def no_victim_finetuning():
# TODO only have one of these, since they're fully determined from each other
shape = "(64,)"
shape_split = make_tuple(shape)
if len(shape_split) == 0:
raise Exception("didn't parse shape right")
shape_target = []
shape_target.extend(shape_split)
shape_target.extend(shape_split)
adv_index = 1
victim_index = 0
gamma = 0.95
lr = 0.0065
batch_size = 2000
adversary_policy = get_pg_pol
victim_policy = get_pg_pol
adversary_trainer = partial(get_pg_train, gamma=gamma, lr=lr, batch_size=batch_size, shape=[])
victim_trainer = partial(get_pg_train, gamma=gamma, lr=lr, batch_size=batch_size, shape=[])
victim_policy_key = "baseline"
adversary_type_key = "total"
dir_loc = TRAINING_SAVE_DIR
env = "tictactoe-v0"
adversary_training_iters = 800
victim_finetuning_iters = 0
outer_loop_iters = 1
dim = 10
# Note that this currently isn't smart enough to ensure that your adversary and victim algs during training are
# compatible with the ones you've chosen here, and they need to be
prev_run = latest_training_sacred_dir(dir_loc)
victim_trainer_func = partial(possibly_pretrained_trainer, load_weights=True) # victim reloads all weights
adversary_trainer_func = partial(possibly_pretrained_trainer, load_weights=False) # adversary reloads no weights
reload_checkpoint_index = None
def paper_config_vul_tot():
shape = "(64,)"
shape_split = make_tuple(shape)
lr = 0.0001
adversary_policy=partial(get_hierarchical_pg_pol, core_policy_hiddens=(shape_split), redundant_policy_hiddens=(shape_split))
adversary_training_iters= 800
outer_loop_iters=1
adversary_type_key = "total"
def paper_config_itr_adv_train_tot():
shape = "(64,)"
shape_split = make_tuple(shape)
lr = 0.0001
adversary_policy = partial(get_hierarchical_pg_pol, core_policy_hiddens=(shape_split), redundant_policy_hiddens=(shape_split))
adversary_trainer_func = partial(possibly_pretrained_trainer, load_weights=False)
adversary_training_iters = 800
outer_loop_iters = 10
victim_finetuning_iters = 10
adversary_type_key = "iat_total"
dir_loc = None
def paper_sys_ablation():
"""
Reinitializes core weights, resets redundant weights to pretrained values, trains both (nothing frozen)
"""
shape = "(64,)"
shape_split = make_tuple(shape)
lr = 0.0001
adversary_policy = partial(get_hierarchical_pg_pol, core_policy_hiddens=(shape_split), redundant_policy_hiddens=(shape_split))
adversary_trainer_func = partial(possibly_pretrained_trainer, load_weights=True,
pretrained_weight_keywords=['core'])
adversary_type_key = "total_core_pretrained"
def paper_config_vul_str():
shape = "(64,)"
shape_split = make_tuple(shape)
lr = 0.0001
adversary_policy = partial(get_hierarchical_pg_pol, core_policy_hiddens=(shape_split), redundant_policy_hiddens=(shape_split),
redundant_trainable=False)
adversary_trainer_func = partial(possibly_pretrained_trainer, load_weights=True,
pretrained_weight_keywords=['redundant'])
adversary_training_iters = 800
outer_loop_iters = 1
adversary_type_key = "strategic"
dir_loc=None