reconstruction_weight=1.0,
kl_weight=1.0,
math_weight=0.0,
train_math=False,
math_A=None,
noisy=True,
build_background_encoder=lambda scope: MLP(n_units=[10, 10], scope=scope),
build_background_decoder=IdentityFunction,
max_possible_objects=None)
simple_config = alg_config.copy(
alg_name="simple",
build_network=simple.SimpleVAE,
render_hook=simple.SimpleVAE_RenderHook(),
build_encoder=networks.Backbone,
build_decoder=networks.InverseBackbone,
n_channels=128,
n_final_layers=3,
kernel_size=1,
pixels_per_cell=(12, 12),
)
baseline_config = alg_config.copy(
alg_name="baseline",
build_network=baseline.Baseline_Network,
render_hook=baseline.Baseline_RenderHook(),
build_object_encoder=lambda scope: MLP(n_units=[512, 256], scope=scope),
build_object_decoder=lambda scope: MLP(n_units=[256, 512], scope=scope),
cc_threshold=1e-3,
object_shape=(21, 21),
)
def test_simple_add(test_config):
# Fully specify the config here so that this test is not affected by config changes external to this file.
config = Config(
env_name="test_simple_add_a2c",
name="test_simple_add_a2c",
get_updater=a2c.A2C,
n_controller_units=32,
batch_size=16,
optimizer_spec="adam",
opt_steps_per_update=20,
sub_batch_size=0,
epsilon=0.2,
lr_schedule=1e-4,
max_steps=501,
build_policy=BuildEpsilonSoftmaxPolicy(),
build_controller=BuildLstmController(),
exploration_schedule=0.1,
val_exploration_schedule=0.0,
actor_exploration_schedule=None,
policy_weight=1.0,
value_weight=0.0,
value_reg_weight=0.0,
entropy_weight=0.01,
split=False,
q_lmbda=1.0,
v_lmbda=1.0,
policy_importance_c=0,
q_importance_c=None,
v_importance_c=None,
max_grad_norm=None,
gamma=1.0,
use_differentiable_loss=False,
use_gpu=False,
display_step=500,
seed=0,
# env-specific
build_env=simple_addition.build_env,
T=30,
curriculum=[
dict(width=1),
dict(width=2),
dict(width=3),
],
base=10,
final_reward=True,
)
config.update(test_config)
n_repeats = 1 # Haven't made it completely deterministic yet, so keep it at 1.
results = defaultdict(int)
threshold = 0.15
for i in range(n_repeats):
config = config.copy()
output = _raw_run(config)
stdout = output.path_for('stdout')
result = _get_deterministic_output(stdout)
results[result] += 1
assert output.history[-1]['best_01_loss'] < threshold
if len(results) != 1:
for r in sorted(results):
print("\n" + "*" * 80)
print("The following occurred {} times:\n".format(results[r]))
print(r)
raise Exception("Results were not deterministic.")
assert len(output.config.curriculum) == 3
config.load_path = output.path_for('weights/best_of_stage_2')
assert os.path.exists(config.load_path + ".index")
assert os.path.exists(config.load_path + ".meta")
# Load one of the hypotheses, train it for a bit, make sure the accuracy is still high.
config.curriculum = [output.config.curriculum[-1]]
config = config.copy()
output = _raw_run(config)
stdout = output.path_for('stdout')
result = _get_deterministic_output(stdout)
results[result] += 1
assert output.history[-1]['best_01_loss'] < threshold
# Load one of the hypotheses, don't train it at all, make sure the accuracy is still high.
config.do_train = False
config = config.copy()
output = _raw_run(config)
stdout = output.path_for('stdout')
result = _get_deterministic_output(stdout)
results[result] += 1
assert output.history[-1]['best_01_loss'] < threshold
grid_config = env_config.copy(
env_name="nips_2018_grid",
build_env=Nips2018Grid,
# dataset params
min_chars=16,
max_chars=25,
n_patch_examples=0,
image_shape=(6 * 14, 6 * 14),
patch_shape=(14, 14),
characters=list(range(10)),
patch_size_std=0.0,
colours="white",
grid_shape=(6, 6),
spacing=(-3, -3),
random_offset_range=(15, 15),
n_distractors_per_image=0,
backgrounds="",
backgrounds_sample_every=False,
background_colours="",
background_cfg=dict(mode="colour", colour="black"),
object_shape=(14, 14),
postprocessing="",
preserve_env=False,
n_train=25000,
n_val=1e3,
eval_step=1000,
display_step=1000,
render_step=5000,
patience=1000000,
max_steps=110000,
)
split=False,
q_lmbda=1.0,
v_lmbda=1.0,
policy_importance_c=0,
q_importance_c=None,
v_importance_c=None,
max_grad_norm=None,
gamma=1.0,
use_differentiable_loss=False,
render_n_rollouts=4,
)
actor_critic_config = config.copy(
exp_name="ActorCritic",
split=True
)
ppo_config = config.copy(
exp_name="PPO",
opt_steps_per_update=10,
epsilon=0.2,
value_weight=0.0,
)
# Same config that is used in the test.
test_config = config.copy(
exp_name="TestA2C",
opt_steps_per_update=20,