def _initialize(self):
# Initializes the non-serialized bits of the class.
self._set_random_seed(self.config.seed)
# state for prediction caching
self._predictions = None
self._cached_predict = False
self._closed = False
self._to_pull = 0
try:
self.estimator_dir = os.path.abspath(
os.path.join(self.config.tensorboard_folder,
str(int(time.time()))))
pathlib.Path(self.estimator_dir).mkdir(parents=True, exist_ok=True)
self._tmp_dir = None
except (TypeError, IOError):
# TypeError --> tensorboard_folder is None
# IOError --> user likely does not have permission to write to the tensorboard_folder directory
# Both cases we can resolve by
self._tmp_dir = tempfile.TemporaryDirectory(prefix="Finetune")
self.estimator_dir = self._tmp_dir.name
LOGGER.info("Saving tensorboard output to {}".format(
self.estimator_dir))
self.saver = Saver(
fallback_filename=self.config.base_model_path,
exclude_matches=None if self.config.save_adam_vars else "Adam",
variable_transforms=[
embedding_preprocessor(self.input_pipeline, self.config)
],
save_dtype=self.config.save_dtype,
target_model_init_from_base_model=self.config.
target_model_init_from_base_model)
def load(path, *args, **kwargs):
"""
Load a saved fine-tuned model from disk. Path provided should be a folder which contains .pkl and tf.Saver() files
:param path: string path name to load model from. Same value as previously provided to :meth:`save`. Must be a folder.
:param **kwargs: key-value pairs of config items to override.
"""
if type(path) != str and not hasattr(path, "write"):
instance = path
raise FinetuneError(
'The .load() method can only be called on the class, not on an instance. Try `{}.load("{}") instead.'
.format(instance.__class__.__name__, args[0]))
assert_valid_config(**kwargs)
saver = Saver()
model = saver.load(path)
# Backwards compatability
# Ensure old models get new default settings
for setting, default in get_default_config().items():
if not hasattr(model.config, setting):
if setting == "add_eos_bos_to_chunk":
model.config.add_eos_bos_to_chunk = False
else:
model.config.update({setting: default})
model.config.update(kwargs)
model.input_pipeline.config = model.config
download_data_if_required(model.config.base_model)
saver.set_fallback(model.config.base_model_path)
model._initialize()
model.saver.variables = saver.variables
model._trained = True
return model
def _initialize(self):
# Initializes the non-serialized bits of the class.
self._set_random_seed(self.config.seed)
download_data_if_required()
self.encoder = TextEncoder()
# symbolic ops
self.logits = None # classification logits
self.target_loss = None # cross-entropy loss
self.lm_losses = None # language modeling losses
self.lm_predict_op = None
self.train = None # gradient + parameter update
self.features = None # hidden representation fed to classifier
self.summaries = None # Tensorboard summaries
self.train_writer = None
self.valid_writer = None
self.predict_params = None
self.train_op = None
self.predict_op = None
self.predict_proba_op = None
self.sess = None
self.noop = tf.no_op()
# indicator vars
self.is_built = False # has tf graph been constructed?
self.is_trained = False # has model been fine-tuned?
self.require_lm = False
def process_embeddings(name, value):
if "/we:0" not in name:
return value
vocab_size = self.encoder.vocab_size
word_embeddings = value[:vocab_size -
len(self.encoder.special_tokens)]
special_embed = value[len(word_embeddings):vocab_size]
positional_embed = value[vocab_size:]
if self.config.interpolate_pos_embed and self.config.max_length != len(
positional_embed):
positional_embed = interpolate_pos_embed(
positional_embed, self.config.max_length)
elif self.config.max_length > len(positional_embed):
raise ValueError(
"Max Length cannot be greater than {} if interploate_pos_embed is turned off"
.format(len(positional_embed)))
else:
positional_embed = positional_embed[:self.config.max_length]
embeddings = np.concatenate(
(word_embeddings, special_embed, positional_embed), axis=0)
return embeddings
self.saver = Saver(
fallback_filename=JL_BASE,
exclude_matches=None if self.config.save_adam_vars else "adam",
variable_transforms=[process_embeddings])
def load(cls, path):
"""
Load a saved fine-tuned model from disk. Path provided should be a folder which contains .pkl and tf.Saver() files
:param path: string path name to load model from. Same value as previously provided to :meth:`save`. Must be a folder.
"""
saver = Saver(JL_BASE)
model = saver.load(path)
model._initialize()
model.saver.variables = saver.variables
return model
def load(cls, path):
"""
Load a saved fine-tuned model from disk. Path provided should be a folder which contains .pkl and tf.Saver() files
:param path: string path name to load model from. Same value as previously provided to :meth:`save`. Must be a folder.
"""
saver = Saver(JL_BASE)
model = saver.load(path)
model._initialize()
model.saver.variables = saver.variables
tf.reset_default_graph()
return model
def load(cls, path, **kwargs):
"""
Load a saved fine-tuned model from disk. Path provided should be a folder which contains .pkl and tf.Saver() files
:param path: string path name to load model from. Same value as previously provided to :meth:`save`. Must be a folder.
:param **kwargs: key-value pairs of config items to override.
"""
assert_valid_config(**kwargs)
download_data_if_required()
saver = Saver()
model = saver.load(path)
model.config.update(kwargs)
saver.set_fallback(model.config.base_model_path)
model._initialize()
model.saver.variables = saver.variables
return model
def _initialize(self):
# Initializes the non-serialized bits of the class.
self._set_random_seed(self.config.seed)
self.estimator_ = None
if self.config.tensorboard_folder is not None:
self.estimator_dir = os.path.abspath(
os.path.join(self.config.tensorboard_folder,
str(int(time.time()))))
pathlib.Path(self.estimator_dir).mkdir(parents=True, exist_ok=True)
self.cleanup_glob = None
else:
self.estimator_dir = tempfile.mkdtemp(prefix="Finetune")
self.cleanup_glob = self.estimator_dir
def process_embeddings(name, value):
if "/we:0" not in name:
return value
vocab_size = ENCODER.vocab_size
word_embeddings = value[:vocab_size - len(ENCODER.special_tokens)]
special_embed = value[len(word_embeddings):vocab_size]
positional_embed = value[vocab_size:]
if self.config.interpolate_pos_embed and self.config.max_length != len(
positional_embed):
positional_embed = interpolate_pos_embed(
positional_embed, self.config.max_length)
elif self.config.max_length > len(positional_embed):
raise ValueError(
"Max Length cannot be greater than {} if interploate_pos_embed is turned off"
.format(len(positional_embed)))
else:
positional_embed = positional_embed[:self.config.max_length]
embeddings = np.concatenate(
(word_embeddings, special_embed, positional_embed), axis=0)
return embeddings
base_model_path = os.path.join(os.path.dirname(__file__), "model",
self.config.base_model_path)
self.saver = Saver(
fallback_filename=base_model_path,
exclude_matches=None if self.config.save_adam_vars else "Adam",
variable_transforms=[process_embeddings],
save_dtype=self.config.save_dtype)
def _initialize(self):
# Initializes the non-serialized bits of the class.
self._set_random_seed(self.config.seed)
download_data_if_required()
self.encoder = TextEncoder()
# symbolic ops
self.logits = None # classification logits
self.target_loss = None # cross-entropy loss
self.lm_losses = None # language modeling losses
self.lm_predict_op = None
self.train = None # gradient + parameter update
self.features = None # hidden representation fed to classifier
self.summaries = None # Tensorboard summaries
self.train_writer = None
self.valid_writer = None
self.predict_params = None
self.train_op = None
self.predict_op = None
self.predict_proba_op = None
self.sess = None
self.noop = tf.no_op()
# indicator vars
self.is_built = False # has tf graph been constructed?
self.is_trained = False # has model been fine-tuned?
self.require_lm = False
def process_embeddings(name, value):
if "/we:0" not in name:
return value
vocab_size = self.encoder.vocab_size
word_embeddings = value[:vocab_size - len(self.encoder.special_tokens)]
special_embed = value[len(word_embeddings): vocab_size]
positional_embed = value[vocab_size:]
if self.config.interpolate_pos_embed and self.config.max_length != len(positional_embed):
positional_embed = interpolate_pos_embed(positional_embed, self.config.max_length)
elif self.config.max_length > len(positional_embed):
raise ValueError("Max Length cannot be greater than {} if interploate_pos_embed is turned off".format(len(positional_embed)))
else:
positional_embed = positional_embed[:self.config.max_length]
embeddings = np.concatenate((word_embeddings, special_embed, positional_embed), axis=0)
return embeddings
self.saver = Saver(
fallback_filename=JL_BASE,
exclude_matches=None if self.config.save_adam_vars else "adam",
variable_transforms=[process_embeddings]
)
def load(path, *args, **kwargs):
"""
Load a saved fine-tuned model from disk. Path provided should be a folder which contains .pkl and tf.Saver() files
:param path: string path name to load model from. Same value as previously provided to :meth:`save`. Must be a folder.
:param **kwargs: key-value pairs of config items to override.
"""
if type(path) != str:
instance = path
raise FinetuneError(
"The .load() method can only be called on the class, not on an instance. Try `{}.load(\"{}\") instead.".format(
instance.__class__.__name__, args[0]
)
)
assert_valid_config(**kwargs)
saver = Saver()
model = saver.load(path)
model.config.update(kwargs)
download_data_if_required(model.config.base_model)
saver.set_fallback(model.config.base_model_path)
model._initialize()
model.saver.variables = saver.variables
return model
class BaseModel(object, metaclass=ABCMeta):
"""
A sklearn-style task agnostic base class for finetuning a Transformer language model.
"""
def __init__(self, config=None, **kwargs):
"""
For a full list of configuration options, see `finetune.config`.
:param config: A config object generated by `finetune.config.get_config` or None (for default config).
:param **kwargs: key-value pairs of config items to override.
"""
weak_self = weakref.ref(self)
def cleanup():
strong_self = weak_self()
if strong_self is not None:
BaseModel.__del__(strong_self)
atexit.register(cleanup)
self.config = config or get_default_config()
self.config.update(kwargs)
if self.config.num_layers_trained != self.config.n_layer and self.config.train_embeddings:
raise ValueError(
"If you are only finetuning a subset of the layers, you cannot finetune embeddings."
)
self.input_pipeline = self._get_input_pipeline()
download_data_if_required()
self._initialize()
@abstractmethod
def _get_input_pipeline(self):
pass
def _initialize(self):
# Initializes the non-serialized bits of the class.
self._set_random_seed(self.config.seed)
self.estimator_ = None
if self.config.tensorboard_folder is not None:
self.estimator_dir = os.path.abspath(
os.path.join(self.config.tensorboard_folder,
str(int(time.time()))))
pathlib.Path(self.estimator_dir).mkdir(parents=True, exist_ok=True)
self.cleanup_glob = None
else:
self.estimator_dir = tempfile.mkdtemp(prefix="Finetune")
self.cleanup_glob = self.estimator_dir
def process_embeddings(name, value):
if "/we:0" not in name:
return value
vocab_size = ENCODER.vocab_size
word_embeddings = value[:vocab_size - len(ENCODER.special_tokens)]
special_embed = value[len(word_embeddings):vocab_size]
positional_embed = value[vocab_size:]
if self.config.interpolate_pos_embed and self.config.max_length != len(
positional_embed):
positional_embed = interpolate_pos_embed(
positional_embed, self.config.max_length)
elif self.config.max_length > len(positional_embed):
raise ValueError(
"Max Length cannot be greater than {} if interploate_pos_embed is turned off"
.format(len(positional_embed)))
else:
positional_embed = positional_embed[:self.config.max_length]
embeddings = np.concatenate(
(word_embeddings, special_embed, positional_embed), axis=0)
return embeddings
base_model_path = os.path.join(os.path.dirname(__file__), "model",
self.config.base_model_path)
self.saver = Saver(
fallback_filename=base_model_path,
exclude_matches=None if self.config.save_adam_vars else "Adam",
variable_transforms=[process_embeddings],
save_dtype=self.config.save_dtype)
@abstractmethod
def _predict_op(self, logits, **kwargs):
raise NotImplementedError
@abstractmethod
def _predict_proba_op(self, logits, **kwargs):
raise NotImplementedError
@abstractmethod
def _target_model(self,
*,
featurizer_state,
targets,
n_outputs,
train=False,
reuse=None,
**kwargs):
# Overridden by subclass to attach a target model onto the shared base featurizer.
raise NotImplementedError
def _n_steps(self, n_examples, batch_size, n_gpus):
steps = int(math.ceil(n_examples / (batch_size * n_gpus)))
return steps
def finetune(self, Xs, Y=None, batch_size=None):
if not callable(Xs) and Y is not None and len(Xs) != len(Y):
raise FinetuneError(
"Mismatch between number of examples ({}) and number of targets ({}) provided."
.format(len(Xs), len(Y)))
batch_size = batch_size or self.config.batch_size
val_input_fn, train_input_fn, val_size, val_interval = self.input_pipeline.get_train_input_fns(
Xs, Y, batch_size=batch_size)
if val_size <= 10 and self.config.keep_best_model:
tf.logging.warning(
"Early stopping / keeping best model with a validation size of {} is likely to case undesired results"
.format(val_size))
steps_per_epoch = self._n_steps(n_examples=self.config.dataset_size,
batch_size=batch_size,
n_gpus=max(
1, len(self.config.visible_gpus)))
num_steps = steps_per_epoch * self.config.n_epochs
estimator = self.get_estimator()
train_hooks = [
self.saver.get_saver_hook(
estimator=estimator,
keep_best_model=self.config.keep_best_model,
steps_per_epoch=steps_per_epoch,
early_stopping_steps=self.config.early_stopping_steps,
eval_frequency=val_interval)
]
if val_size > 0:
train_hooks.append(
tf.contrib.estimator.InMemoryEvaluatorHook(
estimator,
val_input_fn,
every_n_iter=val_interval,
steps=val_size // batch_size))
with warnings.catch_warnings():
warnings.simplefilter("ignore")
estimator.train(train_input_fn, hooks=train_hooks, steps=num_steps)
def get_estimator(self, force_build_lm=False):
conf = tf.ConfigProto(
allow_soft_placement=self.config.soft_device_placement,
log_device_placement=self.config.log_device_placement,
)
num_gpus = len(self.config.visible_gpus)
if num_gpus > 1:
distribute_strategy = PatchedParameterServerStrategy(
num_gpus_per_worker=num_gpus)
else:
distribute_strategy = None
config = tf.estimator.RunConfig(
tf_random_seed=self.config.seed,
save_summary_steps=self.config.val_interval,
save_checkpoints_secs=None,
save_checkpoints_steps=None,
# disable auto summaries
session_config=conf,
log_step_count_steps=100,
train_distribute=distribute_strategy,
keep_checkpoint_max=1)
model_fn = get_model_fn(
target_model_fn=self._target_model,
predict_op=self._predict_op,
predict_proba_op=self._predict_proba_op,
build_target_model=self.input_pipeline.target_dim is not None,
build_lm=force_build_lm or self.config.lm_loss_coef > 0.0
or self.input_pipeline.target_dim is None,
encoder=ENCODER,
target_dim=self.input_pipeline.target_dim,
label_encoder=self.input_pipeline.label_encoder,
saver=self.saver)
return tf.estimator.Estimator(model_dir=self.estimator_dir,
model_fn=model_fn,
config=config,
params=self.config)
def _inference(self, Xs, mode=None):
estimator = self.get_estimator()
input_func = self.input_pipeline.get_predict_input_fn(Xs)
length = len(Xs) if not callable(Xs) else None
pred_gen = list(
map(
lambda y: y[mode] if mode else y,
tqdm.tqdm(estimator.predict(input_fn=input_func,
predict_keys=mode),
total=length,
desc="Inference")))
return pred_gen
def fit(self, *args, **kwargs):
""" An alias for finetune. """
return self.finetune(*args, **kwargs)
def _predict(self, Xs):
raw_preds = self._inference(Xs, PredictMode.NORMAL)
return self.input_pipeline.label_encoder.inverse_transform(
np.asarray(raw_preds))
def predict(self, Xs):
return self._predict(Xs)
def _predict_proba(self, Xs):
"""
Produce raw numeric outputs for proba predictions
"""
raw_preds = self._inference(Xs, PredictMode.PROBAS)
return raw_preds
def predict_proba(self, *args, **kwargs):
"""
The base method for predicting from the model.
"""
raw_probas = self._predict_proba(*args, **kwargs)
classes = self.input_pipeline.label_encoder.classes_
formatted_predictions = []
for probas in raw_probas:
formatted_predictions.append(dict(zip(classes, probas)))
return formatted_predictions
def _featurize(self, Xs):
raw_preds = self._inference(Xs, PredictMode.FEATURIZE)
return np.asarray(raw_preds)
@abstractmethod
def featurize(self, *args, **kwargs):
"""
Base method to get raw features out of the model.
These features are the same that are fed into the target_model.
"""
return self._featurize(*args, **kwargs)
@classmethod
def get_eval_fn(cls):
raise NotImplementedError(
"No default eval function is given, please pass an explicit eval fn to grid_search"
)
def transform(self, *args, **kwargs):
"""
An alias for `featurize`.
"""
return self.featurize(*args, **kwargs)
def _set_random_seed(self, seed=None):
seed = seed or self.config.seed
random.seed(seed)
np.random.seed(seed)
tf.set_random_seed(seed)
def generate_text(self,
seed_text='',
max_length=None,
use_extra_toks=True):
"""
Performs a prediction on the Language modeling objective given some seed text. It uses a noisy greedy decoding.
Temperature parameter for decoding is set in the config.
:param max_length: The maximum length to decode to.
:param seed_text: Defaults to the empty string. This will form the starting point to begin modelling
:return: A string containing the generated text.
"""
def dataset_encoded():
while not dataset_encoded.finished:
yield {
"tokens": arr_encoded.token_ids,
"mask": arr_encoded.mask
}
dataset_encoded.finished = False
def get_input_fn():
types, shapes = self.input_pipeline.feed_shape_type_def()
tf_dataset = Dataset.from_generator(dataset_encoded, types[0],
shapes[0])
return tf_dataset.batch(1)
self.config.use_extra_toks = use_extra_toks
encoded = ENCODER._encode([seed_text])
if encoded == [] and not use_extra_toks:
raise ValueError(
"If you are not using the extra tokens, you must provide some non-empty seed text"
)
start = [ENCODER.start] if use_extra_toks else []
encoded = EncodedOutput(token_ids=start + encoded.token_ids[0])
estimator = self.get_estimator(force_build_lm=True)
predict = estimator.predict(input_fn=get_input_fn, )
EOS = ENCODER.clf_token
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
for i in range(len(encoded.token_ids),
(max_length or self.config.max_length) - 2):
arr_encoded = self.input_pipeline._array_format(encoded)
class_idx = next(predict)[PredictMode.GENERATE_TEXT]
encoded.token_ids.append(class_idx[i])
if encoded.token_ids[-1] == EOS:
break
dataset_encoded.finished = True
del self.config["use_extra_toks"]
return ENCODER.decode(encoded.token_ids)
def __getstate__(self):
"""
Leave serialization of all tf objects to tf
"""
required_fields = ['_load_from_file', 'config', 'input_pipeline']
serialized_state = {
k: v
for k, v in self.__dict__.items() if k in required_fields
}
return serialized_state
def save(self, path):
"""
Saves the state of the model to disk to the folder specific by `path`. If `path` does not exist, it will be auto-created.
Save is performed in two steps:
- Serialize tf graph to disk using tf.Saver
- Serialize python model using pickle
Note:
Does not serialize state of Adam optimizer.
Should not be used to save / restore a training model.
"""
if path is None:
return
path = os.path.abspath(path)
self.saver.save(self, path)
@classmethod
def load(cls, path):
"""
Load a saved fine-tuned model from disk. Path provided should be a folder which contains .pkl and tf.Saver() files
:param path: string path name to load model from. Same value as previously provided to :meth:`save`. Must be a folder.
"""
saver = Saver(JL_BASE)
model = saver.load(path)
model._initialize()
model.saver.variables = saver.variables
return model
@classmethod
def finetune_grid_search(cls,
Xs,
Y,
*,
test_size,
config=None,
eval_fn=None,
probs=False,
return_all=False):
"""
Performs grid search over config items defined using "GridSearchable" objects and returns either full results or
the config object that relates to the best results. The default config contains grid searchable objects for the
most important parameters to search over.
:param Xs: Input text. Either [num_samples] or [sequence, num_samples] for single or multi input models respectively.
:param Y: Targets, A list of targets, [num_samples] that correspond to each sample in Xs.
:param test_size: Int or float. If an int is given this number of samples is used to validate, if a float is
given then that fraction of samples is used.
:param config: A config object, or None to use the default config.
:param eval_fn: An eval function that takes 2 inputs (prediction, truth) and returns a float, with a max value being desired.
:param probs: If true, eval_fn is passed probability outputs from predict_proba, otherwise the output of predict is used.
:param return_all: If True, all results are returned, if False, only the best config is returned.
:return: default is to return the best config object. If return_all is true, it returns a list of tuples of the
form [(config, eval_fn output), ... ]
"""
if isinstance(Xs[0], str):
Xs = [Xs]
config = config or get_default_config()
config.val_size = 0.0
eval_fn = eval_fn or cls.get_eval_fn()
trainXs, testXs, trainY, testY = train_test_split(list_transpose(Xs),
Y,
test_size=test_size,
shuffle=True)
trainXs = list_transpose(trainXs)
testXs = list_transpose(testXs)
gs = config.get_grid_searchable()
ranged_keys = gs.keys()
ranged_iterators = gs.values()
grid_gen = itertools.product(*ranged_iterators)
results = []
for grid_item in grid_gen:
config_ = deepcopy(config)
config_.update(dict(zip(ranged_keys, grid_item)))
instance = cls(config=config_)
instance.finetune(*trainXs, Y=trainY)
if probs:
res = instance.predict_proba(*testXs)
else:
res = instance.predict(*testXs)
results.append((config_, eval_fn(res, testY)))
del instance
if return_all:
return results
return max(results, key=lambda x: x[1])[0]
@classmethod
def finetune_grid_search_cv(cls,
Xs,
Y,
*,
n_splits,
test_size,
config=None,
eval_fn=None,
probs=False,
return_all=False):
"""
Performs cross validated grid search over config items defined using "GridSearchable" objects and returns either full results or
the config object that relates to the best results. The default config contains grid searchable objects for the
most important parameters to search over.
It should be noted that the cv splits are not guaranteed unique, but each split is given to each set of hparams.
:param Xs: Input text. Either [num_samples] or [sequence, num_samples] for single or multi input models respectively.
:param Y: Targets, A list of targets, [num_samples] that correspond to each sample in Xs.
:param n_splits: Number of CV splits to do.
:param test_size: Int or float. If an int is given this number of samples is used to validate, if a float is
given then that fraction of samples is used.
:param config: A config object, or None to use the default config.
:param eval_fn: An eval function that takes 2 batches of outputs and returns a float, with a max value being
desired. An arithmetic mean must make sense for this metric.
:param probs: If true, eval_fn is passed probability outputs from predict_proba, otherwise the output of predict is used.
:param return_all: If True, all results are returned, if False, only the best config is returned.
:return: default is to return the best config object. If return_all is true, it returns a list of tuples of the
form [(config, eval_fn output), ... ]
"""
results = []
for _ in range(n_splits):
res = cls.finetune_grid_search(Xs,
Y,
test_size=test_size,
probs=probs,
eval_fn=eval_fn,
config=config,
return_all=True)
results.append(res)
results = list(zip(*results))
aggregated_results = []
for configuration in results:
config_common = None
sum_res = 0
n_res = 0
for config, result in configuration:
config_common = config_common or config
assert config == config_common
n_res += 1
sum_res += result
aggregated_results.append((config_common, sum_res / n_res))
if return_all:
return aggregated_results
return max(aggregated_results, key=lambda x: x[1])[0]
def __del__(self):
if hasattr(self, 'cleanup_glob') and self.cleanup_glob is not None:
for file_or_folder in glob.glob(self.cleanup_glob):
try:
shutil.rmtree(file_or_folder)
except NotADirectoryError:
os.remove(file_or_folder)
class BaseModel(object, metaclass=ABCMeta):
"""
A sklearn-style task agnostic base class for finetuning a Transformer language model.
"""
defaults = dict()
def __init__(self, **kwargs):
"""
For a full list of configuration options, see `finetune.config`.
:param config: A config object generated by `finetune.config.get_config` or None (for default config).
:param **kwargs: key-value pairs of config items to override.
"""
weak_self = weakref.ref(self)
def cleanup():
strong_self = weak_self()
if strong_self is not None:
BaseModel.__del__(strong_self)
atexit.register(cleanup)
d = deepcopy(self.defaults)
d.update(kwargs)
self.config = get_config(**d)
self.resolved_gpus = None
self.validate_config()
download_data_if_required(self.config.base_model)
self.input_pipeline = self._get_input_pipeline()
self._trained = False
self._initialize()
if self.config.debugging_logs:
os.environ["TF_CPP_MIN_LOG_LEVEL"] = "0"
tf_logging.set_verbosity(tf_logging.DEBUG)
def validate_config(self):
if (self.config.num_layers_trained != self.config.n_layer
and self.config.train_embeddings):
raise ValueError(
"If you are only finetuning a subset of the layers, you cannot finetune embeddings."
)
@abstractmethod
def _get_input_pipeline(self):
pass
def _initialize(self):
# Initializes the non-serialized bits of the class.
self._set_random_seed(self.config.seed)
# state for prediction caching
self._predictions = None
self._cached_predict = False
self._closed = False
self._to_pull = 0
try:
self.estimator_dir = os.path.abspath(
os.path.join(self.config.tensorboard_folder,
str(int(time.time()))))
pathlib.Path(self.estimator_dir).mkdir(parents=True, exist_ok=True)
self._tmp_dir = None
except (TypeError, IOError):
# TypeError --> tensorboard_folder is None
# IOError --> user likely does not have permission to write to the tensorboard_folder directory
# Both cases we can resolve by
self._tmp_dir = tempfile.TemporaryDirectory(prefix="Finetune")
self.estimator_dir = self._tmp_dir.name
LOGGER.info("Saving tensorboard output to {}".format(
self.estimator_dir))
self.saver = Saver(
fallback_filename=self.config.base_model_path,
exclude_matches=None if self.config.save_adam_vars else "Adam",
variable_transforms=[
embedding_preprocessor(self.input_pipeline, self.config)
],
save_dtype=self.config.save_dtype,
target_model_init_from_base_model=self.config.
target_model_init_from_base_model)
@abstractmethod
def _predict_op(self, logits, **kwargs):
raise NotImplementedError
@abstractmethod
def _predict_proba_op(self, logits, **kwargs):
raise NotImplementedError
@abstractmethod
def _target_model(self,
*,
config,
featurizer_state,
targets,
n_outputs,
train=False,
reuse=None,
**kwargs):
# Overridden by subclass to attach a target model onto the shared base featurizer.
raise NotImplementedError
def _pre_target_model_hook(self, featurizer_state):
add_context_embed(featurizer_state)
def _n_steps(self, n_examples, batch_size, n_gpus):
steps = int(math.ceil(n_examples / (batch_size * n_gpus)))
return steps
def finetune(self,
Xs,
Y=None,
batch_size=None,
context=None,
update_hook=None):
if (not callable(Xs) and Y is not None and len(Xs) != len(Y)):
raise FinetuneError(
"Mismatch between number of examples ({}) and number of targets ({}) provided."
.format(len(Xs), len(Y)))
batch_size = batch_size or self.config.batch_size
val_input_fn, train_input_fn, val_size, val_interval = self.input_pipeline.get_train_input_fns(
Xs,
Y,
batch_size=batch_size,
context=context,
update_hook=update_hook)
if self.config.keep_best_model:
if isinstance(val_size, dict):
tf.logging.warning(
"Cannot early stop or keep best model with MTL")
elif val_size <= 10:
tf.logging.warning(
"Early stopping / keeping best model with a validation size of {} is likely to case undesired results"
.format(val_size))
force_build_lm = Y is None
estimator, hooks = self.get_estimator(force_build_lm=force_build_lm)
train_hooks = hooks.copy()
steps_per_epoch = self._n_steps(
n_examples=self.input_pipeline.dataset_size,
batch_size=batch_size,
n_gpus=max(1, len(self.resolved_gpus)),
)
num_steps = steps_per_epoch * self.config.n_epochs
if self.config.tasks is not None:
# Validation with MTL tasks
for task in self.config.tasks:
if val_size[task] > 0:
train_hooks.append(
tf.estimator.experimental.InMemoryEvaluatorHook(
estimator,
val_input_fn[task],
every_n_iter=val_interval[task],
steps=val_size[task] // batch_size,
name=task,
))
train_hooks.append(
tf.estimator.experimental.InMemoryEvaluatorHook(
estimator,
val_input_fn[task + "_train"],
every_n_iter=val_interval[task],
steps=val_size[task] // batch_size,
name=task + "_train",
))
early_stopping_interval = sys.maxsize # turn off early stopping for mtl.
elif val_size > 0:
# Validation with all other tasks.
train_hooks.append(
tf.estimator.experimental.InMemoryEvaluatorHook(
estimator,
val_input_fn,
every_n_iter=val_interval,
steps=math.ceil(val_size / batch_size),
))
early_stopping_interval = val_interval
else:
early_stopping_interval = sys.maxsize
train_hooks.append(
self.saver.get_saver_hook(
estimator=estimator,
keep_best_model=self.config.keep_best_model,
steps_per_epoch=steps_per_epoch,
early_stopping_steps=self.config.early_stopping_steps,
eval_frequency=early_stopping_interval,
cache_weights_to_file=self.config.cache_weights_to_file))
if self.config.in_memory_finetune is not None:
train_hooks.extend(make_in_memory_finetune_hooks(self, estimator))
with warnings.catch_warnings():
warnings.simplefilter("ignore")
if self.config.prefit_init:
tf.logging.info("Starting pre-fit initialisation...")
num_layers_trained = self.config.num_layers_trained
self.config.num_layers_trained = 0
estimator.train(train_input_fn,
hooks=train_hooks,
steps=num_steps)
self.config.num_layers_trained = num_layers_trained
self.saver.variables = {
k: v
for k, v in self.saver.variables.items()
if "adam" not in k and "global_step" not in k
}
for weight in self.saver.variables:
if (weight.startswith("model/target/")):
w = self.saver.variables[weight]
if len(w.shape) == 1:
continue
w_flat = np.reshape(w, [-1, w.shape[-1]])
expectation_of_norm = ((self.config.weight_stddev**2) *
w_flat.shape[0])**0.5
self.saver.variables[weight] = np.reshape(
expectation_of_norm * w_flat /
np.linalg.norm(w_flat, axis=0),
w.shape,
)
tf.logging.info("Finishing pre-fit initialisation...")
estimator.train(train_input_fn, hooks=train_hooks, steps=num_steps)
self._trained = True
def _distribute_strategy(self, visible_gpus):
"""
Select a distribution strategy based on available devices.
Side effect: sets self.resolved_gpus for future use in computing steps per epoch
"""
if isinstance(visible_gpus, (list, tuple)):
resolved_gpus = all_gpus(visible_gpus=tuple(visible_gpus))
else:
resolved_gpus = all_gpus()
resolved_gpus_string = ['/gpu:{}'.format(gpu) for gpu in resolved_gpus]
if len(resolved_gpus_string) == 1:
distribute_strategy = tf.contrib.distribute.OneDeviceStrategy(
resolved_gpus_string[0])
else:
if self.config.per_process_gpu_memory_fraction is not None:
warnings.warn(
"Setting `per_process_gpu_memory_fraction` is currently unsupported in multi-gpu environments."
)
if isinstance(self.config.distribution_strategy, str):
if self.config.distribution_strategy.lower() == "mirrored":
distribute_strategy = tf.distribute.MirroredStrategy()
elif self.config.distribution_strategy.lower(
) == "central_storage":
distribute_strategy = tf.distribute.experimental.CentralStorageStrategy(
resolved_gpus_string or None)
else:
raise FinetuneError(
"Distribute strategy {} is not supported, please try \"mirrored\" or \"central_storage\" or an instance of tf.distribute.Strategy"
)
elif isinstance(self.config.distribution_strategy,
tf.distribute.Strategy):
distribute_strategy = self.config.distribution_strategy
self.resolved_gpus = resolved_gpus
return distribute_strategy
def _get_estimator_config(self):
conf = tf.ConfigProto(
allow_soft_placement=self.config.soft_device_placement,
log_device_placement=self.config.log_device_placement,
)
if self.config.per_process_gpu_memory_fraction is not None:
conf.gpu_options.per_process_gpu_memory_fraction = (
self.config.per_process_gpu_memory_fraction)
optimizer_options = conf.graph_options.optimizer_options
if self.config.xla:
optimizer_options.global_jit_level = tf.OptimizerOptions.ON_1
distribute_strategy = self._distribute_strategy(
self.config.visible_gpus)
config = tf.estimator.RunConfig(
tf_random_seed=self.config.seed,
save_summary_steps=self.config.val_interval,
save_checkpoints_secs=None,
save_checkpoints_steps=None,
# disable auto summaries
session_config=conf,
log_step_count_steps=100,
train_distribute=distribute_strategy,
keep_checkpoint_max=1,
)
return config
def get_estimator(self, force_build_lm=False, build_explain=False):
build_lm = force_build_lm or self.config.lm_loss_coef > 0.0
config = self._get_estimator_config()
model_fn = get_model_fn(
target_model_fn=self._target_model,
pre_target_model_hook=self._pre_target_model_hook,
predict_op=self._predict_op,
predict_proba_op=self._predict_proba_op,
build_target_model=self.input_pipeline.target_dim is not None,
lm_type=self.config.lm_type if build_lm else None,
encoder=self.input_pipeline.text_encoder,
target_dim=self.input_pipeline.target_dim,
label_encoder=self.input_pipeline.label_encoder,
build_explain=build_explain,
n_replicas=max(1, len(self.resolved_gpus)))
hooks = [InitializeHook(self.saver)]
est = tf.estimator.Estimator(
model_dir=self.estimator_dir,
model_fn=model_fn,
config=config,
params=self.config,
)
return est, hooks
def close(self):
self._closed = True
if self._predictions is not None:
# force input fn termination
try:
for _ in self._predictions:
pass
except AttributeError:
pass
self._predictions = None
def _clear_prediction_queue(self):
# Flush examples used to pad the last batch
# of previous call to predict()
for i in range(self._to_pull):
next(self._predictions)
# Reset counter
self._to_pull = 0
def _data_generator(self):
self._cached_example = None
self._to_pull = 0
while not self._closed:
try:
example = self._data.pop(0)
# Ensure examples used for padding match expected input format
if isinstance(example, str):
self._cached_example = ""
elif isinstance(example, (list, tuple)):
self._cached_example = [""] * len(example)
yield example
except IndexError:
# _data_generator was asked for more examples than we had
# Feed a cached example through the input_pipeline
# to fill out the batch, but remember to clear it
# out of the queue later
self._to_pull += 1
yield self._cached_example
@contextmanager
def cached_predict(self):
"""
Context manager that prevents the recreation of the tensorflow graph on every call to BaseModel.predict().
"""
self._cached_predict = True
yield self
self._cached_predict = False
self.close()
def _cached_inference(self,
Xs,
predict_keys=None,
n_examples=None,
update_hook=None):
"""
Ensure graph is not rebuilt on subsequent calls to .predict()
"""
self._data = Xs
self._closed = False
n = n_examples or len(self._data)
if self._predictions is None:
input_fn = self.input_pipeline.get_predict_input_fn(
self._data_generator)
_estimator, hooks = self.get_estimator()
self._predictions = _estimator.predict(input_fn=input_fn,
predict_keys=predict_keys,
hooks=hooks)
self._clear_prediction_queue()
predictions = [None] * n
for i in ProgressBar(range(n),
total=n,
desc="Inference",
update_hook=update_hook):
y = next(self._predictions)
try:
y = y[predict_keys[0]] if len(predict_keys) == 1 else y
except ValueError:
raise FinetuneError(
"Cannot call `predict()` on a model that has not been fit."
)
predictions[i] = y
return predictions
def _inference(self,
Xs,
predict_keys=None,
n_examples=None,
context=None,
update_hook=None):
Xs = self.input_pipeline._format_for_inference(Xs)
if self._cached_predict:
return self._cached_inference(Xs=Xs,
predict_keys=predict_keys,
n_examples=n_examples,
update_hook=update_hook)
else:
input_fn = self.input_pipeline.get_predict_input_fn(
Xs, context=context)
estimator, hooks = self.get_estimator(
build_explain=PredictMode.EXPLAIN in predict_keys)
length = len(Xs) if not callable(Xs) else None
prediction_iterator = estimator.predict(input_fn=input_fn,
predict_keys=predict_keys,
hooks=hooks)
predictions = ProgressBar(prediction_iterator,
total=n_examples or length,
desc="Inference",
update_hook=update_hook)
try:
return [
pred[predict_keys[0]] if len(predict_keys) == 1 else pred
for pred in predictions
]
except ValueError:
raise FinetuneError(
"Cannot call `predict()` on a model that has not been fit."
)
def fit(self, *args, **kwargs):
""" An alias for finetune. """
return self.finetune(*args, **kwargs)
def _predict(self, Xs, context=None, **kwargs):
raw_preds = self._inference(Xs,
predict_keys=[PredictMode.NORMAL],
context=context,
**kwargs)
return self.input_pipeline.label_encoder.inverse_transform(
np.asarray(raw_preds))
def predict(self, Xs, context=None, **kwargs):
return self._predict(Xs, context=context, **kwargs)
def _predict_proba(self, Xs, context=None, **kwargs):
"""
Produce raw numeric outputs for proba predictions
"""
raw_preds = self._inference(Xs,
predict_keys=[PredictMode.PROBAS],
context=context,
**kwargs)
return raw_preds
def predict_proba(self, *args, **kwargs):
"""
The base method for predicting from the model.
"""
raw_probas = self._predict_proba(*args, **kwargs)
classes = self.input_pipeline.label_encoder.classes_
formatted_predictions = []
for probas in raw_probas:
formatted_predictions.append(dict(zip(classes, probas)))
return formatted_predictions
def attention_weights(self, Xs):
if self.config.base_model in [GPTModel, GPTModelSmall]:
raw_preds = self._inference(Xs,
predict_keys=[PredictMode.ATTENTION])
return raw_preds
raise NotImplementedError(
"'attention_weights' only supported for GPTModel and GPTModelSmall base models."
)
def _featurize(self, Xs, **kwargs):
raw_preds = self._inference(Xs,
predict_keys=[PredictMode.FEATURIZE],
**kwargs)
return np.asarray(raw_preds)
def _featurize_sequence(self, Xs, **kwargs):
raw_preds = self._inference(Xs,
predict_keys=[PredictMode.SEQUENCE],
**kwargs)
return np.asarray(raw_preds)
def featurize(self, *args, **kwargs):
"""
Base method to get raw features out of the model.
These features are the same features that are fed into the target_model.
"""
return self._featurize(*args, **kwargs)
def featurize_sequence(self, *args, **kwargs):
"""
Base method to get raw token-level features out of the model.
These features are the same features that are fed into the target_model.
"""
return self._featurize_sequence(*args, **kwargs)
@classmethod
def get_eval_fn(cls):
raise NotImplementedError(
"No default eval function is given, please pass an explicit eval fn to grid_search"
)
def transform(self, *args, **kwargs):
"""
An alias for `featurize`.
"""
return self.featurize(*args, **kwargs)
def _set_random_seed(self, seed=None):
seed = seed or self.config.seed
random.seed(seed)
np.random.seed(seed)
tf.set_random_seed(seed)
def generate_text(self,
seed_text="",
max_length=None,
use_extra_toks=None):
"""
Performs a prediction on the Language modeling objective given some seed text. It uses a noisy greedy decoding.
Temperature parameter for decoding is set in the config.
:param max_length: The maximum length to decode to.
:param seed_text: Defaults to the empty string. This will form the starting point to begin modelling
:return: A string containing the generated text.
"""
if use_extra_toks is None:
use_extra_toks = self._trained
def dataset_encoded():
while not dataset_encoded.finished:
yield {
"tokens": arr_encoded.token_ids,
"mask": arr_encoded.mask
}
dataset_encoded.finished = False
def get_input_fn():
types, shapes = self.input_pipeline.feed_shape_type_def()
tf_dataset = Dataset.from_generator(dataset_encoded, types[0],
shapes[0])
return tf_dataset.batch(1)
self.config.use_extra_toks = use_extra_toks
encoded = self.input_pipeline.text_encoder._encode([seed_text])
if encoded.token_ids == [] and not use_extra_toks:
raise ValueError(
"If you are not using the extra tokens, you must provide some non-empty seed text"
)
start = [self.input_pipeline.text_encoder.start_token
] if use_extra_toks else []
token_ids = start
if encoded.token_ids is not None and len(encoded.token_ids):
token_ids += encoded.token_ids[0]
encoded = EncodedOutput(token_ids=token_ids)
estimator, hooks = self.get_estimator(force_build_lm=True)
predict = estimator.predict(input_fn=get_input_fn,
predict_keys=[PredictMode.GENERATE_TEXT],
hooks=hooks)
EOS = self.input_pipeline.text_encoder.end_token
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
for i in range(
len(encoded.token_ids) - 1,
(max_length or self.config.max_length) - 2):
arr_encoded = self.input_pipeline._array_format(encoded)
class_idx = next(predict)[PredictMode.GENERATE_TEXT]
encoded.token_ids.append(class_idx[-1])
if encoded.token_ids[-1] == EOS:
break
dataset_encoded.finished = True
del self.config["use_extra_toks"]
return self.input_pipeline.text_encoder.decode(encoded.token_ids)
def __getstate__(self):
"""
Leave serialization of all tf objects to tf
"""
required_fields = [
"_load_from_file", "config", "input_pipeline", "_trained"
]
serialized_state = {
k: v
for k, v in self.__dict__.items() if k in required_fields
}
return serialized_state
def save(self, path):
"""
Saves the state of the model to disk to the folder specific by `path`. If `path` does not exist, it will be auto-created.
Save is performed in two steps:
- Serialize tf graph to disk using tf.Saver
- Serialize python model using pickle
Note:
Does not serialize state of Adam optimizer.
Should not be used to save / restore a training model.
"""
if path is None:
return
if isinstance(path, str):
path = os.path.abspath(path)
self.saver.save(self, path)
def create_base_model(self, filename, exists_ok=False):
"""
Saves the current weights into the correct file format to be used as a base model.
:param filename: the path to save the base model relative to finetune's base model filestore.
:param exists_ok: Whether to replace the model if it exists.
"""
base_model_path = os.path.join(os.path.dirname(__file__), "model",
filename)
if not exists_ok and os.path.exists(base_model_path):
base_model_path = base_model_path + str(int(time.time()))
LOGGER.warning(
"Cannot overwrite model {}, set exists_ok to overwrite, saving as {} to avoid loss of data."
.format(filename, base_model_path))
if not self.saver.variables:
raise FinetuneError(
"Cannot save a base model with no weights changed. Call fit before creating a base model."
)
weights_stripped = {
k: v
for k, v in self.saver.variables.items()
if "featurizer" in k and "Adam" not in k
}
joblib.dump(weights_stripped, base_model_path)
def load(path, *args, **kwargs):
"""
Load a saved fine-tuned model from disk. Path provided should be a folder which contains .pkl and tf.Saver() files
:param path: string path name to load model from. Same value as previously provided to :meth:`save`. Must be a folder.
:param **kwargs: key-value pairs of config items to override.
"""
if type(path) != str and not hasattr(path, "write"):
instance = path
raise FinetuneError(
'The .load() method can only be called on the class, not on an instance. Try `{}.load("{}") instead.'
.format(instance.__class__.__name__, args[0]))
assert_valid_config(**kwargs)
saver = Saver()
model = saver.load(path)
# Backwards compatability
# Ensure old models get new default settings
for setting, default in get_default_config().items():
if not hasattr(model.config, setting):
if setting == "add_eos_bos_to_chunk":
model.config.add_eos_bos_to_chunk = False
else:
model.config.update({setting: default})
model.config.update(kwargs)
model.input_pipeline.config = model.config
download_data_if_required(model.config.base_model)
saver.set_fallback(model.config.base_model_path)
model._initialize()
model.saver.variables = saver.variables
model._trained = True
return model
@classmethod
def finetune_grid_search(cls,
Xs,
Y,
*,
test_size,
eval_fn=None,
probs=False,
return_all=False,
**kwargs):
"""
Performs grid search over config items defined using "GridSearchable" objects and returns either full results or
the config object that relates to the best results. The default config contains grid searchable objects for the
most important parameters to search over.
:param Xs: Input text. Either [num_samples] or [sequence, num_samples] for single or multi input models respectively.
:param Y: Targets, A list of targets, [num_samples] that correspond to each sample in Xs.
:param test_size: Int or float. If an int is given this number of samples is used to validate, if a float is
given then that fraction of samples is used.
:param eval_fn: An eval function that takes 2 inputs (prediction, truth) and returns a float, with a max value being desired.
:param probs: If true, eval_fn is passed probability outputs from predict_proba, otherwise the output of predict is used.
:param return_all: If True, all results are returned, if False, only the best config is returned.
:param kwargs: Keyword arguments to pass to get_config()
:return: default is to return the best config object. If return_all is true, it returns a list of tuples of the
form [(config, eval_fn output), ... ]
"""
if isinstance(Xs[0], str):
Xs = [Xs]
config = get_config(**kwargs)
config.val_size = 0.0
eval_fn = eval_fn or cls.get_eval_fn()
trainXs, testXs, trainY, testY = train_test_split(list_transpose(Xs),
Y,
test_size=test_size,
shuffle=True)
trainXs = list_transpose(trainXs)
testXs = list_transpose(testXs)
gs = config.get_grid_searchable()
ranged_keys = gs.keys()
ranged_iterators = gs.values()
grid_gen = itertools.product(*ranged_iterators)
results = []
for grid_item in grid_gen:
config_ = deepcopy(config)
config_.update(dict(zip(ranged_keys, grid_item)))
instance = cls(config=config_)
instance.finetune(*trainXs, Y=trainY)
if probs:
res = instance.predict_proba(*testXs)
else:
res = instance.predict(*testXs)
results.append((config_, eval_fn(res, testY)))
del instance
if return_all:
return results
return max(results, key=lambda x: x[1])[0]
@classmethod
def finetune_grid_search_cv(cls,
Xs,
Y,
*,
n_splits,
test_size,
eval_fn=None,
probs=False,
return_all=False,
**kwargs):
"""
Performs cross validated grid search over config items defined using "GridSearchable" objects and returns either full results or
the config object that relates to the best results. The default config contains grid searchable objects for the
most important parameters to search over.
It should be noted that the cv splits are not guaranteed unique, but each split is given to each set of hparams.
:param Xs: Input text. Either [num_samples] or [sequence, num_samples] for single or multi input models respectively.
:param Y: Targets, A list of targets, [num_samples] that correspond to each sample in Xs.
:param n_splits: Number of CV splits to do.
:param test_size: Int or float. If an int is given this number of samples is used to validate, if a float is
given then that fraction of samples is used.
:param eval_fn: An eval function that takes 2 batches of outputs and returns a float, with a max value being
desired. An arithmetic mean must make sense for this metric.
:param probs: If true, eval_fn is passed probability outputs from predict_proba, otherwise the output of predict is used.
:param return_all: If True, all results are returned, if False, only the best config is returned.
:param kwargs: Keyword arguments to pass to get_config()
:return: default is to return the best config object. If return_all is true, it returns a list of tuples of the
form [(config, eval_fn output), ... ]
"""
results = []
for _ in range(n_splits):
res = cls.finetune_grid_search(Xs,
Y,
test_size=test_size,
probs=probs,
eval_fn=eval_fn,
return_all=True,
**kwargs)
results.append(res)
results = list(zip(*results))
aggregated_results = []
for configuration in results:
config_common = None
sum_res = 0
n_res = 0
for config, result in configuration:
config_common = config_common or config
assert config == config_common
n_res += 1
sum_res += result
aggregated_results.append((config_common, sum_res / n_res))
if return_all:
return aggregated_results
return max(aggregated_results, key=lambda x: x[1])[0]
def process_long_sequence(self, X, context=None, **kwargs):
arr_encoded = [
self.input_pipeline._text_to_ids(x)
for x in self.input_pipeline._format_for_inference(X)
]
flat_array_encoded = []
sequence_id = []
for i, ae in enumerate(arr_encoded):
for sample in ae:
flat_array_encoded.append(sample)
sequence_id.append(i)
labels, batch_probas = [], []
for pred in self._inference(
X,
predict_keys=[PredictMode.PROBAS, PredictMode.NORMAL],
n_examples=len(flat_array_encoded),
context=context,
**kwargs):
normal_pred = pred[PredictMode.NORMAL]
if not hasattr(self, 'multi_label'):
normal_pred = np.expand_dims(normal_pred, 0)
labels.append(
self.input_pipeline.label_encoder.inverse_transform(
normal_pred))
batch_probas.append(pred[PredictMode.PROBAS])
if not batch_probas:
batch_probas = [None] * len(labels)
for chunk_idx, (label_seq,
proba_seq) in enumerate(zip(labels, batch_probas)):
position_seq = flat_array_encoded[chunk_idx].char_locs
start_of_doc = chunk_idx == 0 or sequence_id[
chunk_idx - 1] != sequence_id[chunk_idx]
end_of_doc = (chunk_idx + 1 == len(flat_array_encoded) or
sequence_id[chunk_idx] != sequence_id[chunk_idx + 1])
yield position_seq, start_of_doc, end_of_doc, label_seq, proba_seq
def __del__(self):
if hasattr(self, "_tmp_dir") and self._tmp_dir is not None:
self._tmp_dir.cleanup()
class BaseModel(object, metaclass=ABCMeta):
"""
A sklearn-style class for finetuning a Transformer language model on a classification task.
"""
def __init__(self, config=None, **kwargs):
"""
For a full list of configuration options, see `finetune.config`.
:param config: A config object generated by `finetune.config.get_config` or None (for default config).
:param **kwargs: key-value pairs of config items to override.
"""
tf.reset_default_graph()
self.config = config or get_default_config()
self.config.update(kwargs)
if self.config.num_layers_trained != self.config.n_layer and self.config.train_embeddings:
raise ValueError("If you are only finetuning a subset of the layers, you cannot finetune embeddings.")
self.label_encoder = None
self._initialize()
self.target_dim = None
self._load_from_file = False
def _initialize(self):
# Initializes the non-serialized bits of the class.
self._set_random_seed(self.config.seed)
download_data_if_required()
self.encoder = TextEncoder()
# symbolic ops
self.logits = None # classification logits
self.target_loss = None # cross-entropy loss
self.lm_losses = None # language modeling losses
self.lm_predict_op = None
self.train = None # gradient + parameter update
self.features = None # hidden representation fed to classifier
self.summaries = None # Tensorboard summaries
self.train_writer = None
self.valid_writer = None
self.predict_params = None
self.train_op = None
self.predict_op = None
self.predict_proba_op = None
self.sess = None
self.noop = tf.no_op()
# indicator vars
self.is_built = False # has tf graph been constructed?
self.is_trained = False # has model been fine-tuned?
self.require_lm = False
def process_embeddings(name, value):
if "/we:0" not in name:
return value
vocab_size = self.encoder.vocab_size
word_embeddings = value[:vocab_size - len(self.encoder.special_tokens)]
special_embed = value[len(word_embeddings): vocab_size]
positional_embed = value[vocab_size:]
if self.config.interpolate_pos_embed and self.config.max_length != len(positional_embed):
positional_embed = interpolate_pos_embed(positional_embed, self.config.max_length)
elif self.config.max_length > len(positional_embed):
raise ValueError("Max Length cannot be greater than {} if interploate_pos_embed is turned off".format(len(positional_embed)))
else:
positional_embed = positional_embed[:self.config.max_length]
embeddings = np.concatenate((word_embeddings, special_embed, positional_embed), axis=0)
return embeddings
self.saver = Saver(
fallback_filename=JL_BASE,
exclude_matches=None if self.config.save_adam_vars else "adam",
variable_transforms=[process_embeddings]
)
def _format_for_encoding(self, Xs):
"""
Most subclasses take in inputs as:
List (batch) of list (docs)
Encode_multi_input expect the following format:
List (batch) of list (docs) of list (subseqs) of text
This method is responsible for standardizing inputs to the above format
"""
return [[[x] for x in X] for X in Xs]
def _text_to_ids(self, Xs, Y=None, max_length=None):
# Maps lists of text to formatted numpy arrays of token ids and loss-masks marking the lengths of the sequences.
max_length = max_length or self.config.max_length
# If 1d array of text is passed, coerce into multifield format
if len(Xs) and isinstance(Xs[0], (bytes, str)):
Xs = [[x] for x in Xs]
Xs = self._format_for_encoding(Xs)
if self.config.chunk_long_sequences and len(Xs[0]) == 1:
# can only chunk single sequence inputs
chunk_size = max_length - 2
step_size = chunk_size // 3
encoded = self.encoder.encode_multi_input(Xs, Y=Y, max_length=sys.maxsize)
d = defaultdict(list)
for idx in range(len(encoded.token_ids)):
starts = list(range(0, len(encoded.token_ids[idx]), step_size))
for start in starts:
end = start + chunk_size
for field in EncodedOutput._fields:
field_value = getattr(encoded, field)
if field_value is not None:
d[field].append(field_value[idx][start:end])
encoder_out = EncodedOutput(**d)
return self._array_format(encoder_out)
else:
encoder_out = self.encoder.encode_multi_input(Xs, Y=Y, max_length=max_length)
return self._array_format(encoder_out)
@abstractmethod
def _predict_op(self, logits, **kwargs):
raise NotImplementedError
@abstractmethod
def _predict_proba_op(self, logits, **kwargs):
raise NotImplementedError
@abstractmethod
def _target_model(self, *, featurizer_state, targets, n_outputs, train=False, reuse=None, **kwargs):
# Overridden by subclass to attach a target model onto the shared base featurizer.
raise NotImplementedError
@abstractmethod
def _target_encoder(self):
# Overridden by subclass to produce the right target encoding for a given target model.
raise NotImplementedError
def _eval(self, *tensors, feed_dict):
"""
Evaluate the value of each of the provided tensors.
Returns a `dict` that maps from tensor to result value.
If any result value is None, that result is excluded from the results `dict`.
"""
tensors = [
tensor if tensor is not None else self.noop
for tensor in tensors
]
values = self.sess.run(tensors, feed_dict=feed_dict)
return {
tensor: value
for tensor, value in zip(tensors, values)
if value is not None
}
def finetune(self, Xs, Y=None, batch_size=None):
arr_encoded = self._text_to_ids(Xs)
return self._training_loop(
arr_encoded,
Y=Y,
batch_size=batch_size,
)
def _training_loop(self, arr_encoded, Y=None, batch_size=None):
self.label_encoder = self._target_encoder()
idxs = list(range(len(arr_encoded.token_ids)))
train_idxs, val_idxs = train_test_split(idxs, test_size=self.config.val_size)
if Y is None:
# only language model will be trained, mock fake target of right length
train_Y = np.asarray([[]] * len(train_idxs))
val_Y = np.asarray([[]] * len(val_idxs))
target_dim = None
else:
Y = np.asarray(Y)
train_Y = self.label_encoder.fit_transform(Y[train_idxs])
val_Y = self.label_encoder.transform(Y[val_idxs])
target_dim = self.label_encoder.target_dim
batch_size = batch_size or self.config.batch_size
n_batch_train = batch_size * max(len(self.config.visible_gpus), 1)
n_examples = len(train_idxs)
n_updates_total = (n_examples // n_batch_train) * self.config.n_epochs
train_dataset = (arr_encoded.token_ids[train_idxs], arr_encoded.mask[train_idxs], train_Y)
val_dataset = (arr_encoded.token_ids[val_idxs], arr_encoded.mask[val_idxs], val_Y)
self._build_model(n_updates_total=n_updates_total, target_dim=target_dim)
self.is_trained = True
avg_train_loss = None
avg_val_loss = None
global_step = 0
best_val_loss = float("inf")
val_window = [float("inf")] * self.config.val_window_size
for i in range(self.config.n_epochs):
iterator = iter_data(
*train_dataset,
n_batch=n_batch_train,
tqdm_desc="Epoch {}".format(i),
verbose=self.config.verbose
)
for (xmb, mmb, ymb) in iterator:
feed_dict = {
self.X: xmb,
self.M: mmb,
}
if target_dim:
feed_dict[self.Y] = ymb
global_step += 1
if global_step % self.config.val_interval == 0:
feed_dict[self.do_dropout] = DROPOUT_OFF
outputs = self._eval(self.summaries, feed_dict=feed_dict)
if self.train_writer is not None:
self.train_writer.add_summary(outputs.get(self.summaries), global_step)
sum_val_loss = 0
for xval, mval, yval in iter_data(*val_dataset, n_batch=n_batch_train, verbose=self.config.verbose,
tqdm_desc="Validation"):
feed_dict = {
self.X: xval,
self.M: mval,
self.do_dropout: DROPOUT_OFF
}
if target_dim:
feed_dict[self.Y] = yval
outputs = self._eval(self.target_loss, self.summaries, feed_dict=feed_dict)
if self.valid_writer is not None:
self.valid_writer.add_summary(outputs.get(self.summaries), global_step)
val_cost = outputs.get(self.target_loss, 0)
sum_val_loss += val_cost
if avg_val_loss is None:
avg_val_loss = val_cost
else:
avg_val_loss = (
avg_val_loss * self.config.rolling_avg_decay
+ val_cost * (1 - self.config.rolling_avg_decay)
)
val_window.append(sum_val_loss)
val_window.pop(0)
if np.mean(val_window) <= best_val_loss:
best_val_loss = np.mean(val_window)
if self.config.autosave_path is not None:
self.save(self.config.autosave_path)
tqdm.tqdm.write("Train loss: {}\t Validation loss: {}".format(avg_train_loss, avg_val_loss))
feed_dict[self.do_dropout] = DROPOUT_ON
outputs = self._eval(self.target_loss, self.train_op, feed_dict=feed_dict)
cost = outputs.get(self.target_loss, 0)
if avg_train_loss is None:
avg_train_loss = cost
else:
avg_train_loss = avg_train_loss * self.config.rolling_avg_decay + cost * (
1 - self.config.rolling_avg_decay)
return self
def fit(self, *args, **kwargs):
""" An alias for finetune. """
return self.finetune(*args, **kwargs)
def _predict(self, Xs, max_length=None):
predictions = []
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
max_length = max_length or self.config.max_length
for xmb, mmb in self._infer_prep(Xs, max_length=max_length):
output = self._eval(self.predict_op,
feed_dict={
self.X: xmb,
self.M: mmb,
self.do_dropout: DROPOUT_OFF
}
)
prediction = output.get(self.predict_op)
formatted_predictions = self.label_encoder.inverse_transform(prediction)
predictions.append(formatted_predictions)
return np.concatenate(predictions).tolist()
def predict(self, Xs, max_length=None):
return self._predict(Xs, max_length=max_length)
def _predict_proba(self, Xs, max_length=None):
"""
Produce raw numeric outputs for proba predictions
"""
predictions = []
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
max_length = max_length or self.config.max_length
for xmb, mmb in self._infer_prep(Xs, max_length=max_length):
output = self._eval(
self.predict_proba_op,
feed_dict={
self.X: xmb,
self.M: mmb,
self.do_dropout: DROPOUT_OFF
}
)
probas = output.get(self.predict_proba_op)
predictions.extend(probas)
return predictions
def predict_proba(self, *args, **kwargs):
"""
The base method for predicting from the model.
"""
raw_probas = self._predict_proba(*args, **kwargs)
classes = self.label_encoder.classes_
formatted_predictions = []
for probas in raw_probas:
formatted_predictions.append(
dict(zip(classes, probas))
)
return formatted_predictions
def _featurize(self, Xs, max_length=None):
features = []
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
max_length = max_length or self.config.max_length
for xmb, mmb in self._infer_prep(Xs, max_length=max_length):
feature_batch = self.sess.run(self.features, {
self.X: xmb,
self.M: mmb,
self.do_dropout: DROPOUT_OFF
})
features.append(feature_batch)
return np.concatenate(features)
@abstractmethod
def featurize(self, *args, **kwargs):
"""
Base method to get raw features out of the model.
These features are the same that are fed into the target_model.
"""
return self._featurize(*args, **kwargs)
@classmethod
def get_eval_fn(cls):
raise NotImplementedError("No default eval function is given, please pass an explicit eval fn to grid_search")
def transform(self, *args, **kwargs):
"""
An alias for `featurize`.
"""
return self.featurize(*args, **kwargs)
def _infer_prep(self, Xs, max_length=None):
max_length = max_length or self.config.max_length
arr_encoded = self._text_to_ids(Xs, max_length=max_length)
n_batch_train = self.config.batch_size * max(len(self.config.visible_gpus), 1)
self._build_model(n_updates_total=0, target_dim=self.target_dim, train=False)
yield from iter_data(arr_encoded.token_ids, arr_encoded.mask, n_batch=n_batch_train,
verbose=self.config.verbose)
def _array_format(self, encoded_output):
"""
Returns numpy array of token idxs and corresponding mask
Returned `x` array contains two channels:
0: byte-pair encoding embedding
1: positional embedding
"""
n = len(encoded_output.token_ids)
seq_lengths = [len(x) for x in encoded_output.token_ids]
x = np.zeros((n, self.config.max_length, 2), dtype=np.int32)
mask = np.zeros((n, self.config.max_length), dtype=np.float32)
labels_arr = np.full((n, self.config.max_length), PAD_TOKEN,
dtype='object') if encoded_output.labels is not None else None
for i, seq_length in enumerate(seq_lengths):
# BPE embedding
x[i, :seq_length, 0] = encoded_output.token_ids[i]
# masking: value of 1 means "consider this in cross-entropy LM loss"
mask[i, 1:seq_length] = 1
if encoded_output.labels:
labels_arr[i, :seq_length] = encoded_output.labels[i]
# positional_embeddings
x[:, :, 1] = np.arange(self.encoder.vocab_size, self.encoder.vocab_size + self.config.max_length)
return ArrayEncodedOutput(
token_ids=x,
tokens=encoded_output.tokens,
labels=labels_arr,
char_locs=encoded_output.char_locs,
mask=mask,
)
def _compile_train_op(self, *, params, grads, n_updates_total):
grads = average_grads(grads)
if self.config.summarize_grads:
self.summaries += tf.contrib.training.add_gradients_summaries(grads)
grads = [grad for grad, param in grads]
self.train_op = AdamWeightDecay(
params=params,
grads=grads,
lr=self.config.lr,
schedule=partial(schedules[self.config.lr_schedule], warmup=self.config.lr_warmup),
t_total=n_updates_total,
l2=self.config.l2_reg,
max_grad_norm=self.config.max_grad_norm,
vector_l2=self.config.vector_l2,
b1=self.config.b1,
b2=self.config.b2,
e=self.config.epsilon,
pretrained_weights=self.saver.get_pretrained_weights(),
deviation_regularization=self.config.regularize_deviation
)
def _construct_graph(self, n_updates_total, target_dim=None, train=True):
gpu_grads = []
self.summaries = []
# store whether or not graph was previously compiled with dropout
self.train = train
self._define_placeholders(target_dim=target_dim)
aggregator = defaultdict(list)
train_loss_tower = 0
gpus = self.config.visible_gpus
n_splits = max(len(gpus), 1)
# multi-GPU setup, using CPU as param server is most efficient unless system has direct GPU connections
# single GPU, no need to use a different GPU as a parameter server
params_device = 'cpu' if len(gpus) != 1 else gpus[0]
# decide on setting for language model loss coefficient
# if the language model loss does not contribute to overall loss,
# remove the language model computation from the graph
lm_loss_coef = self.config.lm_loss_coef
if target_dim is None:
lm_loss_coef = 1.0
compile_lm = (train and lm_loss_coef > 0) or self.require_lm
for i, (X, M, Y) in enumerate(soft_split(self.X, self.M, self.Y, n_splits=n_splits)):
do_reuse = True if i > 0 else tf.AUTO_REUSE
if gpus:
device = tf.device(assign_to_gpu(gpus[i], params_device=params_device))
else:
device = tf.device('cpu')
scope = tf.variable_scope(tf.get_variable_scope(), reuse=do_reuse)
with device, scope:
featurizer_state = featurizer(
X,
config=self.config,
encoder=self.encoder,
dropout_placeholder=self.do_dropout,
train=train,
reuse=do_reuse
)
if compile_lm:
language_model_state = language_model(
X=X,
M=M,
config=self.config,
embed_weights=featurizer_state['embed_weights'],
hidden=featurizer_state['sequence_features'],
reuse=do_reuse
)
train_loss = lm_loss_coef * tf.reduce_mean(language_model_state['losses'])
aggregator['lm_losses'].append(language_model_state['losses'])
lm_logits = language_model_state["logits"]
aggregator["lm_model"].append(sample_with_temperature(lm_logits, self.config.lm_temp))
else:
train_loss = 0
aggregator['features'].append(featurizer_state['features'])
if target_dim is not None:
with tf.variable_scope('model/target'):
target_model_state = self._target_model(
featurizer_state=featurizer_state,
targets=Y,
n_outputs=target_dim,
train=train,
reuse=do_reuse,
max_length=self.config.max_length
)
train_loss += (1 - lm_loss_coef) * tf.reduce_mean(target_model_state['losses'])
train_loss_tower += train_loss
aggregator['logits'].append(target_model_state['logits'])
aggregator['target_losses'].append(target_model_state['losses'])
params = find_trainable_variables("model")
grads = tf.gradients(train_loss, params)
grads = list(zip(grads, params))
gpu_grads.append(grads)
with tf.device(params_device):
self.features = tf.concat(aggregator['features'], axis=0)
if compile_lm:
self.lm_predict_op = tf.concat(aggregator["lm_model"], 0)
self.lm_losses = tf.concat(aggregator['lm_losses'], axis=0)
self.lm_loss = tf.reduce_mean(self.lm_losses)
self.summaries.append(tf.summary.scalar('LanguageModelLoss', self.lm_loss))
if train:
self._compile_train_op(
params=params,
grads=gpu_grads,
n_updates_total=n_updates_total
)
if target_dim is not None:
self.logits = tf.concat(aggregator['logits'], axis=0)
self.target_losses = concat_or_stack(aggregator['target_losses'])
self.predict_op = self._predict_op(
self.logits, **target_model_state.get("predict_params", {})
)
self.predict_proba_op = self._predict_proba_op(
self.logits, **target_model_state.get("predict_params", {})
)
self.target_loss = tf.reduce_mean(self.target_losses)
self.summaries.append(tf.summary.scalar('TargetModelLoss', self.target_loss))
self.summaries.append(tf.summary.scalar('TotalLoss', train_loss_tower / n_splits))
self.summaries = tf.summary.merge(self.summaries) if self.summaries else self.noop
def _build_model(self, n_updates_total, target_dim, train=True):
"""
Construct tensorflow symbolic graph.
"""
if not self.is_trained or train != self.train or self.target_dim != target_dim:
# reconstruct graph to include/remove dropout
# if `train` setting has changed
self._construct_graph(n_updates_total, target_dim, train=train)
self._initialize_session()
self.saver.initialize(self.sess)
self.target_dim = target_dim
if train:
if self.config.tensorboard_folder is not None:
if not os.path.exists(self.config.tensorboard_folder):
os.mkdir(self.config.tensorboard_folder)
self.train_writer = tf.summary.FileWriter(self.config.tensorboard_folder + '/train', self.sess.graph)
self.valid_writer = tf.summary.FileWriter(self.config.tensorboard_folder + '/valid', self.sess.graph)
self.is_built = True
def _initialize_session(self):
if self.sess is None:
gpus = self.config.visible_gpus
os.environ['CUDA_VISIBLE_DEVICES'] = ",".join([str(gpu) for gpu in gpus])
conf = tf.ConfigProto(allow_soft_placement=self.config.soft_device_placement,
log_device_placement=self.config.log_device_placement)
self.sess = tf.Session(config=conf)
def _set_random_seed(self, seed=None):
seed = seed or self.config.seed
random.seed(seed)
np.random.seed(seed)
tf.set_random_seed(seed)
def _target_placeholder(self, target_dim=None):
return tf.placeholder(tf.float32, [None, target_dim or 1]) # classification targets
def _define_placeholders(self, target_dim=None):
# tf placeholders
self.X = tf.placeholder(tf.int32, [None, self.config.max_length, 2]) # token idxs (BPE embedding + positional)
self.M = tf.placeholder(tf.float32, [None, self.config.max_length]) # sequence mask
# when target dim is not set, an array of [None] targets is passed as a placeholder
self.do_dropout = tf.placeholder(tf.float32) # 1 for do dropout and 0 to not do dropout
self.Y = self._target_placeholder(target_dim=target_dim)
def generate_text(self, seed_text='', max_length=None):
"""
Performs a prediction on the Language modeling objective given some seed text. It uses a noisy greedy decoding.
Temperature parameter for decoding is set in the config.
:param max_length: The maximum length to decode to.
:param seed_text: Defaults to the empty string. This will form the starting point to begin modelling
:return: A string containing the generated text.
"""
self.require_lm = True
encoded = self.encoder._encode([seed_text])
self._build_model(n_updates_total=0, target_dim=self.target_dim, train=False)
string = [self.encoder['_start_']] + encoded.token_ids[0]
EOS = self.encoder['_classify_']
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
for i in range(len(encoded.token_ids), (max_length or self.config.max_length) - 1):
arr_encoded = self._array_format(encoded)
class_idx = self.sess.run(self.lm_predict_op, {self.X: arr_encoded.token_ids, self.M: arr_encoded.mask})
string.append(class_idx[i])
if string[-1] == EOS:
break
return self.encoder.decode(string)
def __getstate__(self):
"""
Leave serialization of all tf objects to tf
"""
required_fields = [
'label_encoder', 'target_dim', '_load_from_file', 'config', 'target_type',
]
serialized_state = {
k: v for k, v in self.__dict__.items()
if k in required_fields
}
return serialized_state
def save(self, path):
"""
Saves the state of the model to disk to the folder specific by `path`. If `path` does not exist, it will be auto-created.
Save is performed in two steps:
- Serialize tf graph to disk using tf.Saver
- Serialize python model using pickle
Note:
Does not serialize state of Adam optimizer.
Should not be used to save / restore a training model.
"""
if path is None:
return
path = os.path.abspath(path)
self.saver.save(self, path)
self._load_from_file = False
@classmethod
def load(cls, path):
"""
Load a saved fine-tuned model from disk. Path provided should be a folder which contains .pkl and tf.Saver() files
:param path: string path name to load model from. Same value as previously provided to :meth:`save`. Must be a folder.
"""
saver = Saver(JL_BASE)
model = saver.load(path)
model._initialize()
model.saver.variables = saver.variables
tf.reset_default_graph()
return model
@classmethod
def finetune_grid_search(cls, Xs, Y, *, test_size, config=None, eval_fn=None, probs=False, return_all=False):
"""
Performs grid search over config items defined using "GridSearchable" objects and returns either full results or
the config object that relates to the best results. The default config contains grid searchable objects for the
most important parameters to search over.
:param Xs: Input text. Either [num_samples] or [sequence, num_samples] for single or multi input models respectively.
:param Y: Targets, A list of targets, [num_samples] that correspond to each sample in Xs.
:param test_size: Int or float. If an int is given this number of samples is used to validate, if a float is
given then that fraction of samples is used.
:param config: A config object, or None to use the default config.
:param eval_fn: An eval function that takes 2 inputs (prediction, truth) and returns a float, with a max value being desired.
:param probs: If true, eval_fn is passed probability outputs from predict_proba, otherwise the output of predict is used.
:param return_all: If True, all results are returned, if False, only the best config is returned.
:return: default is to return the best config object. If return_all is true, it returns a list of tuples of the
form [(config, eval_fn output), ... ]
"""
if isinstance(Xs[0], str):
Xs = [Xs]
config = config or get_default_config()
config.val_size = 0.0
eval_fn = eval_fn or cls.get_eval_fn()
trainXs, testXs, trainY, testY = train_test_split(list_transpose(Xs), Y, test_size=test_size, shuffle=True)
trainXs = list_transpose(trainXs)
testXs = list_transpose(testXs)
gs = config.get_grid_searchable()
ranged_keys = gs.keys()
ranged_iterators = gs.values()
grid_gen = itertools.product(*ranged_iterators)
results = []
for grid_item in grid_gen:
config_ = deepcopy(config)
config_.update(dict(zip(ranged_keys, grid_item)))
instance = cls(config=config_)
instance.finetune(*trainXs, Y=trainY)
if probs:
res = instance.predict_proba(*testXs)
else:
res = instance.predict(*testXs)
results.append((config_, eval_fn(res, testY)))
del instance
if return_all:
return results
return max(results, key=lambda x: x[1])[0]
@classmethod
def finetune_grid_search_cv(cls, Xs, Y, *, n_splits, test_size, config=None, eval_fn=None, probs=False, return_all=False):
"""
Performs cross validated grid search over config items defined using "GridSearchable" objects and returns either full results or
the config object that relates to the best results. The default config contains grid searchable objects for the
most important parameters to search over.
It should be noted that the cv splits are not guaranteed unique, but each split is given to each set of hparams.
:param Xs: Input text. Either [num_samples] or [sequence, num_samples] for single or multi input models respectively.
:param Y: Targets, A list of targets, [num_samples] that correspond to each sample in Xs.
:param n_splits: Number of CV splits to do.
:param test_size: Int or float. If an int is given this number of samples is used to validate, if a float is
given then that fraction of samples is used.
:param config: A config object, or None to use the default config.
:param eval_fn: An eval function that takes 2 batches of outputs and returns a float, with a max value being
desired. An arithmetic mean must make sense for this metric.
:param probs: If true, eval_fn is passed probability outputs from predict_proba, otherwise the output of predict is used.
:param return_all: If True, all results are returned, if False, only the best config is returned.
:return: default is to return the best config object. If return_all is true, it returns a list of tuples of the
form [(config, eval_fn output), ... ]
"""
results = []
for _ in range(n_splits):
res = cls.finetune_grid_search(Xs, Y, test_size=test_size, probs=probs, eval_fn=eval_fn, config=config,
return_all=True)
results.append(res)
results = list(zip(*results))
aggregated_results = []
for configuration in results:
config_common = None
sum_res = 0
n_res = 0
for config, result in configuration:
config_common = config_common or config
assert config == config_common
n_res += 1
sum_res += result
aggregated_results.append((config_common, sum_res/n_res))
if return_all:
return aggregated_results
return max(aggregated_results, key=lambda x: x[1])[0]
def __del__(self):
try:
if self.sess is not None:
self.sess.close()
except AttributeError:
pass
class BaseModel(object, metaclass=ABCMeta):
"""
A sklearn-style class for finetuning a Transformer language model on a classification task.
"""
def __init__(self, config=None, **kwargs):
"""
For a full list of configuration options, see `finetune.config`.
:param config: A config object generated by `finetune.config.get_config` or None (for default config).
:param **kwargs: key-value pairs of config items to override.
"""
tf.reset_default_graph()
self.config = config or get_default_config()
self.config.update(kwargs)
if self.config.num_layers_trained != self.config.n_layer and self.config.train_embeddings:
raise ValueError(
"If you are only finetuning a subset of the layers, you cannot finetune embeddings."
)
self.label_encoder = None
self._initialize()
self.target_dim = None
self._load_from_file = False
def _initialize(self):
# Initializes the non-serialized bits of the class.
self._set_random_seed(self.config.seed)
download_data_if_required()
self.encoder = TextEncoder()
# symbolic ops
self.logits = None # classification logits
self.target_loss = None # cross-entropy loss
self.lm_losses = None # language modeling losses
self.lm_predict_op = None
self.train = None # gradient + parameter update
self.features = None # hidden representation fed to classifier
self.summaries = None # Tensorboard summaries
self.train_writer = None
self.valid_writer = None
self.predict_params = None
self.train_op = None
self.predict_op = None
self.predict_proba_op = None
self.sess = None
self.noop = tf.no_op()
# indicator vars
self.is_built = False # has tf graph been constructed?
self.is_trained = False # has model been fine-tuned?
self.require_lm = False
def process_embeddings(name, value):
if "/we:0" not in name:
return value
vocab_size = self.encoder.vocab_size
word_embeddings = value[:vocab_size -
len(self.encoder.special_tokens)]
special_embed = value[len(word_embeddings):vocab_size]
positional_embed = value[vocab_size:]
if self.config.interpolate_pos_embed and self.config.max_length != len(
positional_embed):
positional_embed = interpolate_pos_embed(
positional_embed, self.config.max_length)
elif self.config.max_length > len(positional_embed):
raise ValueError(
"Max Length cannot be greater than {} if interploate_pos_embed is turned off"
.format(len(positional_embed)))
else:
positional_embed = positional_embed[:self.config.max_length]
embeddings = np.concatenate(
(word_embeddings, special_embed, positional_embed), axis=0)
return embeddings
self.saver = Saver(
fallback_filename=JL_BASE,
exclude_matches=None if self.config.save_adam_vars else "adam",
variable_transforms=[process_embeddings])
def _format_for_encoding(self, Xs):
"""
Most subclasses take in inputs as:
List (batch) of list (docs)
Encode_multi_input expect the following format:
List (batch) of list (docs) of list (subseqs) of text
This method is responsible for standardizing inputs to the above format
"""
return [[[x] for x in X] for X in Xs]
def _text_to_ids(self, Xs, Y=None, max_length=None):
# Maps lists of text to formatted numpy arrays of token ids and loss-masks marking the lengths of the sequences.
max_length = max_length or self.config.max_length
# If 1d array of text is passed, coerce into multifield format
if len(Xs) and isinstance(Xs[0], (bytes, str)):
Xs = [[x] for x in Xs]
Xs = self._format_for_encoding(Xs)
if self.config.chunk_long_sequences and len(Xs[0]) == 1:
# can only chunk single sequence inputs
chunk_size = max_length - 2
step_size = chunk_size // 3
encoded = self.encoder.encode_multi_input(Xs,
Y=Y,
max_length=sys.maxsize)
d = defaultdict(list)
for idx in range(len(encoded.token_ids)):
starts = list(range(0, len(encoded.token_ids[idx]), step_size))
for start in starts:
end = start + chunk_size
for field in EncodedOutput._fields:
field_value = getattr(encoded, field)
if field_value is not None:
d[field].append(field_value[idx][start:end])
encoder_out = EncodedOutput(**d)
return self._array_format(encoder_out)
else:
encoder_out = self.encoder.encode_multi_input(
Xs, Y=Y, max_length=max_length)
return self._array_format(encoder_out)
@abstractmethod
def _predict_op(self, logits, **kwargs):
raise NotImplementedError
@abstractmethod
def _predict_proba_op(self, logits, **kwargs):
raise NotImplementedError
@abstractmethod
def _target_model(self,
*,
featurizer_state,
targets,
n_outputs,
train=False,
reuse=None,
**kwargs):
# Overridden by subclass to attach a target model onto the shared base featurizer.
raise NotImplementedError
@abstractmethod
def _target_encoder(self):
# Overridden by subclass to produce the right target encoding for a given target model.
raise NotImplementedError
def _eval(self, *tensors, feed_dict):
"""
Evaluate the value of each of the provided tensors.
Returns a `dict` that maps from tensor to result value.
If any result value is None, that result is excluded from the results `dict`.
"""
tensors = [
tensor if tensor is not None else self.noop for tensor in tensors
]
values = self.sess.run(tensors, feed_dict=feed_dict)
return {
tensor: value
for tensor, value in zip(tensors, values) if value is not None
}
def finetune(self, Xs, Y=None, batch_size=None):
arr_encoded = self._text_to_ids(Xs)
return self._training_loop(
arr_encoded,
Y=Y,
batch_size=batch_size,
)
def _training_loop(self, arr_encoded, Y=None, batch_size=None):
self.label_encoder = self._target_encoder()
idxs = list(range(len(arr_encoded.token_ids)))
train_idxs, val_idxs = train_test_split(idxs,
test_size=self.config.val_size)
if Y is None:
# only language model will be trained, mock fake target of right length
train_Y = np.asarray([[]] * len(train_idxs))
val_Y = np.asarray([[]] * len(val_idxs))
target_dim = None
else:
Y = np.asarray(Y)
train_Y = self.label_encoder.fit_transform(Y[train_idxs])
val_Y = self.label_encoder.transform(Y[val_idxs])
target_dim = self.label_encoder.target_dim
batch_size = batch_size or self.config.batch_size
n_batch_train = batch_size * max(len(self.config.visible_gpus), 1)
n_examples = len(train_idxs)
n_updates_total = (n_examples // n_batch_train) * self.config.n_epochs
train_dataset = (arr_encoded.token_ids[train_idxs],
arr_encoded.mask[train_idxs], train_Y)
val_dataset = (arr_encoded.token_ids[val_idxs],
arr_encoded.mask[val_idxs], val_Y)
self._build_model(n_updates_total=n_updates_total,
target_dim=target_dim)
self.is_trained = True
avg_train_loss = None
avg_val_loss = None
global_step = 0
best_val_loss = float("inf")
val_window = [float("inf")] * self.config.val_window_size
for i in range(self.config.n_epochs):
iterator = iter_data(*train_dataset,
n_batch=n_batch_train,
tqdm_desc="Epoch {}".format(i),
verbose=self.config.verbose)
for (xmb, mmb, ymb) in iterator:
feed_dict = {
self.X: xmb,
self.M: mmb,
}
if target_dim:
feed_dict[self.Y] = ymb
global_step += 1
if global_step % self.config.val_interval == 0:
feed_dict[self.do_dropout] = DROPOUT_OFF
outputs = self._eval(self.summaries, feed_dict=feed_dict)
if self.train_writer is not None:
self.train_writer.add_summary(
outputs.get(self.summaries), global_step)
sum_val_loss = 0
for xval, mval, yval in iter_data(
*val_dataset,
n_batch=n_batch_train,
verbose=self.config.verbose,
tqdm_desc="Validation"):
feed_dict = {
self.X: xval,
self.M: mval,
self.do_dropout: DROPOUT_OFF
}
if target_dim:
feed_dict[self.Y] = yval
outputs = self._eval(self.target_loss,
self.summaries,
feed_dict=feed_dict)
if self.valid_writer is not None:
self.valid_writer.add_summary(
outputs.get(self.summaries), global_step)
val_cost = outputs.get(self.target_loss, 0)
sum_val_loss += val_cost
if avg_val_loss is None:
avg_val_loss = val_cost
else:
avg_val_loss = (
avg_val_loss * self.config.rolling_avg_decay +
val_cost * (1 - self.config.rolling_avg_decay))
val_window.append(sum_val_loss)
val_window.pop(0)
if np.mean(val_window) <= best_val_loss:
best_val_loss = np.mean(val_window)
if self.config.autosave_path is not None:
self.save(self.config.autosave_path)
tqdm.tqdm.write(
"Train loss: {}\t Validation loss: {}".format(
avg_train_loss, avg_val_loss))
feed_dict[self.do_dropout] = DROPOUT_ON
outputs = self._eval(self.target_loss,
self.train_op,
feed_dict=feed_dict)
cost = outputs.get(self.target_loss, 0)
if avg_train_loss is None:
avg_train_loss = cost
else:
avg_train_loss = avg_train_loss * self.config.rolling_avg_decay + cost * (
1 - self.config.rolling_avg_decay)
return self
def fit(self, *args, **kwargs):
""" An alias for finetune. """
return self.finetune(*args, **kwargs)
def _predict(self, Xs, max_length=None):
predictions = []
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
max_length = max_length or self.config.max_length
for xmb, mmb in self._infer_prep(Xs, max_length=max_length):
output = self._eval(self.predict_op,
feed_dict={
self.X: xmb,
self.M: mmb,
self.do_dropout: DROPOUT_OFF
})
prediction = output.get(self.predict_op)
formatted_predictions = self.label_encoder.inverse_transform(
prediction)
predictions.append(formatted_predictions)
return np.concatenate(predictions).tolist()
def predict(self, Xs, max_length=None):
return self._predict(Xs, max_length=max_length)
def _predict_proba(self, Xs, max_length=None):
"""
Produce raw numeric outputs for proba predictions
"""
predictions = []
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
max_length = max_length or self.config.max_length
for xmb, mmb in self._infer_prep(Xs, max_length=max_length):
output = self._eval(self.predict_proba_op,
feed_dict={
self.X: xmb,
self.M: mmb,
self.do_dropout: DROPOUT_OFF
})
probas = output.get(self.predict_proba_op)
predictions.extend(probas)
return predictions
def predict_proba(self, *args, **kwargs):
"""
The base method for predicting from the model.
"""
raw_probas = self._predict_proba(*args, **kwargs)
classes = self.label_encoder.classes_
formatted_predictions = []
for probas in raw_probas:
formatted_predictions.append(dict(zip(classes, probas)))
return formatted_predictions
def _featurize(self, Xs, max_length=None):
features = []
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
max_length = max_length or self.config.max_length
for xmb, mmb in self._infer_prep(Xs, max_length=max_length):
feature_batch = self.sess.run(self.features, {
self.X: xmb,
self.M: mmb,
self.do_dropout: DROPOUT_OFF
})
features.append(feature_batch)
return np.concatenate(features)
@abstractmethod
def featurize(self, *args, **kwargs):
"""
Base method to get raw features out of the model.
These features are the same that are fed into the target_model.
"""
return self._featurize(*args, **kwargs)
@classmethod
def get_eval_fn(cls):
raise NotImplementedError(
"No default eval function is given, please pass an explicit eval fn to grid_search"
)
def transform(self, *args, **kwargs):
"""
An alias for `featurize`.
"""
return self.featurize(*args, **kwargs)
def _infer_prep(self, Xs, max_length=None):
max_length = max_length or self.config.max_length
arr_encoded = self._text_to_ids(Xs, max_length=max_length)
n_batch_train = self.config.batch_size * max(
len(self.config.visible_gpus), 1)
self._build_model(n_updates_total=0,
target_dim=self.target_dim,
train=False)
yield from iter_data(arr_encoded.token_ids,
arr_encoded.mask,
n_batch=n_batch_train,
verbose=self.config.verbose)
def _array_format(self, encoded_output):
"""
Returns numpy array of token idxs and corresponding mask
Returned `x` array contains two channels:
0: byte-pair encoding embedding
1: positional embedding
"""
n = len(encoded_output.token_ids)
seq_lengths = [len(x) for x in encoded_output.token_ids]
x = np.zeros((n, self.config.max_length, 2), dtype=np.int32)
mask = np.zeros((n, self.config.max_length), dtype=np.float32)
labels_arr = np.full(
(n, self.config.max_length), PAD_TOKEN,
dtype='object') if encoded_output.labels is not None else None
for i, seq_length in enumerate(seq_lengths):
# BPE embedding
x[i, :seq_length, 0] = encoded_output.token_ids[i]
# masking: value of 1 means "consider this in cross-entropy LM loss"
mask[i, 1:seq_length] = 1
if encoded_output.labels:
labels_arr[i, :seq_length] = encoded_output.labels[i]
# positional_embeddings
x[:, :,
1] = np.arange(self.encoder.vocab_size,
self.encoder.vocab_size + self.config.max_length)
return ArrayEncodedOutput(
token_ids=x,
tokens=encoded_output.tokens,
labels=labels_arr,
char_locs=encoded_output.char_locs,
mask=mask,
)
def _compile_train_op(self, *, params, grads, n_updates_total):
grads = average_grads(grads)
if self.config.summarize_grads:
self.summaries += tf.contrib.training.add_gradients_summaries(
grads)
grads = [grad for grad, param in grads]
self.train_op = AdamWeightDecay(
params=params,
grads=grads,
lr=self.config.lr,
schedule=partial(schedules[self.config.lr_schedule],
warmup=self.config.lr_warmup),
t_total=n_updates_total,
l2=self.config.l2_reg,
max_grad_norm=self.config.max_grad_norm,
vector_l2=self.config.vector_l2,
b1=self.config.b1,
b2=self.config.b2,
e=self.config.epsilon,
pretrained_weights=self.saver.get_pretrained_weights(),
deviation_regularization=self.config.regularize_deviation)
def _construct_graph(self, n_updates_total, target_dim=None, train=True):
gpu_grads = []
self.summaries = []
# store whether or not graph was previously compiled with dropout
self.train = train
self._define_placeholders(target_dim=target_dim)
aggregator = defaultdict(list)
train_loss_tower = 0
gpus = self.config.visible_gpus
n_splits = max(len(gpus), 1)
# multi-GPU setup, using CPU as param server is most efficient unless system has direct GPU connections
# single GPU, no need to use a different GPU as a parameter server
params_device = 'cpu' if len(gpus) != 1 else gpus[0]
# decide on setting for language model loss coefficient
# if the language model loss does not contribute to overall loss,
# remove the language model computation from the graph
lm_loss_coef = self.config.lm_loss_coef
if target_dim is None:
lm_loss_coef = 1.0
compile_lm = (train and lm_loss_coef > 0) or self.require_lm
for i, (X, M, Y) in enumerate(
soft_split(self.X, self.M, self.Y, n_splits=n_splits)):
do_reuse = True if i > 0 else tf.AUTO_REUSE
if gpus:
device = tf.device(
assign_to_gpu(gpus[i], params_device=params_device))
else:
device = tf.device('cpu')
scope = tf.variable_scope(tf.get_variable_scope(), reuse=do_reuse)
with device, scope:
featurizer_state = featurizer(
X,
config=self.config,
encoder=self.encoder,
dropout_placeholder=self.do_dropout,
train=train,
reuse=do_reuse)
if compile_lm:
language_model_state = language_model(
X=X,
M=M,
config=self.config,
embed_weights=featurizer_state['embed_weights'],
hidden=featurizer_state['sequence_features'],
reuse=do_reuse)
train_loss = lm_loss_coef * tf.reduce_mean(
language_model_state['losses'])
aggregator['lm_losses'].append(
language_model_state['losses'])
lm_logits = language_model_state["logits"]
aggregator["lm_model"].append(
sample_with_temperature(lm_logits,
self.config.lm_temp))
else:
train_loss = 0
aggregator['features'].append(featurizer_state['features'])
if target_dim is not None:
with tf.variable_scope('model/target'):
target_model_state = self._target_model(
featurizer_state=featurizer_state,
targets=Y,
n_outputs=target_dim,
train=train,
reuse=do_reuse,
max_length=self.config.max_length)
train_loss += (1 - lm_loss_coef) * tf.reduce_mean(
target_model_state['losses'])
train_loss_tower += train_loss
aggregator['logits'].append(target_model_state['logits'])
aggregator['target_losses'].append(
target_model_state['losses'])
params = find_trainable_variables("model")
grads = tf.gradients(train_loss, params)
grads = list(zip(grads, params))
gpu_grads.append(grads)
with tf.device(params_device):
self.features = tf.concat(aggregator['features'], axis=0)
if compile_lm:
self.lm_predict_op = tf.concat(aggregator["lm_model"], 0)
self.lm_losses = tf.concat(aggregator['lm_losses'], axis=0)
self.lm_loss = tf.reduce_mean(self.lm_losses)
self.summaries.append(
tf.summary.scalar('LanguageModelLoss', self.lm_loss))
if train:
self._compile_train_op(params=params,
grads=gpu_grads,
n_updates_total=n_updates_total)
if target_dim is not None:
self.logits = tf.concat(aggregator['logits'], axis=0)
self.target_losses = concat_or_stack(
aggregator['target_losses'])
self.predict_op = self._predict_op(
self.logits,
**target_model_state.get("predict_params", {}))
self.predict_proba_op = self._predict_proba_op(
self.logits,
**target_model_state.get("predict_params", {}))
self.target_loss = tf.reduce_mean(self.target_losses)
self.summaries.append(
tf.summary.scalar('TargetModelLoss', self.target_loss))
self.summaries.append(
tf.summary.scalar('TotalLoss',
train_loss_tower / n_splits))
self.summaries = tf.summary.merge(
self.summaries) if self.summaries else self.noop
def _build_model(self, n_updates_total, target_dim, train=True):
"""
Construct tensorflow symbolic graph.
"""
if not self.is_trained or train != self.train or self.target_dim != target_dim:
# reconstruct graph to include/remove dropout
# if `train` setting has changed
self._construct_graph(n_updates_total, target_dim, train=train)
self._initialize_session()
self.saver.initialize(self.sess)
self.target_dim = target_dim
if train:
if self.config.tensorboard_folder is not None:
if not os.path.exists(self.config.tensorboard_folder):
os.mkdir(self.config.tensorboard_folder)
self.train_writer = tf.summary.FileWriter(
self.config.tensorboard_folder + '/train', self.sess.graph)
self.valid_writer = tf.summary.FileWriter(
self.config.tensorboard_folder + '/valid', self.sess.graph)
self.is_built = True
def _initialize_session(self):
if self.sess is None:
gpus = self.config.visible_gpus
os.environ['CUDA_VISIBLE_DEVICES'] = ",".join(
[str(gpu) for gpu in gpus])
conf = tf.ConfigProto(
allow_soft_placement=self.config.soft_device_placement,
log_device_placement=self.config.log_device_placement)
self.sess = tf.Session(config=conf)
def _set_random_seed(self, seed=None):
seed = seed or self.config.seed
random.seed(seed)
np.random.seed(seed)
tf.set_random_seed(seed)
def _target_placeholder(self, target_dim=None):
return tf.placeholder(
tf.float32, [None, target_dim or 1]) # classification targets
def _define_placeholders(self, target_dim=None):
# tf placeholders
self.X = tf.placeholder(tf.int32,
[None, self.config.max_length, 2
]) # token idxs (BPE embedding + positional)
self.M = tf.placeholder(
tf.float32, [None, self.config.max_length]) # sequence mask
# when target dim is not set, an array of [None] targets is passed as a placeholder
self.do_dropout = tf.placeholder(
tf.float32) # 1 for do dropout and 0 to not do dropout
self.Y = self._target_placeholder(target_dim=target_dim)
def generate_text(self, seed_text='', max_length=None):
"""
Performs a prediction on the Language modeling objective given some seed text. It uses a noisy greedy decoding.
Temperature parameter for decoding is set in the config.
:param max_length: The maximum length to decode to.
:param seed_text: Defaults to the empty string. This will form the starting point to begin modelling
:return: A string containing the generated text.
"""
self.require_lm = True
encoded = self.encoder._encode([seed_text])
self._build_model(n_updates_total=0,
target_dim=self.target_dim,
train=False)
string = [self.encoder['_start_']] + encoded.token_ids[0]
EOS = self.encoder['_classify_']
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
for i in range(len(encoded.token_ids),
(max_length or self.config.max_length) - 1):
arr_encoded = self._array_format(encoded)
class_idx = self.sess.run(self.lm_predict_op, {
self.X: arr_encoded.token_ids,
self.M: arr_encoded.mask
})
string.append(class_idx[i])
if string[-1] == EOS:
break
return self.encoder.decode(string)
def __getstate__(self):
"""
Leave serialization of all tf objects to tf
"""
required_fields = [
'label_encoder',
'target_dim',
'_load_from_file',
'config',
'target_type',
]
serialized_state = {
k: v
for k, v in self.__dict__.items() if k in required_fields
}
return serialized_state
def save(self, path):
"""
Saves the state of the model to disk to the folder specific by `path`. If `path` does not exist, it will be auto-created.
Save is performed in two steps:
- Serialize tf graph to disk using tf.Saver
- Serialize python model using pickle
Note:
Does not serialize state of Adam optimizer.
Should not be used to save / restore a training model.
"""
if path is None:
return
path = os.path.abspath(path)
self.saver.save(self, path)
self._load_from_file = False
@classmethod
def load(cls, path):
"""
Load a saved fine-tuned model from disk. Path provided should be a folder which contains .pkl and tf.Saver() files
:param path: string path name to load model from. Same value as previously provided to :meth:`save`. Must be a folder.
"""
saver = Saver(JL_BASE)
model = saver.load(path)
model._initialize()
model.saver.variables = saver.variables
tf.reset_default_graph()
return model
@classmethod
def finetune_grid_search(cls,
Xs,
Y,
*,
test_size,
config=None,
eval_fn=None,
probs=False,
return_all=False):
"""
Performs grid search over config items defined using "GridSearchable" objects and returns either full results or
the config object that relates to the best results. The default config contains grid searchable objects for the
most important parameters to search over.
:param Xs: Input text. Either [num_samples] or [sequence, num_samples] for single or multi input models respectively.
:param Y: Targets, A list of targets, [num_samples] that correspond to each sample in Xs.
:param test_size: Int or float. If an int is given this number of samples is used to validate, if a float is
given then that fraction of samples is used.
:param config: A config object, or None to use the default config.
:param eval_fn: An eval function that takes 2 inputs (prediction, truth) and returns a float, with a max value being desired.
:param probs: If true, eval_fn is passed probability outputs from predict_proba, otherwise the output of predict is used.
:param return_all: If True, all results are returned, if False, only the best config is returned.
:return: default is to return the best config object. If return_all is true, it returns a list of tuples of the
form [(config, eval_fn output), ... ]
"""
if isinstance(Xs[0], str):
Xs = [Xs]
config = config or get_default_config()
config.val_size = 0.0
eval_fn = eval_fn or cls.get_eval_fn()
trainXs, testXs, trainY, testY = train_test_split(list_transpose(Xs),
Y,
test_size=test_size,
shuffle=True)
trainXs = list_transpose(trainXs)
testXs = list_transpose(testXs)
gs = config.get_grid_searchable()
ranged_keys = gs.keys()
ranged_iterators = gs.values()
grid_gen = itertools.product(*ranged_iterators)
results = []
for grid_item in grid_gen:
config_ = deepcopy(config)
config_.update(dict(zip(ranged_keys, grid_item)))
instance = cls(config=config_)
instance.finetune(*trainXs, Y=trainY)
if probs:
res = instance.predict_proba(*testXs)
else:
res = instance.predict(*testXs)
results.append((config_, eval_fn(res, testY)))
del instance
if return_all:
return results
return max(results, key=lambda x: x[1])[0]
@classmethod
def finetune_grid_search_cv(cls,
Xs,
Y,
*,
n_splits,
test_size,
config=None,
eval_fn=None,
probs=False,
return_all=False):
"""
Performs cross validated grid search over config items defined using "GridSearchable" objects and returns either full results or
the config object that relates to the best results. The default config contains grid searchable objects for the
most important parameters to search over.
It should be noted that the cv splits are not guaranteed unique, but each split is given to each set of hparams.
:param Xs: Input text. Either [num_samples] or [sequence, num_samples] for single or multi input models respectively.
:param Y: Targets, A list of targets, [num_samples] that correspond to each sample in Xs.
:param n_splits: Number of CV splits to do.
:param test_size: Int or float. If an int is given this number of samples is used to validate, if a float is
given then that fraction of samples is used.
:param config: A config object, or None to use the default config.
:param eval_fn: An eval function that takes 2 batches of outputs and returns a float, with a max value being
desired. An arithmetic mean must make sense for this metric.
:param probs: If true, eval_fn is passed probability outputs from predict_proba, otherwise the output of predict is used.
:param return_all: If True, all results are returned, if False, only the best config is returned.
:return: default is to return the best config object. If return_all is true, it returns a list of tuples of the
form [(config, eval_fn output), ... ]
"""
results = []
for _ in range(n_splits):
res = cls.finetune_grid_search(Xs,
Y,
test_size=test_size,
probs=probs,
eval_fn=eval_fn,
config=config,
return_all=True)
results.append(res)
results = list(zip(*results))
aggregated_results = []
for configuration in results:
config_common = None
sum_res = 0
n_res = 0
for config, result in configuration:
config_common = config_common or config
assert config == config_common
n_res += 1
sum_res += result
aggregated_results.append((config_common, sum_res / n_res))
if return_all:
return aggregated_results
return max(aggregated_results, key=lambda x: x[1])[0]
def __del__(self):
try:
if self.sess is not None:
self.sess.close()
except AttributeError:
pass