def class_avg_finetune(model,
texts,
labels,
nb_classes,
batch_size,
method,
epoch_size=5000,
nb_epochs=1000,
error_checking=True,
verbose=True):
""" Compiles and finetunes the given model.
# Arguments:
model: Model to be finetuned
texts: List of three lists, containing tokenized inputs for training,
validation and testing (in that order).
labels: List of three lists, containing labels for training,
validation and testing (in that order).
nb_classes: Number of classes in the dataset.
batch_size: Batch size.
method: Finetuning method to be used. For available methods, see
FINETUNING_METHODS in global_variables.py. Note that the model
should be defined accordingly (see docstring for deepmoji_transfer())
epoch_size: Number of samples in an epoch.
nb_epochs: Number of epochs. Doesn't matter much as early stopping is used.
error_checking: If set to True, warnings will be printed when the label
list has the wrong dimensions.
verbose: Verbosity flag.
# Returns:
Model after finetuning,
score after finetuning using the class average F1 metric.
"""
if method not in FINETUNING_METHODS:
raise ValueError('ERROR (class_avg_tune_trainable): '
'Invalid method parameter. '
'Available options: {}'.format(FINETUNING_METHODS))
(X_train, y_train) = (texts[0], labels[0])
(X_val, y_val) = (texts[1], labels[1])
(X_test, y_test) = (texts[2], labels[2])
checkpoint_path = '{}/deepmoji-checkpoint-{}.hdf5' \
.format(WEIGHTS_DIR, str(uuid.uuid4()))
f1_init_path = '{}/deepmoji-f1-init-{}.hdf5' \
.format(WEIGHTS_DIR, str(uuid.uuid4()))
# Check dimension of labels
if error_checking:
# Binary classification has two classes but one value
expected_shape = 1 if nb_classes == 2 else nb_classes
for ls in [y_train, y_val, y_test]:
if len(ls.shape) <= 1 or not ls.shape[1] == expected_shape:
print('WARNING (class_avg_tune_trainable): '
'The dimension of the provided '
'labels do not match the expected value. '
'Expected: {}, actual: {}'.format(
expected_shape, ls.shape[1]))
break
if method in ['last', 'new']:
lr = 0.001
elif method in ['full', 'chain-thaw']:
lr = 0.0001
loss = 'binary_crossentropy'
# Freeze layers if using last
if method == 'last':
model = freeze_layers(model, unfrozen_keyword='softmax')
# Compile model, for chain-thaw we compile it later (after freezing)
if method != 'chain-thaw':
adam = Adam(clipnorm=1, lr=lr)
model.compile(loss=loss, optimizer=adam, metrics=['accuracy'])
# Training
if verbose:
print('Method: {}'.format(method))
print('Classes: {}'.format(nb_classes))
if method == 'chain-thaw':
result = class_avg_chainthaw(model,
nb_classes=nb_classes,
train=(X_train, y_train),
val=(X_val, y_val),
test=(X_test, y_test),
batch_size=batch_size,
loss=loss,
epoch_size=epoch_size,
nb_epochs=nb_epochs,
checkpoint_weight_path=checkpoint_path,
f1_init_weight_path=f1_init_path,
verbose=verbose)
else:
result = class_avg_tune_trainable(
model,
nb_classes=nb_classes,
train=(X_train, y_train),
val=(X_val, y_val),
test=(X_test, y_test),
epoch_size=epoch_size,
nb_epochs=nb_epochs,
batch_size=batch_size,
init_weight_path=f1_init_path,
checkpoint_weight_path=checkpoint_path,
verbose=verbose)
return model, result
def class_avg_finetune(model, texts, labels, nb_classes, batch_size,
method, epoch_size=5000,
nb_epochs=1000, error_checking=True,
verbose=True):
""" Compiles and finetunes the given model.
# Arguments:
model: Model to be finetuned
texts: List of three lists, containing tokenized inputs for training,
validation and testing (in that order).
labels: List of three lists, containing labels for training,
validation and testing (in that order).
nb_classes: Number of classes in the dataset.
batch_size: Batch size.
method: Finetuning method to be used. For available methods, see
FINETUNING_METHODS in global_variables.py. Note that the model
should be defined accordingly (see docstring for deepmoji_transfer())
epoch_size: Number of samples in an epoch.
nb_epochs: Number of epochs. Doesn't matter much as early stopping is used.
error_checking: If set to True, warnings will be printed when the label
list has the wrong dimensions.
verbose: Verbosity flag.
# Returns:
Model after finetuning,
score after finetuning using the class average F1 metric.
"""
if method not in FINETUNING_METHODS:
raise ValueError('ERROR (class_avg_tune_trainable): '
'Invalid method parameter. '
'Available options: {}'.format(FINETUNING_METHODS))
(X_train, y_train) = (texts[0], labels[0])
(X_val, y_val) = (texts[1], labels[1])
(X_test, y_test) = (texts[2], labels[2])
checkpoint_path = '{}/deepmoji-checkpoint-{}.hdf5' \
.format(WEIGHTS_DIR, str(uuid.uuid4()))
f1_init_path = '{}/deepmoji-f1-init-{}.hdf5' \
.format(WEIGHTS_DIR, str(uuid.uuid4()))
# Check dimension of labels
if error_checking:
# Binary classification has two classes but one value
expected_shape = 1 if nb_classes == 2 else nb_classes
for ls in [y_train, y_val, y_test]:
if len(ls.shape) <= 1 or not ls.shape[1] == expected_shape:
print('WARNING (class_avg_tune_trainable): '
'The dimension of the provided '
'labels do not match the expected value. '
'Expected: {}, actual: {}'
.format(expected_shape, ls.shape[1]))
break
if method in ['last', 'new']:
lr = 0.001
elif method in ['full', 'chain-thaw']:
lr = 0.0001
loss = 'binary_crossentropy'
# Freeze layers if using last
if method == 'last':
model = freeze_layers(model, unfrozen_keyword='softmax')
# Compile model, for chain-thaw we compile it later (after freezing)
if method != 'chain-thaw':
adam = Adam(clipnorm=1, lr=lr)
model.compile(loss=loss, optimizer=adam, metrics=['accuracy'])
# Training
if verbose:
print('Method: {}'.format(method))
print('Classes: {}'.format(nb_classes))
if method == 'chain-thaw':
result = class_avg_chainthaw(model, nb_classes=nb_classes,
train=(X_train, y_train),
val=(X_val, y_val),
test=(X_test, y_test),
batch_size=batch_size, loss=loss,
epoch_size=epoch_size,
nb_epochs=nb_epochs,
checkpoint_weight_path=checkpoint_path,
f1_init_weight_path=f1_init_path,
verbose=verbose)
else:
result = class_avg_tune_trainable(model, nb_classes=nb_classes,
train=(X_train, y_train),
val=(X_val, y_val),
test=(X_test, y_test),
epoch_size=epoch_size,
nb_epochs=nb_epochs,
batch_size=batch_size,
init_weight_path=f1_init_path,
checkpoint_weight_path=checkpoint_path,
verbose=verbose)
return model, result