def get_config(
files_list,
input_names=["state_1", "state_2"],
output_names=["Qvalue_1", "Qvalue_2"],
agents=2,
):
"""This is only used during training."""
expreplay = ExpReplay(
predictor_io_names=(input_names, output_names),
player=get_player(task="train", files_list=files_list, agents=agents),
state_shape=IMAGE_SIZE,
batch_size=BATCH_SIZE,
memory_size=MEMORY_SIZE,
init_memory_size=INIT_MEMORY_SIZE,
init_exploration=1.0,
update_frequency=UPDATE_FREQ,
history_len=FRAME_HISTORY,
agents=agents,
)
return TrainConfig(
# dataflow=expreplay,
data=QueueInput(expreplay),
model=Model(agents=agents),
callbacks=[
ModelSaver(),
PeriodicTrigger(
RunOp(DQNModel.update_target_param, verbose=True),
# update target network every 10k steps
every_k_steps=10000 // UPDATE_FREQ,
),
expreplay,
ScheduledHyperParamSetter("learning_rate", [(60, 4e-4),
(100, 2e-4)]),
ScheduledHyperParamSetter(
ObjAttrParam(expreplay, "exploration"),
# 1->0.1 in the first million steps
[(0, 1), (10, 0.1), (320, 0.01)],
interp="linear",
),
PeriodicTrigger(
Evaluator(
nr_eval=EVAL_EPISODE,
input_names=input_names,
output_names=output_names,
files_list=files_list,
get_player_fn=get_player,
agents=agents,
),
every_k_epochs=EPOCHS_PER_EVAL,
),
HumanHyperParamSetter("learning_rate"),
],
steps_per_epoch=STEPS_PER_EPOCH,
max_epoch=1000,
)
def get_config():
"""This is only used during training."""
expreplay = ExpReplay(predictor_io_names=(['state'], ['Qvalue']),
player=get_player(directory=data_dir,
task='train',
files_list=train_data_fpaths),
state_shape=OBSERVATION_DIMS,
batch_size=BATCH_SIZE,
memory_size=MEMORY_SIZE,
init_memory_size=INIT_MEMORY_SIZE,
init_exploration=1.0,
update_frequency=UPDATE_FREQ,
frame_history_len=FRAME_HISTORY)
return TrainConfig(
# dataflow=expreplay,
data=QueueInput(expreplay),
model=Model(),
callbacks=[ # TODO: periodically save videos
ModelSaver(checkpoint_dir="model_checkpoints",
keep_checkpoint_every_n_hours=0.25,
max_to_keep=1000), # TODO: og was just ModelSaver()
PeriodicTrigger(
RunOp(DQNModel.update_target_param, verbose=True),
# update target network every 10k/freq steps
every_k_steps=10000 // UPDATE_FREQ),
# expreplay,
ScheduledHyperParamSetter('learning_rate', [(60, 4e-4),
(100, 2e-4)]),
ScheduledHyperParamSetter(
ObjAttrParam(expreplay, 'exploration'),
# 1->0.1 in the first 10M steps
[(0, 1), (100, 0.1), (120, 0.01)],
interp='linear'),
PeriodicTrigger( # runs exprelay._trigger()
expreplay, every_k_steps=5000),
PeriodicTrigger(
# eval_model_multithread(pred, EVAL_EPISODE, get_player)
Evaluator(nr_eval=EVAL_EPISODE,
input_names=['state'],
output_names=['Qvalue'],
directory=data_dir,
files_list=test_data_fpaths,
get_player_fn=get_player),
every_k_steps=10000 // UPDATE_FREQ),
HumanHyperParamSetter('learning_rate'),
],
steps_per_epoch=STEPS_PER_EPOCH,
max_epoch=NUM_EPOCHS,
)
def get_config():
"""This is only used during training."""
expreplay = ExpReplay(predictor_io_names=(['state'], ['Qvalue']),
player=get_player(directory=data_dir,
task='train',
files_list=train_list),
state_shape=IMAGE_SIZE,
batch_size=BATCH_SIZE,
memory_size=MEMORY_SIZE,
init_memory_size=INIT_MEMORY_SIZE,
init_exploration=1.0,
update_frequency=UPDATE_FREQ,
history_len=FRAME_HISTORY)
return TrainConfig(
# dataflow=expreplay,
data=QueueInput(expreplay),
model=Model(),
callbacks=[
ModelSaver(),
PeriodicTrigger(
RunOp(DQNModel.update_target_param, verbose=True),
# update target network every 10k steps
every_k_steps=10000 // UPDATE_FREQ),
expreplay,
ScheduledHyperParamSetter('learning_rate', [(60, 4e-4),
(100, 2e-4)]),
ScheduledHyperParamSetter(
ObjAttrParam(expreplay, 'exploration'),
# 1->0.1 in the first million steps
[(0, 1), (10, 0.1), (320, 0.01)],
interp='linear'),
PeriodicTrigger(Evaluator(nr_eval=EVAL_EPISODE,
input_names=['state'],
output_names=['Qvalue'],
directory=data_dir,
files_list=test_list,
get_player_fn=get_player),
every_k_epochs=EPOCHS_PER_EVAL),
HumanHyperParamSetter('learning_rate'),
],
steps_per_epoch=STEPS_PER_EPOCH,
max_epoch=1000,
)
def get_config(files_list, data_type, trainable_variables):
"""This is only used during training."""
expreplay = ExpReplay(
predictor_io_names=(['state'], ['Qvalue']),
player=get_player(task='train',
files_list=files_list,
data_type=data_type),
state_shape=IMAGE_SIZE,
batch_size=BATCH_SIZE,
memory_size=MEMORY_SIZE,
init_memory_size=INIT_MEMORY_SIZE,
init_exploration=0.8, #0.0
###############################################################################
# HITL UPDATE
update_frequency=INIT_UPDATE_FREQ,
###############################################################################
history_len=FRAME_HISTORY,
arg_type=data_type)
return TrainConfig(
# dataflow=expreplay,
data=QueueInput(expreplay),
model=Model(IMAGE_SIZE, FRAME_HISTORY, METHOD, NUM_ACTIONS, GAMMA,
trainable_variables),
callbacks=[
ModelSaver(),
PeriodicTrigger(
RunOp(DQNModel.update_target_param, verbose=True),
# update target network every 10k steps
every_k_steps=10000 // UPDATE_FREQ),
expreplay,
ScheduledHyperParamSetter('learning_rate', [(60, 4e-4),
(100, 2e-4)]),
ScheduledHyperParamSetter(
ObjAttrParam(expreplay, 'exploration'),
# 1->0.1 in the first million steps
[(0, 0.8), (1000000, 0.1), (32000000, 0.01)],
interp='linear',
step_based=True),
###############################################################################
# HITL UPDATE
# Here the number of steps taken in the environment is increased from 0, during
# the pretraining phase, to 4 to allow the agent to take 4 steps in the env
# between each TD update.
ScheduledHyperParamSetter(ObjAttrParam(expreplay,
'update_frequency'),
[(0, INIT_UPDATE_FREQ),
(NUM_PRETRAIN, UPDATE_FREQ)],
interp=None,
step_based=True),
###############################################################################
PeriodicTrigger(Evaluator(nr_eval=EVAL_EPISODE,
input_names=['state'],
output_names=['Qvalue'],
files_list=files_list,
data_type=data_type,
get_player_fn=get_player),
every_k_steps=STEPS_PER_EVAL),
HumanHyperParamSetter('learning_rate'),
],
steps_per_epoch=STEPS_PER_EPOCH,
max_epoch=MAX_EPOCHS,
)