import os
import random
import weakref
import atexit
import warnings
import itertools
import math
from abc import ABCMeta, abstractmethod
from copy import deepcopy
import tempfile
import time
import sys
import shutil
import glob
from contextlib import contextmanager
import pathlib
import logging
import tqdm
import numpy as np
import tensorflow as tf
from tensorflow.data import Dataset
from tensorflow.contrib.distribute import OneDeviceStrategy
from sklearn.model_selection import train_test_split
import joblib
import finetune
from finetune.util import list_transpose
from finetune.encoding.input_encoder import EncodedOutput
from finetune.config import get_config, all_gpus, assert_valid_config
from finetune.saver import Saver, InitializeHook
from finetune.errors import FinetuneError
from finetune.model import get_model_fn, PredictMode
from finetune.util.download import download_data_if_required
from finetune.util.estimator import PatchedParameterServerStrategy
from finetune.util.positional_embeddings import embedding_preprocessor
from finetune.base_models import GPTModel, GPTModelSmall
LOGGER = logging.getLogger('finetune')
class BaseModel(object, metaclass=ABCMeta):
"""
A sklearn-style task agnostic base class for finetuning a Transformer language model.
"""
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)
self.config = get_config(**kwargs)
self.resolved_gpus = None
self.validate_config()
self.input_pipeline = self._get_input_pipeline()
download_data_if_required(self.config.base_model)
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
)
@abstractmethod
def _predict_op(self, logits, **kwargs):
raise NotImplementedError
@abstractmethod
def _predict_proba_op(self, logits, **kwargs):
raise NotImplementedError
@staticmethod
@abstractmethod
def _target_model(*, 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 _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 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=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
)
)
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)
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()
num_gpus = len(resolved_gpus)
if num_gpus > 1:
distribute_strategy = PatchedParameterServerStrategy(
visible_gpus=resolved_gpus
)
elif num_gpus == 1:
gpu = resolved_gpus[0]
distribute_strategy = OneDeviceStrategy(device='/gpu:{}'.format(gpu))
else:
distribute_strategy = OneDeviceStrategy(device='/cpu:0')
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,
)
conf.gpu_options.per_process_gpu_memory_fraction = (
self.config.per_process_gpu_memory_fraction
)
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):
config = self._get_estimator_config()
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,
encoder=self.input_pipeline.text_encoder,
target_dim=self.input_pipeline.target_dim,
label_encoder=self.input_pipeline.label_encoder,
saver=self.saver,
build_explain=build_explain
)
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 get_separate_estimators(self, force_build_lm = False):
fns = get_separate_model_fns(
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,
encoder=self.input_pipeline.text_encoder,
target_dim=self.input_pipeline.target_dim,
label_encoder=self.input_pipeline.label_encoder,
saver=self.saver
)
featurizer_est = tf.estimator.Estimator(
model_dir=self.estimator_dir,
model_fn=fns['featurizer_model_fn'],
config=config,
params=self.config
)
target_est = tf.estimator.Estimator(
model_dir=self.estimator_dir,
model_fn=fns['target_model_fn'],
config=config,
params=self.config
)
hooks = [InitializeHook(self.saver)]
return featurizer_est, target_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):
"""
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 tqdm.tqdm(range(n), total=n, desc="Inference"):
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):
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)
else:
input_fn = self.input_pipeline.get_predict_input_fn(Xs)
estimator, hooks = self.get_estimator(build_explain=PredictMode.EXPLAIN in predict_keys)
length = len(Xs) if not callable(Xs) else None
predictions = tqdm.tqdm(
estimator.predict(
input_fn=input_fn, predict_keys=predict_keys, hooks=hooks
),
total=length,
desc="Inference"
)
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):
raw_preds = self._inference(Xs, predict_keys=[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, predict_keys=[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 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):
raw_preds = self._inference(Xs, predict_keys=[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 = self.input_pipeline.text_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 = [self.input_pipeline.text_encoder.start] if use_extra_toks else []
encoded = EncodedOutput(token_ids=start + encoded.token_ids[0])
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.clf_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[i])
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'
]
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)
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:
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
@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 __del__(self):
if hasattr(self, '_tmp_dir') and self._tmp_dir is not None:
self._tmp_dir.cleanup()