def testSetFromMap(self):
hparams = hparam.HParams(a=1, b=2.0, c='tanh')
hparams.override_from_dict({'a': -2, 'c': 'identity'})
self.assertDictEqual({'a': -2, 'c': 'identity', 'b': 2.0}, hparams.values())
hparams = hparam.HParams(x=1, b=2.0, d=[0.5])
hparams.override_from_dict({'d': [0.1, 0.2, 0.3]})
self.assertDictEqual({'d': [0.1, 0.2, 0.3], 'x': 1, 'b': 2.0},
hparams.values())
def resnet_weight():
hp = hparam.HParams()
hp.add_hparam("strategy", "weight")
hp.add_hparam("black_list", ["logits", "bias"])
hp.add_hparam("white_list", ["td_conv"])
hp.add_hparam("sparsities", [0.1 * i for i in range(10)])
return hp
def testLossSingleWeights(self):
"""Ensure _loss_single() respects optional 'weights' argument."""
with tf.Graph().as_default():
with self.test_session() as sess:
batch_size = 2
sequence_size = 16
vocab_size = 3
model_hparams = hparam.HParams(
prepend_mode="none",
loss={},
weights_fn={},
label_smoothing=0.0,
shared_embedding_and_softmax_weights=False)
ph = problem_hparams.TestProblem(
vocab_size, vocab_size).get_hparams(model_hparams)
model = t2t_model.T2TModel(model_hparams, problem_hparams=ph)
logits = tf.zeros((batch_size, sequence_size, 1, 1, vocab_size))
feature = tf.ones((batch_size, sequence_size, 1, 1))
# all-zero weights == zero loss.
weights = tf.zeros((batch_size, sequence_size))
loss_num, loss_denom = model._loss_single(
logits, "targets", feature, weights=weights)
self.assertAllClose(tf.zeros_like(loss_num), sess.run(loss_num))
self.assertAllClose(tf.zeros_like(loss_denom), sess.run(loss_denom))
# non-zero weights > zero loss.
weights = tf.ones((batch_size, sequence_size))
loss_num, loss_denom = model._loss_single(
logits, "targets", feature, weights=weights)
self.assertAllLess(0.0, sess.run(loss_num))
self.assertAllClose(batch_size * sequence_size, sess.run(loss_denom))
def testLists(self):
hparams = hparam.HParams(aaa=[1], b=[2.0, 3.0], c_c=['relu6'])
self.assertDictEqual({
'aaa': [1],
'b': [2.0, 3.0],
'c_c': ['relu6']
}, hparams.values())
self.assertEqual([1], hparams.aaa)
self.assertEqual([2.0, 3.0], hparams.b)
self.assertEqual(['relu6'], hparams.c_c)
hparams.parse('aaa=[12]')
self.assertEqual([12], hparams.aaa)
hparams.parse('aaa=[12,34,56]')
self.assertEqual([12, 34, 56], hparams.aaa)
hparams.parse('c_c=[relu4,relu12],b=[1.0]')
self.assertEqual(['relu4', 'relu12'], hparams.c_c)
self.assertEqual([1.0], hparams.b)
hparams.parse('c_c=[],aaa=[-34]')
self.assertEqual([-34], hparams.aaa)
self.assertEqual([], hparams.c_c)
hparams.parse('c_c=[_12,3\'4"],aaa=[+3]')
self.assertEqual([3], hparams.aaa)
self.assertEqual(['_12', '3\'4"'], hparams.c_c)
with self.assertRaisesRegexp(ValueError, 'Unknown hyperparameter'):
hparams.parse('x=[123]')
with self.assertRaisesRegexp(ValueError, 'Could not parse'):
hparams.parse('aaa=[poipoi]')
with self.assertRaisesRegexp(ValueError, 'Could not parse'):
hparams.parse('aaa=[1.0]')
with self.assertRaisesRegexp(ValueError, 'Could not parse'):
hparams.parse('b=[12x]')
with self.assertRaisesRegexp(ValueError, 'Could not parse'):
hparams.parse('b=[relu]')
with self.assertRaisesRegexp(ValueError, 'Must pass a list'):
hparams.parse('aaa=123')
def shake_shake_fgsm():
aparams = hparam.HParams()
aparams.attack = "fgsm"
aparams.attack_epsilons = [(i + 1) * 0.1 for i in range(12)]
aparams.add_hparam("clip_min", 0.0)
aparams.add_hparam("clip_max", 255.0)
return aparams
def testEmpty(self):
hparams = hparam.HParams()
self.assertDictEqual({}, hparams.values())
hparams.parse('')
self.assertDictEqual({}, hparams.values())
with self.assertRaisesRegexp(ValueError, 'Unknown hyperparameter'):
hparams.parse('xyz=123')
def default_model_hparams():
return hparam.HParams(
max_input_seq_length=0,
max_target_seq_length=0,
prepend_mode="none",
split_to_length=0,
data_dir=None)
def create_hparams_from_json(json_path, hparams=None):
"""Loading hparams from json; can also start from hparams if specified."""
tf.logging.info("Loading hparams from existing json %s" % json_path)
with tf.gfile.Open(json_path, "r") as f:
hparams_values = json.load(f)
# Prevent certain keys from overwriting the passed-in hparams.
# TODO(trandustin): Remove this hack after registries are available to avoid
# saving them as functions.
if hparams:
hparams_values.pop("bottom", None)
hparams_values.pop("loss", None)
hparams_values.pop("name", None)
hparams_values.pop("top", None)
hparams_values.pop("weights_fn", None)
new_hparams = hparam.HParams(**hparams_values)
# Some keys are in new_hparams but not hparams, so we need to be more
# careful than simply using parse_json() from HParams
if hparams: # hparams specified, so update values from json
for key in sorted(new_hparams.values().keys()):
if hasattr(hparams, key): # Overlapped keys
value = getattr(hparams, key)
new_value = getattr(new_hparams, key)
if value != new_value: # Different values
tf.logging.info("Overwrite key %s: %s -> %s" % (
key, value, new_value))
setattr(hparams, key, new_value)
else:
hparams = new_hparams
return hparams
def _default_hparams():
"""A set of basic model hyperparameters."""
return hparam.HParams(
# Use this parameter to get comparable perplexity numbers with different
# tokenizations. This value should be set to the ratio of the number of
# tokens in the test set according to the tokenization used to the number
# of tokens in the test set in the "official" tokenization. For
# example, if we are using a word-piece based model and we want to
# compute per-word perplexity, then we set loss_multiplier to the number
# of wordpieces per word in the test set.
loss_multiplier=1.0,
# Use this parameter to allow for larger sequences in the batch. Without
# the use of this parameter, the size of the inner two dimensions will
# be used to judge the sequence length.
batch_size_multiplier=1,
# During inference for autoregressive problems, if the batch_size is 1,
# the inference will stop when the model predict a text_encoder.EOS_ID
# token.
stop_at_eos=False,
# Modalities used to map from features to a space compatible with
# chosen model architecture. It comprises key-value pairs of a feature
# name (str) and its modality type.
modality={},
vocab_size={},
# Identifiers used to tell the model which input/target space will be
# expected. For example, it can tell that we expect French as characters
# as output, or Spanish as sound. Spaces defined as constants in SpaceID
# class.
input_space_id=SpaceID.GENERIC,
target_space_id=SpaceID.GENERIC)
def rlmf_original():
return hparam.HParams(
game="pong",
sticky_actions=False,
base_algo="ppo",
base_algo_params="ppo_original_params",
batch_size=16,
eval_batch_size=2,
frame_stack_size=4,
eval_sampling_temps=[0.0, 0.2, 0.5, 0.8, 1.0, 2.0],
max_num_noops=8,
eval_max_num_noops=8,
eval_rl_env_max_episode_steps=1000,
resize_height_factor=2,
resize_width_factor=2,
distributional_size=1, # In distributional RL, number of buckets.
distributional_subscale=0.04, # How to scale values to buckets.
distributional_threshold=0.0, # Optimism threshold for experiments.
grayscale=0,
rl_env_max_episode_steps=-1,
# If set, use this as the gym env name, instead of changing game mode etc.
rl_env_name="",
# Controls whether we should derive observation space, do some
# pre-processing etc. See T2TGymEnv._derive_observation_space.
rl_should_derive_observation_space=True,
aunused=0, # unused param for multi-run settings.
)
def dqn_atari_base():
# These params are based on agents/dqn/configs/dqn.gin
# with some modifications taking into account our code
return hparam.HParams(
agent_gamma=0.99,
agent_update_horizon=1,
agent_min_replay_history=20000, # agent steps
agent_update_period=4,
agent_target_update_period=8000, # agent steps
agent_epsilon_train=0.01,
agent_epsilon_eval=0.001,
agent_epsilon_decay_period=250000, # agent steps
agent_generates_trainable_dones=True,
agent_type="VanillaDQN", # one of ["Rainbow", "VanillaDQN"]
optimizer_class="RMSProp",
optimizer_learning_rate=0.00025,
optimizer_decay=0.95,
optimizer_momentum=0.0,
optimizer_epsilon=0.00001,
optimizer_centered=True,
# TODO(kozak): change names maybe replay_buffer -> agent?
# Also batch_size is now buffer_batch_size in _DQNAgent.
replay_buffer_replay_capacity=1000000,
replay_buffer_buffer_batch_size=32,
time_limit=27000,
save_every_steps=50000,
num_frames=int(20 * 1e6),
# TODO(konradczechowski) this is not used in trainer_model_free, clean
# this up after evaluation refactor
eval_episodes_num=3,
)
def testBoolParsing(self):
for value in 'true', 'false', 'True', 'False', '1', '0':
for initial in False, True:
hparams = hparam.HParams(use_gpu=initial)
hparams.parse('use_gpu=' + value)
self.assertEqual(hparams.use_gpu, value
in ['True', 'true', '1'])
def create_simulated_env(
output_dir, grayscale, resize_width_factor, resize_height_factor,
frame_stack_size, generative_model, generative_model_params,
random_starts=True, which_epoch_data="last", **other_hparams
):
""""Create SimulatedEnv with minimal subset of hparams."""
# We need these, to initialize T2TGymEnv, but these values (hopefully) have
# no effect on player.
a_bit_risky_defaults = {
"game": "pong", # assumes that T2TGymEnv has always reward_range (-1,1)
"real_batch_size": 1,
"rl_env_max_episode_steps": -1,
"max_num_noops": 0
}
for key in a_bit_risky_defaults:
if key not in other_hparams:
other_hparams[key] = a_bit_risky_defaults[key]
hparams = hparam.HParams(
grayscale=grayscale,
resize_width_factor=resize_width_factor,
resize_height_factor=resize_height_factor,
frame_stack_size=frame_stack_size,
generative_model=generative_model,
generative_model_params=generative_model_params,
**other_hparams
)
return load_data_and_make_simulated_env(
output_dir, wm_dir=None, hparams=hparams,
which_epoch_data=which_epoch_data,
random_starts=random_starts)
def testCreateOutputTrainMode(self, likelihood, num_mixtures, depth):
batch = 1
height = 8
width = 8
channels = 3
rows = height
if likelihood == common_image_attention.DistributionType.CAT:
cols = channels * width
else:
cols = width
hparams = hparam.HParams(
hidden_size=2,
likelihood=likelihood,
num_channels=channels,
mode=tf_estimator.ModeKeys.TRAIN,
num_mixtures=num_mixtures,
)
decoder_output = tf.random_normal(
[batch, rows, cols, hparams.hidden_size])
targets = tf.random_uniform([batch, height, width, channels],
minval=-1.,
maxval=1.)
output = common_image_attention.create_output(decoder_output, rows,
cols, targets, hparams)
if hparams.likelihood == common_image_attention.DistributionType.CAT:
self.assertEqual(output.shape,
(batch, height, width, channels, depth))
else:
self.assertEqual(output.shape, (batch, height, width, depth))
def testSetHParamTypeMismatch(self):
hparams = hparam.HParams(
int_=1, str_='str', bool_=True, float_=1.1, list_int=[1, 2], none=None)
with self.assertRaises(ValueError):
hparams.set_hparam('str_', 2.2)
with self.assertRaises(ValueError):
hparams.set_hparam('int_', False)
with self.assertRaises(ValueError):
hparams.set_hparam('bool_', 1)
with self.assertRaises(ValueError):
hparams.set_hparam('int_', 2.2)
with self.assertRaises(ValueError):
hparams.set_hparam('list_int', [2, 3.3])
with self.assertRaises(ValueError):
hparams.set_hparam('int_', '2')
# Casting int to float is OK
hparams.set_hparam('float_', 1)
# Getting stuck with NoneType :(
hparams.set_hparam('none', '1')
self.assertEqual('1', hparams.none)
def testSetHParam(self):
hparams = hparam.HParams(aaa=1, b=2.0, c_c='relu6', d=True)
self.assertDictEqual({
'aaa': 1,
'b': 2.0,
'c_c': 'relu6',
'd': True
}, hparams.values())
self.assertEqual(1, hparams.aaa)
self.assertEqual(2.0, hparams.b)
self.assertEqual('relu6', hparams.c_c)
hparams.set_hparam('aaa', 12)
hparams.set_hparam('b', 3.0)
hparams.set_hparam('c_c', 'relu4')
hparams.set_hparam('d', False)
self.assertDictEqual({
'aaa': 12,
'b': 3.0,
'c_c': 'relu4',
'd': False
}, hparams.values())
self.assertEqual(12, hparams.aaa)
self.assertEqual(3.0, hparams.b)
self.assertEqual('relu4', hparams.c_c)
def testFunction(self):
def f(x):
return x
hparams = hparam.HParams(function=f)
self.assertEqual(hparams.function, f)
json_str = hparams.to_json()
self.assertEqual(json_str, '{}')
def resnet_fgsm():
aparams = hparam.HParams()
aparams.attack = "fgsm"
aparams.epsilon_name = "eps"
aparams.attack_epsilons = [i * 0.8 for i in range(20)]
aparams.add_hparam("clip_min", 0.0)
aparams.add_hparam("clip_max", 255.0)
return aparams
def decode_hparams(overrides=""):
"""Hyperparameters for decoding."""
hp = hparam.HParams(
decode_mode=0,
save_images=False,
log_results=True,
extra_length=100,
min_length_ratio=0.0,
batch_size=0,
beam_size=4,
alpha=0.6,
eos_penalty=0.0,
block_size=0,
guess_and_check_top_k=0,
guess_and_check_epsilon=-1,
insertion_parallel=False,
return_beams=False,
write_beam_scores=False,
max_input_size=-1,
identity_output=False,
num_samples=-1, # Number of examples to decode.
delimiter="\n",
decode_to_file="", # str. Prefix for filename to write decodings to.
decode_reference="", # str. Filename to read references from.
decode_in_memory=False,
# How much decode should wait for the next checkpoint
decode_timeout_mins=240,
summaries_log_dir="decode", # Directory to write hook summaries.
shards=1, # How many shards of data to decode (treating 1 as None).
shard_id=0, # Which shard are we decoding if more than 1 above.
shards_start_offset=0, # Number of the first shard to decode.
shard_google_format=False, # If True use Google shard naming format.
num_decodes=1, # Number of times to go over the dataset.
force_decode_length=False,
display_decoded_images=False,
# Multi-problem decoding task id.
multiproblem_task_id=-1,
# Used for video decoding.
frames_per_second=10,
skip_eos_postprocess=False,
# Creates a blue/red border covering border_percent of the frame.
border_percent=2,
# Maximum number of videos displayed.
# number of videos displayed = max_display_outputs * max_display_decodes
max_display_outputs=10,
max_display_decodes=5,
# Used in computation of VGG feature based video metrics.
# Set this to be the path to a trained VGG ckpt to output
# useful metrics.
vgg_ckpt_path="",
# Used for MLPerf compliance logging.
mlperf_decode_step=0.0,
mlperf_threshold=25.0,
mlperf_success=False,
# A comma-delimited list of additional infer() outputs to be exported.
export_extra_infer_outputs="")
hp.parse(overrides)
return hp
def start_tf():
import problems
global problem
global request_fn
problem = registry.problem("translate_many_to_many")
hparams = hparam.HParams(data_dir=DATA_DIR) # ./vocab.txt
problem.get_hparams(hparams)
request_fn = make_request_fn()
print("TF initialized OK")
def merge_unscoped_hparams(scopes_and_hparams):
"""Merge multiple HParams into one with scopes."""
merged_values = {}
for (scope, hparams) in scopes_and_hparams:
for key, value in six.iteritems(hparams.values()):
scoped_key = "%s.%s" % (scope, key)
merged_values[scoped_key] = value
return hparam.HParams(**merged_values)
def copy_hparams(hparams):
hp_vals = hparams.values()
new_hparams = hparam.HParams(**hp_vals)
other_attrs = ["problem", "problem_hparams"]
for attr in other_attrs:
attr_val = getattr(hparams, attr, None)
if attr_val is not None:
setattr(new_hparams, attr, attr_val)
return new_hparams
def main_setup():
#try:
global request_fn
global problem_hp
usr_dir.import_usr_dir(t2t_usr_dir)
problem_hp = registry.problem(problem)
hparams = hparam.HParams(data_dir=os.path.expanduser(data_dir))
problem_hp.get_hparams(hparams)
request_fn = make_request_fn()
def split_scoped_hparams(scopes, merged_hparams):
"""Split single HParams with scoped keys into multiple."""
split_values = {scope: {} for scope in scopes}
merged_values = merged_hparams.values()
for scoped_key, value in six.iteritems(merged_values):
scope = scoped_key.split(".")[0]
key = scoped_key[len(scope) + 1:]
split_values[scope][key] = value
return [hparam.HParams(**split_values[scope]) for scope in scopes]
def planner_tiny():
return hparam.HParams(
num_rollouts=1,
planning_horizon=2,
rollout_agent_type="random",
batch_size=1,
env_type="simulated",
uct_const=0.0,
uniform_first_action=True,
)
def planner_base():
return hparam.HParams(
num_rollouts=96,
batch_size=96,
planning_horizon=8,
rollout_agent_type="policy",
env_type="simulated",
uct_const=0.,
uniform_first_action=True,
)
def planner_small():
return hparam.HParams(
num_rollouts=64,
planning_horizon=16,
rollout_agent_type="policy",
batch_size=64,
env_type="simulated",
uct_const=0.0,
uniform_first_action=True,
)
def testSummarizeLosses(self):
with tf.Graph().as_default():
model = t2t_model.T2TModel(hparam.HParams())
losses = {"training": tf.random_normal([]),
"extra": tf.random_normal([])}
outputs = model._summarize_losses(losses)
self.assertIsNone(outputs, None)
self.assertEqual(
len(tf.get_collection(tf.GraphKeys.SUMMARIES, scope="losses")),
len(losses))
def testImagenetMultiResolutionPreprocessExample(self, resize_method):
example = {"inputs": tf.random_uniform([64, 64, 3], minval=-1.)}
mode = tf.estimator.ModeKeys.TRAIN
hparams = hparam.HParams(resolutions=[8, 16, 32])
if resize_method is not None:
hparams.resize_method = resize_method
problem = imagenet.ImageImagenetMultiResolutionGen()
preprocessed_example = problem.preprocess_example(example, mode, hparams)
self.assertLen(preprocessed_example, 1)
self.assertEqual(preprocessed_example["inputs"].shape, (42, 32, 3))
def main_setup():
#try:
global request_fn
global problem_hp
validate_flags()
usr_dir.import_usr_dir(FLAGS.t2t_usr_dir)
problem_hp = registry.problem(FLAGS.problem)
#print(FLAGS.data_dir, '<---')
hparams = hparam.HParams(data_dir=os.path.expanduser(FLAGS.data_dir))
problem_hp.get_hparams(hparams)
request_fn = make_request_fn()
