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_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