def test_freeze_layers(): """ Correct layers are frozen. """ model = deepmoji_transfer(5, 30) keyword = 'softmax' model = freeze_layers(model, unfrozen_keyword=keyword) for layer in model.layers: if layer is not None and len(layer.trainable_weights): if keyword in layer.name: assert layer.trainable else: assert not layer.trainable
def test_freeze_layers(): """ Correct layers are frozen. """ model = deepmoji_transfer(5, 30) keyword = 'softmax' model = freeze_layers(model, unfrozen_keyword=keyword) for layer in model.layers: if layer is not None and len(layer.trainable_weights): if keyword in layer.name: assert layer.trainable else: assert not layer.trainable
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