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?
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 _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]
)
class BaseModel(object, metaclass=ABCMeta):
"""
A sklearn-style class for finetuning a Transformer language model on a classification task.
:param config: A config object, or None to use the default config.
:param **kwargs: key-value pairs of config items to override.
"""
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 _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.clf_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
# indicator vars
self.is_built = False # has tf graph been constructed?
self.is_trained = False # has model been fine-tuned?
def _text_to_ids(self, *Xs, 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
assert len(Xs) == 1, "This implementation assumes a single Xs"
encoder_out = self.encoder.encode_for_classification(Xs[0], max_length=max_length)
return self._array_format(encoder_out)
def _target_model(self, featurizer_state, targets, n_outputs, train=False, reuse=None, **kwargs):
# Conditionally constructs the default model for each of the main ways in which finetune can be used.
# Can be overridden to use a different target model.
if self.target_type == CLASSIFICATION:
return classifier(featurizer_state['features'], targets, n_outputs, self.do_dropout, config=self.config,
train=train, reuse=reuse, **kwargs)
elif self.target_type == REGRESSION:
return regressor(featurizer_state['features'], targets, n_outputs, self.do_dropout, config=self.config,
train=train, reuse=reuse, **kwargs)
elif self.target_type == SEQUENCE_LABELING:
return sequence_labeler(featurizer_state['sequence_features'], targets, n_outputs, self.do_dropout,
config=self.config,
train=train, reuse=reuse, **kwargs)
else:
raise InvalidTargetType(self.target_type)
def _predict_ops(self, logits, **kwargs):
# Gets the correct prediction methods for each of the main ways that the model can be used.
if self.target_type == CLASSIFICATION:
return tf.argmax(logits, -1), tf.nn.softmax(logits, -1)
elif self.target_type == REGRESSION:
return logits, tf.constant(0) # TODO: Find something better than constant 0 for predict proba.
elif self.target_type == SEQUENCE_LABELING:
return sequence_decode(logits, kwargs.get("transition_matrix"))
else:
raise InvalidTargetType(self.target_type)
def _get_target_encoder(self):
# Gets the correct target encoder for the problem.
if self.target_type == CLASSIFICATION:
return OneHotLabelEncoder()
elif self.target_type == REGRESSION:
return RegressionEncoder()
elif self.target_type == SEQUENCE_LABELING:
return SequenceLabelingEncoder()
else:
raise InvalidTargetType(self.target_type)
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 tf.no_op()
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, batch_size)
def _training_loop(self, arr_encoded, Y, batch_size=None):
batch_size = batch_size or self.config.batch_size
self.label_encoder = self._get_target_encoder()
n_batch_train = batch_size * max(len(get_available_gpus(self.config)), 1)
train_x, train_mask = arr_encoded.token_ids, arr_encoded.mask
n_examples = train_x.shape[0]
n_updates_total = (n_examples // n_batch_train) * self.config.n_epochs
if Y is not None:
Y = self.label_encoder.fit_transform(Y)
target_dim = len(self.label_encoder.target_dim)
else:
# only language model will be trained, mock fake target
Y = [[None]] * n_examples
target_dim = None
self._build_model(n_updates_total=n_updates_total, target_dim=target_dim)
dataset = (train_x, train_mask, Y)
x_tr, x_va, m_tr, m_va, y_tr, y_va = train_test_split(*dataset, test_size=self.config.val_size,
random_state=self.config.seed)
dataset = (x_tr, m_tr, y_tr)
val_dataset = (x_va, m_va, y_va)
self.is_trained = True
avg_train_loss = 0
avg_val_loss = 0
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):
for xmb, mmb, ymb in iter_data(*dataset, n_batch=n_batch_train, verbose=self.config.verbose):
global_step += 1
if global_step % self.config.val_interval == 0:
tqdm.tqdm.write("Train loss is :{}, Val loss is :{}".format(avg_train_loss, avg_val_loss))
outputs = self._eval(
self.summaries,
feed_dict={
self.X: xmb,
self.M: mmb,
self.Y: ymb,
self.do_dropout: DROPOUT_OFF
}
)
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"):
outputs = self._eval(
self.clf_loss,
self.summaries,
feed_dict={
self.X: xval,
self.M: mval,
self.Y: yval,
self.do_dropout: DROPOUT_OFF
}
)
if self.valid_writer is not None:
self.valid_writer.add_summary(outputs.get(self.summaries), global_step)
val_cost = outputs.get(self.clf_loss, 0)
sum_val_loss += val_cost
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.save_best_model:
self.save(self.config.autosave_path)
outputs = self._eval(
self.clf_loss,
self.train_op,
feed_dict={
self.X: xmb,
self.M: mmb,
self.Y: ymb,
self.do_dropout: DROPOUT_ON
}
)
cost = outputs.get(self.clf_loss, 0)
avg_train_loss = avg_train_loss * self.config.rolling_avg_decay + cost * (
1 - self.config.rolling_avg_decay)
return self
@abstractmethod
def finetune(self, *args, **kwargs):
""" The base method for finetuning the model. """
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
}
)
class_idx = output.get(self.predict_op)
class_labels = self.label_encoder.inverse_transform(class_idx)
predictions.append(class_labels)
return np.concatenate(predictions).tolist()
@abstractmethod
def predict(self, *args, **kwargs):
""" The base method for predicting from the model. """
def _predict_proba(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_proba_op,
feed_dict={
self.X: xmb,
self.M: mmb,
self.do_dropout: DROPOUT_OFF
}
)
probas = output.get(self.predict_proba_op)
classes = self.label_encoder.target_dim
if self.target_type == ANNOTATION:
# sequence predictions
predictions.extend([
dict(zip(classes, proba.tolist()))
for proba in probas
])
elif self.target_type == CLASSIFICATION:
# sequence predictions
predictions.extend([
[
dict(zip(classes, proba_t.tolist()))
for proba_t in proba_seq
] for proba_seq in probas
])
else:
raise AssertionError(
"Target type `{}` does not support the predict_proba method".format(self.target_type))
)
return predictions
@abstractmethod
def predict_proba(self, *args, **kwargs):
""" Base method for predicting, with probabilites. (when available) """
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.
"""
def transform(self, *args, **kwargs):
"""
An alias for `featurize`.
"""
return self.featurize(*args, **kwargs)
def _infer_prep(self, *X, max_length=None):
max_length = max_length or self.config.max_length
arr_encoded = self._text_to_ids(*X, max_length=max_length)
n_batch_train = self.config.batch_size * max(len(get_available_gpus(self.config)), 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 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
)
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.
"""
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 _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?
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
Xs = self._format_for_encoding(*Xs)
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):
if len(Xs) > 1 and Y is None:
Y = Xs[-1]
Xs = Xs[:-1]
fit_language_model_only = (Y is None)
arr_encoded = self._text_to_ids(*Xs)
labels = None if fit_language_model_only else Y
return self._training_loop(
arr_encoded,
Y=labels,
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):
for (xmb, mmb, ymb) in iter_data(*train_dataset, n_batch=n_batch_train, verbose=self.config.verbose):
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.save_best_model:
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)
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, initial_params):
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=initial_params,
deviation_regularization=self.config.regularize_deviation
)
def _construct_graph(self, n_updates_total, target_dim=None, train=True, pre_trained_weights=None):
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]
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
)
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
)
lm_loss_coef = self.config.lm_loss_coef
if target_dim is None:
lm_loss_coef = 1.0
train_loss = lm_loss_coef * tf.reduce_mean(language_model_state['losses'])
aggregator['features'].append(featurizer_state['features'])
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))
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.lm_predict_op = tf.concat(aggregator["lm_model"], 0)
self.features = tf.concat(aggregator['features'], axis=0)
self.lm_losses = tf.concat(aggregator['lm_losses'], axis=0)
if train:
self._compile_train_op(
params=params,
grads=gpu_grads,
n_updates_total=n_updates_total,
initial_params=pre_trained_weights
)
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.lm_loss = tf.reduce_mean(self.lm_losses)
self.summaries.append(tf.summary.scalar('TargetModelLoss', self.target_loss))
self.summaries.append(tf.summary.scalar('LanguageModelLoss', self.lm_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.
"""
pre_trained_weights = self._load_base_model()
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, pre_trained_weights=pre_trained_weights, train=train)
self._initialize_session()
# Optionally load saved model
if self._load_from_file:
self._load_finetuned_model()
elif not self.is_trained:
self._init_from_pretrained(pre_trained_weights["init_params"])
# Must be set after loading saved models -- we used the stored value of target_dim to
# infer what portions of the graph to attempt to load from the checkpoint
self.target_dim = target_dim
guarantee_initialized_variables(self.sess, keys=['model/target', 'adam'])
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.
"""
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 _load_base_model(self):
"""
Load serialized base model parameters from file.
"""
with open(SHAPES_PATH) as shapes_file:
shapes = json.load(shapes_file)
offsets = np.cumsum([np.prod(shape) for shape in shapes])
init_params = [np.load(PARAM_PATH.format(n)) for n in range(10)]
init_params = np.split(np.concatenate(init_params, 0), offsets)[:-1]
init_params = [param.reshape(shape) for param, shape in zip(init_params, shapes)]
if self.config.interpolate_pos_embed:
init_params[0] = interpolate_pos_embed(init_params[0], self.config.max_length)
elif self.config.max_length > 512:
raise ValueError("Max Length cannot be greater than 512 if interploate_pos_embed is turned off")
else:
init_params[0] = init_params[0][:self.config.max_length]
special_embed = (np.random.randn(len(self.encoder.special_tokens),
self.config.n_embed) * self.config.weight_stddev).astype(np.float32)
reg_mask = np.concatenate(
[np.ones_like(init_params[1]), np.zeros_like(special_embed), np.ones_like(init_params[0])], 0)
init_params[0] = np.concatenate([init_params[1], special_embed, init_params[0]], 0)
del init_params[1]
return {
"init_params": init_params,
"mask": itertools.chain([reg_mask], itertools.repeat(1.0, len(init_params) - 1))
}
def _init_from_pretrained(self, init_params):
""" Load pre-trained weights into the tensors """
pretrained_params = find_trainable_variables("model", exclude="model/clf")
self.sess.run(tf.global_variables_initializer())
self.sess.run([p.assign(ip) for p, ip in zip(pretrained_params, init_params)])
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
# Setting self._load_from_file indicates that we should load the saved model from
# disk at next training / inference. It is set temporarily so that the serialized
# model includes this information.
path = os.path.abspath(path)
self._load_from_file = path
saver = tf.train.Saver(tf.trainable_variables())
path_obj = Path(path)
if path_obj.exists():
if not path_obj.is_dir():
raise ValueError("Invalid save location: {}. A file already exists at this location.".format(path))
else:
path_obj.mkdir(parents=True)
saver.save(self.sess, os.path.join(path, SAVE_PREFIX))
pickle.dump(self, open(
os.path.join(self._load_from_file, '{}.pkl'.format(SAVE_PREFIX)), 'wb')
)
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.
"""
pickle_path = os.path.join(path, '{}.pkl'.format(SAVE_PREFIX))
model = pickle.load(open(pickle_path, 'rb'))
model._initialize()
tf.reset_default_graph()
return model
def _load_finetuned_model(self):
var_list = find_trainable_variables('model', exclude='model/target')
if self.target_dim is not None:
var_list.extend(find_trainable_variables('model/target'))
saver = tf.train.Saver(var_list=var_list)
saver.restore(self.sess, os.path.join(self._load_from_file, SAVE_PREFIX))
self._load_from_file = False
self.is_trained = True
from abc import ABCMeta, abstractmethod
import tqdm
import numpy as np
import tensorflow as tf
from tensorflow.python.data import Dataset
from sklearn.model_selection import train_test_split
from finetune.errors import FinetuneError
from finetune.config import PAD_TOKEN
from finetune.encoding import TextEncoder, ArrayEncodedOutput, EncodedOutput
from finetune.imbalance import compute_class_weights
ENCODER = TextEncoder()
LOGGER = logging.getLogger('finetune')
class BasePipeline(metaclass=ABCMeta):
def __init__(self, config):
self.config = config
self.label_encoder = None
self.target_dim = None
self.pad_idx_ = None
self.rebuild = False
self.epoch = 0
@abstractmethod
def _target_encoder(self):
# Overridden by subclass to produce the right target encoding for a given target model.
class BaseModel(object, metaclass=ABCMeta):
"""
A sklearn-style class for finetuning a Transformer language model on a classification task.
:param config: A config object, or None to use the default config.
:param **kwargs: key-value pairs of config items to override.
"""
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 _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.clf_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
# indicator vars
self.is_built = False # has tf graph been constructed?
self.is_trained = False # has model been fine-tuned?
def _text_to_ids(self, *Xs, 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
assert len(Xs) == 1, "This implementation assumes a single Xs"
token_idxs = self.encoder.encode_for_classification(
Xs[0], max_length=max_length)
seq_array = self._array_format(token_idxs)
return seq_array.token_ids, seq_array.mask
def _target_model(self,
featurizer_state,
targets,
n_outputs,
train=False,
reuse=None,
**kwargs):
# Conditionally constructs the default model for each of the main ways in which finetune can be used.
# Can be overridden to use a different target model.
if self.target_type == CLASSIFICATION:
return classifier(featurizer_state['features'],
targets,
n_outputs,
self.do_dropout,
config=self.config,
train=train,
reuse=reuse,
**kwargs)
elif self.target_type == REGRESSION:
return regressor(featurizer_state['features'],
targets,
n_outputs,
self.do_dropout,
config=self.config,
train=train,
reuse=reuse,
**kwargs)
elif self.target_type == SEQUENCE_LABELING:
return sequence_labeler(featurizer_state['sequence_features'],
targets,
n_outputs,
self.do_dropout,
config=self.config,
train=train,
reuse=reuse,
**kwargs)
else:
raise InvalidTargetType(self.target_type)
def _predict_ops(self, logits, **kwargs):
# Gets the correct prediction methods for each of the main ways that the model can be used.
if self.target_type == CLASSIFICATION:
return tf.argmax(logits, -1), tf.nn.softmax(logits, -1)
elif self.target_type == REGRESSION:
return logits, tf.constant(
0
) # TODO: Find something better than constant 0 for predict proba.
elif self.target_type == SEQUENCE_LABELING:
return sequence_decode(logits, kwargs.get("transition_matrix"))
else:
raise InvalidTargetType(self.target_type)
def _get_target_encoder(self):
# Gets the correct target encoder for the problem.
if self.target_type == CLASSIFICATION:
return OneHotLabelEncoder()
elif self.target_type == REGRESSION:
return RegressionEncoder()
elif self.target_type == SEQUENCE_LABELING:
return SequenceLabelingEncoder()
else:
raise InvalidTargetType(self.target_type)
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 tf.no_op() 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):
train_x, train_mask = self._text_to_ids(*Xs)
return self._training_loop(train_x, train_mask, Y, batch_size)
def _training_loop(self, train_x, train_mask, Y, batch_size=None):
batch_size = batch_size or self.config.batch_size
self.label_encoder = self._get_target_encoder()
n_batch_train = batch_size * max(len(get_available_gpus(self.config)),
1)
n_examples = train_x.shape[0]
n_updates_total = (n_examples // n_batch_train) * self.config.n_epochs
if Y is not None:
Y = self.label_encoder.fit_transform(Y)
target_dim = len(self.label_encoder.target_dim)
else:
# only language model will be trained, mock fake target
Y = [[None]] * n_examples
target_dim = None
self._build_model(n_updates_total=n_updates_total,
target_dim=target_dim)
dataset = (train_x, train_mask, Y)
x_tr, x_va, m_tr, m_va, y_tr, y_va = train_test_split(
*dataset,
test_size=self.config.val_size,
random_state=self.config.seed)
dataset = (x_tr, m_tr, y_tr)
val_dataset = (x_va, m_va, y_va)
self.is_trained = True
avg_train_loss = 0
avg_val_loss = 0
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):
for xmb, mmb, ymb in iter_data(*dataset,
n_batch=n_batch_train,
verbose=self.config.verbose):
global_step += 1
if global_step % self.config.val_interval == 0:
tqdm.tqdm.write(
"Train loss is :{}, Val loss is :{}".format(
avg_train_loss, avg_val_loss))
outputs = self._eval(self.summaries,
feed_dict={
self.X: xmb,
self.M: mmb,
self.Y: ymb,
self.do_dropout: DROPOUT_OFF
})
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"):
outputs = self._eval(self.clf_loss,
self.summaries,
feed_dict={
self.X: xval,
self.M: mval,
self.Y: yval,
self.do_dropout: DROPOUT_OFF
})
if self.valid_writer is not None:
self.valid_writer.add_summary(
outputs.get(self.summaries), global_step)
val_cost = outputs.get(self.clf_loss, 0)
sum_val_loss += val_cost
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.save_best_model:
self.save(self.config.autosave_path)
outputs = self._eval(self.clf_loss,
self.train_op,
feed_dict={
self.X: xmb,
self.M: mmb,
self.Y: ymb,
self.do_dropout: DROPOUT_ON
})
cost = outputs.get(self.clf_loss, 0)
avg_train_loss = avg_train_loss * self.config.rolling_avg_decay + cost * (
1 - self.config.rolling_avg_decay)
return self
@abstractmethod
def finetune(self, *args, **kwargs):
""" The base method for finetuning the model. """
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
})
class_idx = output.get(self.predict_op)
class_labels = self.label_encoder.inverse_transform(class_idx)
predictions.append(class_labels)
return np.concatenate(predictions).tolist()
@abstractmethod
def predict(self, *args, **kwargs):
""" The base method for predicting from the model. """
def _predict_proba(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_proba_op,
feed_dict={
self.X: xmb,
self.M: mmb,
self.do_dropout: DROPOUT_OFF
})
probas = output.get(self.predict_proba_op)
classes = self.label_encoder.target_dim
predictions.extend(
[dict(zip(classes, proba)) for proba in probas])
return predictions
@abstractmethod
def predict_proba(self, *args, **kwargs):
""" Base method for predicting, with probabilites. (when available) """
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
})
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.
"""
def transform(self, *args, **kwargs):
"""
An alias for `featurize`.
"""
return self.featurize(*args, **kwargs)
def _infer_prep(self, *X, max_length=None):
max_length = max_length or self.config.max_length
infer_x, infer_mask = self._text_to_ids(*X, max_length=max_length)
n_batch_train = self.config.batch_size * max(
len(get_available_gpus(self.config)), 1)
self._build_model(n_updates_total=0,
target_dim=self.target_dim,
train=False)
yield from iter_data(infer_x,
infer_mask,
n_batch=n_batch_train,
verbose=self.config.verbose)
def _array_format(self, token_idxs, labels=None):
"""
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(token_idxs)
seq_lengths = [len(x) for x in token_idxs]
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 labels else None
for i, seq_length in enumerate(seq_lengths):
# BPE embedding
x[i, :seq_length, 0] = token_idxs[i]
# masking: value of 1 means "consider this in cross-entropy LM loss"
mask[i, 1:seq_length] = 1
if labels:
labels_arr[i, :seq_length] = labels[i]
# positional_embeddings
x[:, :,
1] = np.arange(self.encoder.vocab_size,
self.encoder.vocab_size + self.config.max_length)
return SequenceArray(token_ids=x, mask=mask, labels=labels_arr)
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)
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.target_dim = target_dim
self._define_placeholders()
aggregator = defaultdict(list)
train_loss_tower = 0
gpus = get_available_gpus(self.config)
n_splits = max(len(gpus), 1)
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=gpus[0]))
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)
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)
lm_loss_coef = self.config.lm_loss_coef
if target_dim is None:
lm_loss_coef = 1.0
train_loss = lm_loss_coef * tf.reduce_mean(
language_model_state['losses'])
aggregator['features'].append(featurizer_state['features'])
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))
if target_dim is not None:
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
params = find_trainable_variables("model")
grads = tf.gradients(train_loss, params)
grads = list(zip(grads, params))
gpu_grads.append(grads)
aggregator['logits'].append(target_model_state['logits'])
aggregator['clf_losses'].append(
target_model_state['losses'])
self.lm_predict_op = tf.concat(aggregator["lm_model"], 0)
self.features = tf.concat(aggregator['features'], axis=0)
self.lm_losses = tf.concat(aggregator['lm_losses'], axis=0)
if target_dim is not None:
self.logits = tf.concat(aggregator['logits'], axis=0)
self.clf_losses = concat_or_stack(aggregator['clf_losses'])
self.predict_op, self.predict_proba_op = self._predict_ops(
self.logits, **target_model_state.get("predict_params", {}))
self._compile_train_op(params=params,
grads=gpu_grads,
n_updates_total=n_updates_total)
self.clf_loss = tf.reduce_mean(self.clf_losses)
self.lm_loss = tf.reduce_mean(self.lm_losses)
self.summaries.append(
tf.summary.scalar('TargetModelLoss', self.clf_loss))
self.summaries.append(
tf.summary.scalar('LanguageModelLoss', self.lm_loss))
self.summaries.append(
tf.summary.scalar('TotalLoss', train_loss_tower / n_splits))
self.summaries = tf.summary.merge(self.summaries)
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)
# Optionally load saved model
if self._load_from_file:
self._load_finetuned_model()
elif not self.is_trained:
self._load_base_model()
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):
gpus = get_available_gpus(self.config)
os.environ['CUDA_VISIBLE_DEVICES'] = ",".join(
[str(gpu) for gpu in gpus])
conf = tf.ConfigProto(allow_soft_placement=True)
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 _define_placeholders(self):
# 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.Y = tf.stop_gradient(
tf.placeholder(
tf.float32,
[None, self.target_dim or 1])) # classification targets
self.do_dropout = tf.placeholder(
tf.float32) # 1 for do dropout and 0 to not do dropout
if self.target_type == SEQUENCE_LABELING:
self.Y = tf.placeholder(
tf.int32,
[None, self.config.max_length]) # classification targets
else:
self.Y = tf.placeholder(
tf.float32, [None, self.target_dim]) # classification targets
def generate_text(self, max_length=None, seed_text=''):
"""
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.
"""
seed_text_tokens = self.encoder._encode([seed_text]).token_ids
self._build_model(n_updates_total=0,
target_dim=self.target_dim,
train=False)
string = [self.encoder['_start_']] + seed_text_tokens[0]
eos = self.encoder['_classify_']
with warnings.catch_warnings():
warnings.filterwarnings("ignore")
for i in range(len(seed_text_tokens[0]),
(max_length or self.config.max_length) - 1):
model_input = self._array_format([string])
class_idx = self.sess.run(self.lm_predict_op, {
self.X: model_input.token_ids,
self.M: model_input.mask
})
string.append(class_idx[i])
if string[-1] == eos:
break
return self.encoder.decode(string)
def _load_base_model(self):
"""
Load serialized base model parameters into tf Tensors
"""
pretrained_params = find_trainable_variables('model',
exclude='model/clf')
self._initialize_session()
self.sess.run(tf.global_variables_initializer())
with open(SHAPES_PATH) as shapes_file:
shapes = json.load(shapes_file)
offsets = np.cumsum([np.prod(shape) for shape in shapes])
init_params = [np.load(PARAM_PATH.format(n)) for n in range(10)]
init_params = np.split(np.concatenate(init_params, 0),
offsets)[:-1]
init_params = [
param.reshape(shape)
for param, shape in zip(init_params, shapes)
]
init_params[0] = init_params[0][:self.config.max_length]
special_embed = (np.random.randn(len(self.encoder.special_tokens),
self.config.n_embed) *
self.config.weight_stddev).astype(np.float32)
init_params[0] = np.concatenate(
[init_params[1], special_embed, init_params[0]], 0)
del init_params[1]
self.sess.run([
p.assign(ip) for p, ip in zip(pretrained_params, init_params)
])
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 in a location with name :param: path. The model is saved in several files
and :param: path is used as a prefix.
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
# Setting self._load_from_file indicates that we should load the saved model from
# disk at next training / inference. It is set temporarily so that the serialized
# model includes this information.
self._load_from_file = path
saver = tf.train.Saver(tf.trainable_variables())
saver.save(self.sess, path)
pickle.dump(self, open(self._load_from_file + '.pkl', 'wb'))
self._load_from_file = False
@classmethod
def load(cls, path):
"""
Load a saved fine-tuned model from disk.
:param path: string path name to load model from. Same value as previously provided to :meth:`save`.
"""
if not path.endswith('.pkl'):
path += '.pkl'
model = pickle.load(open(path, 'rb'))
model._initialize()
tf.reset_default_graph()
return model
def _load_finetuned_model(self):
self._initialize_session()
saver = tf.train.Saver(tf.trainable_variables())
saver.restore(self.sess, self._load_from_file)
self._load_from_file = False
self.is_trained = True
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