FEED_TYPE_IMAGE"}'
interp = AgentInterpretation(learner, ds_type=DatasetType.Train)
interp.plot_rewards(cumulative=True, per_episode=True, group_name=meta)
group_interp.add_interpretation(interp)
group_interp.to_pickle(
f'../docs_src/data/mountaincarcontinuous_{model.name.lower()}/',
f'{model.name.lower()}_{meta}')
[
g.write('../res/run_gifs/mountaincarcontinuous')
for g in interp.generate_gif()
]
data.close()
del learner
del model
del data
def test_dqn_models_mountaincar(model_cls, s_format, experience):
group_interp = GroupAgentInterpretation()
for i in range(5):
learn = trained_learner(model_cls,
'MountainCar-v0',
s_format,
experience,
bs=32,
layers=[24, 12],
memory_size=1000000,
decay=0.00001,
copy_over_frequency=1000)
meta = f'{experience.__name__}_{"FEED_TYPE_STATE" if s_format == FEED_TYPE_STATE else "FEED_TYPE_IMAGE"}'
interp = AgentInterpretation(learn, ds_type=DatasetType.Train)
interp.plot_rewards(cumulative=True, per_episode=True, group_name=meta)
group_interp.add_interpretation(interp)
filename = f'{learn.model.name.lower()}_{meta}'
group_interp.to_pickle(
f'../docs_src/data/mountaincar_{learn.model.name.lower()}/',
filename)
[g.write('../res/run_gifs/mountaincar') for g in interp.generate_gif()]
del learn