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()
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 __init__(self, config=None, **kwargs):
self.config = config or get_default_config()
self.config.override_from_dict(kwargs)
self.label_encoder = None
self._initialize()
self.target_dim = None
self._load_from_file = False
self.target_type = None
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]
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.
"""
self.config = config or get_default_config()
self.config.update(kwargs)
self.label_encoder = None
self._initialize()
self.target_dim = None
self._load_from_file = False
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]
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 __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
Download Stanford Sentiment Treebank to data directory
"""
path = Path(self.filename)
path.parent.mkdir(parents=True, exist_ok=True)
remote_url = "https://s3.amazonaws.com/enso-data/multinli.dev.csv"
response = requests.get(remote_url)
open(DATA_PATH, 'wb').write(response.content)
if __name__ == "__main__":
# Train and evaluate on MultiNLI
dataset = MultiNLI(nrows=10).dataframe
trainX1, testX1, trainX2, testX2, trainY, testY = train_test_split(
dataset.x1, dataset.x2, dataset.target, test_size=0.3, random_state=42)
base_conf = get_default_config()
res = MultiFieldClassifier.finetune_grid_search_cv(
[dataset.x1, dataset.x2],
dataset.target,
n_splits=2,
config=base_conf,
eval_fn=lambda y1, y2: np.mean(np.asarray(y1) == np.asarray(y2)),
test_size=0.1)
model = MultiFieldClassifier(res)
model.fit(trainX1, trainX2, Y=trainY)
acc = np.mean(
np.asarray(model.predict(testX1, testX2)) == np.asarray(testY))
print('Test Accuracy: {:0.2f} with config {}'.format(acc, res))
return CHECKSUM
def download(self):
"""
Download Stanford Sentiment Treebank to data directory
"""
path = Path(self.filename)
path.parent.mkdir(parents=True, exist_ok=True)
remote_url = "https://s3.amazonaws.com/enso-data/multinli.dev.csv"
response = requests.get(remote_url)
open(DATA_PATH, 'wb').write(response.content)
if __name__ == "__main__":
# Train and evaluate on MultiNLI
dataset = MultiNLI(nrows=10).dataframe
trainX1, testX1, trainX2, testX2, trainY, testY = train_test_split(
dataset.x1, dataset.x2, dataset.target, test_size=0.3, random_state=42
)
base_conf = get_default_config()
res = MultifieldClassifier.finetune_grid_search_cv([dataset.x1, dataset.x2], dataset.target, n_splits=2, config=base_conf,
eval_fn=lambda y1, y2: np.mean(np.asarray(y1) == np.asarray(y2)),
test_size=0.1)
model = MultifieldClassifier(res)
model.fit(trainX1, trainX2, Y=trainY)
acc = np.mean(np.asarray(model.predict(testX1, testX2)) == np.asarray(testY))
print('Test Accuracy: {:0.2f} with config {}'.format(acc, res))