models.py

from ..imports import *
from .. import utils as U
from .wrn import create_wide_residual_network




PRETRAINED_RESNET50 = 'pretrained_resnet50'
PRETRAINED_MOBILENET = 'pretrained_mobilenet'
PRETRAINED_INCEPTION = 'pretrained_inception'
RESNET50 = 'resnet50'
MOBILENET = 'mobilenet'
INCEPTION = 'inception'
CNN = 'default_cnn'
WRN22 = 'wrn22'
PRETRAINED_MODELS = [PRETRAINED_RESNET50, PRETRAINED_MOBILENET, PRETRAINED_INCEPTION]
PREDEFINED_MODELS = PRETRAINED_MODELS + [RESNET50, MOBILENET, INCEPTION]
IMAGE_CLASSIFIERS = {
                     PRETRAINED_RESNET50: '50-layer Residual Network (pretrained on ImageNet)',
                     RESNET50:  '50-layer Resididual Network (randomly initialized)',
                     PRETRAINED_MOBILENET: 'MobileNet Neural Network (pretrained on ImageNet)',
                     MOBILENET:  'MobileNet Neural Network (randomly initialized)',
                     PRETRAINED_INCEPTION: 'Inception Version 3  (pretrained on ImageNet)',
                     INCEPTION:  'Inception Version 3 (randomly initialized)',
                     WRN22: '22-layer Wide Residual Network (randomly initialized)',
                     CNN : 'a default LeNet-like Convolutional Neural Network'}

def print_image_classifiers():
    for k,v in IMAGE_CLASSIFIERS.items():
        print("%s: %s" % (k,v))


def print_image_regression_models():
    for k,v in IMAGE_CLASSIFIERS.items():
        print("%s: %s" % (k,v))


def pretrained_datagen(data, name):
    if not data or not U.is_iter(data): return
    idg = data.image_data_generator
    if name == PRETRAINED_RESNET50:
        idg.preprocessing_function = pre_resnet50
        idg.ktrain_preproc = 'resnet50'
        idg.rescale=None
        idg.featurewise_center=False
        idg.samplewise_center=False
        idg.featurewise_std_normalization=False
        idg.samplewise_std_normalization=False
        idg.zca_whitening = False
    elif name == PRETRAINED_MOBILENET:
        idg.preprocessing_function = pre_mobilenet
        idg.ktrain_preproc = 'mobilenet'
        idg.rescale=None
        idg.featurewise_center=False
        idg.samplewise_center=False
        idg.featurewise_std_normalization=False
        idg.samplewise_std_normalization=False
        idg.zca_whitening = False
    elif name == PRETRAINED_INCEPTION:
        idg.preprocessing_function = pre_inception
        idg.ktrain_preproc = 'inception'
        idg.rescale=None
        idg.featurewise_center=False
        idg.samplewise_center=False
        idg.featurewise_std_normalization=False
        idg.samplewise_std_normalization=False
        idg.zca_whitening = False

    return



def image_classifier(name,
                     train_data,
                     val_data=None,
                     freeze_layers=None, 
                     metrics=['accuracy'],
                     optimizer_name = U.DEFAULT_OPT,
                     multilabel=None,
                     multigpu_number=None, 
                     pt_fc = [],
                     pt_ps = [],
                     verbose=1):

    """
    Returns a pre-trained ResNet50 model ready to be fine-tuned
    for multi-class classification. By default, all layers are
    trainable/unfrozen.


    Args:
        name (string): one of {'pretrained_resnet50', 'resnet50', 'default_cnn'}
        train_data (image.Iterator): train data. Note: Will be manipulated here!
        val_data (image.Iterator): validation data.  Note: Will be manipulated here!
        freeze_layers (int):  number of beginning layers to make untrainable
                            If None, then all layers except new Dense layers
                            will be frozen/untrainable.
        metrics (list):  metrics to use
        optimizer_name(str): name of Keras optimizer (e.g., 'adam', 'sgd')
        multilabel(bool):  If True, model will be build to support
                           multilabel classificaiton (labels are not mutually exclusive).
                           If False, binary/multiclassification model will be returned.
                           If None, multilabel status will be inferred from data.
        multigpu_number (int): Repicate model on this many GPUS.
                               Must either be None or greater than 1.
                               If greater than 1, must meet system specifications.
                               If None, model is not replicated on multiple GPUS.
        pt_fc (list of ints): number of hidden units in extra Dense layers
                                before final Dense layer of pretrained model.
                                Only takes effect if name in PRETRAINED_MODELS
        pt_ps (list of floats): dropout probabilities to use before
                                each extra Dense layer in pretrained model.
                                Only takes effect if name in PRETRAINED_MODELS
        verbose (int):         verbosity
    Return:
        model(Model):  the compiled model ready to be fine-tuned/trained

        
    """
    return image_model(name, train_data, val_data=val_data, freeze_layers=freeze_layers,
                       metrics=metrics, optimizer_name=optimizer_name, multilabel=multilabel,
                       multigpu_number=multigpu_number,
                       pt_fc=pt_fc, pt_ps=pt_ps, verbose=verbose)




def image_regression_model(name,
                          train_data,
                          val_data=None,
                          freeze_layers=None, 
                          metrics=['mae'],
                          optimizer_name = U.DEFAULT_OPT,
                          multigpu_number=None, 
                          pt_fc = [],
                          pt_ps = [],
                          verbose=1):

    """
    Returns a pre-trained ResNet50 model ready to be fine-tuned
    for multi-class classification. By default, all layers are
    trainable/unfrozen.


    Args:
        name (string): one of {'pretrained_resnet50', 'resnet50', 'default_cnn'}
        train_data (image.Iterator): train data. Note: Will be manipulated here!
        val_data (image.Iterator): validation data.  Note: Will be manipulated here!
        freeze_layers (int):  number of beginning layers to make untrainable
                            If None, then all layers except new Dense layers
                            will be frozen/untrainable.
        metrics (list):  metrics to use
        optimizer_name(str): name of Keras optimizer (e.g., 'adam', 'sgd')
        multilabel(bool):  If True, model will be build to support
                           multilabel classificaiton (labels are not mutually exclusive).
                           If False, binary/multiclassification model will be returned.
                           If None, multilabel status will be inferred from data.
        multigpu_number (int): Repicate model on this many GPUS.
                               Must either be None or greater than 1.
                               If greater than 1, must meet system specifications.
                               If None, model is not replicated on multiple GPUS.
        pt_fc (list of ints): number of hidden units in extra Dense layers
                                before final Dense layer of pretrained model.
                                Only takes effect if name in PRETRAINED_MODELS
        pt_ps (list of floats): dropout probabilities to use before
                                each extra Dense layer in pretrained model.
                                Only takes effect if name in PRETRAINED_MODELS
        verbose (int):         verbosity
    Return:
        model(Model):  the compiled model ready to be fine-tuned/trained

        
    """


    return image_model(name, train_data, val_data=val_data, freeze_layers=freeze_layers,
                       metrics=metrics, optimizer_name=optimizer_name, multilabel=False,
                       multigpu_number=multigpu_number,
                       pt_fc=pt_fc, pt_ps=pt_ps, verbose=verbose)



def image_model( name,
                 train_data,
                 val_data=None,
                 freeze_layers=None, 
                 metrics=['accuracy'],
                 optimizer_name = U.DEFAULT_OPT,
                 multilabel=None,
                 multigpu_number=None, 
                 pt_fc = [],
                 pt_ps = [],
                 verbose=1):

    """
    Returns a pre-trained ResNet50 model ready to be fine-tuned
    for multi-class classification or regression. By default, all layers are
    trainable/unfrozen.


    Args:
        name (string): one of {'pretrained_resnet50', 'resnet50', 'default_cnn'}
        train_data (image.Iterator): train data. Note: Will be manipulated here!
        val_data (image.Iterator): validation data.  Note: Will be manipulated here!
        freeze_layers (int):  number of beginning layers to make untrainable
                            If None, then all layers except new Dense layers
                            will be frozen/untrainable.
        metrics (list):  metrics to use
        optimizer_name(str): name of Keras optimizer (e.g., 'adam', 'sgd')
        multilabel(bool):  If True, model will be build to support
                           multilabel classificaiton (labels are not mutually exclusive).
                           If False, binary/multiclassification model will be returned.
                           If None, multilabel status will be inferred from data.
        multigpu_number (int): Repicate model on this many GPUS.
                               Must either be None or greater than 1.
                               If greater than 1, must meet system specifications.
                               If None, model is not replicated on multiple GPUS.
        pt_fc (list of ints): number of hidden units in extra Dense layers
                                before final Dense layer of pretrained model.
                                Only takes effect if name in PRETRAINED_MODELS
        pt_ps (list of floats): dropout probabilities to use before
                                each extra Dense layer in pretrained model.
                                Only takes effect if name in PRETRAINED_MODELS
        verbose (int):         verbosity
    Return:
        model(Model):  the compiled model ready to be fine-tuned/trained

        
    """
    # arg check
    U.data_arg_check(train_data=train_data, train_required=True)
    if name not in list(IMAGE_CLASSIFIERS.keys()):
        raise ValueError('Unknown or unsupported model: %s' % (name))


    if not U.is_iter(train_data):
        raise ValueError('train_data must be an Keras iterator ' +\
                         '(e.g., DirectoryIterator, DataframIterator, '+ \
                         'NumpyArrayIterator) - please use the ktrain.data.images_from* ' +\
                         'functions')



    # set pretrained flag
    pretrained = True if name in PRETRAINED_MODELS else False

    # adjust freeze_layers with warning
    if not pretrained and freeze_layers is not None and freeze_layers > 0:
        warnings.warn('Only pretrained models (e.g., pretrained_resnet50) support freeze_layers. ' +\
                      'Setting freeze_layers to 0. Use one of the following models if' +\
                      'desiring a model pretrained on ImageNet: %s' % (PRETRAINED_MODELS))
        freeze_layers = 0

    if pretrained and val_data is None:
        raise ValueError('val_data is required if selecting a pretrained model, '+\
                         'as normalization scheme will be altered.')

    # adjust the data augmentation based on model selected
    if pretrained:
        pretrained_datagen(train_data, name)
        pretrained_datagen(val_data, name)
        U.vprint('The normalization scheme has been changed for use with a %s' % (name) +\
                ' model. If you decide to use a different model, please reload your' +\
                ' dataset with a ktrain.vision.data.images_from* function.\n', verbose=verbose)

    # determine if multilabel
    if multilabel is None:
        multilabel = U.is_multilabel(train_data)
    is_regression=False
    if not multilabel and len(train_data[0][-1].shape) == 1: is_regression=True

    # set loss and acivations
    loss_func = 'categorical_crossentropy'
    activation = 'softmax'
    if multilabel:
        loss_func = 'binary_crossentropy'
        activation = 'sigmoid'
    elif is_regression:
        loss_func = 'mse'
        activation = None
        if metrics == ['accuracy']: metrics = ['mae']

    U.vprint("Is Multi-Label? %s" % (multilabel), verbose=verbose)
    U.vprint("Is Regression? %s" % (is_regression), verbose=verbose)


    # determine number of classes and shape
    num_classes = 1 if is_regression else U.nclasses_from_data(train_data)
    input_shape = U.shape_from_data(train_data)



    #------------
    # build model
    #------------
    if type(multigpu_number) == type(1) and multigpu_number < 2:
        raise ValuError('multigpu_number must either be None or > 1')
    if multigpu_number is not None and multigpu_number >1:
        with tf.device("/cpu:0"):
            model = build_visionmodel(name,
                                      num_classes,
                                      input_shape=input_shape,
                                      freeze_layers=freeze_layers,
                                      activation=activation,
                                      pt_fc = pt_fc,
                                      pt_ps = pt_ps)
        parallel_model = multi_gpu_model(model, gpus=multigpu_number)
        parallel_model.compile(optimizer=optimizer_name, 
                               loss='categorical_crossentropy', metrics=metrics)
        return parallel_model
    else:
        model = build_visionmodel(name,
                                  num_classes,
                                  input_shape=input_shape,
                                  freeze_layers=freeze_layers,
                                  activation=activation,
                                  pt_fc = pt_fc,
                                  pt_ps = pt_ps)
        model.compile(optimizer=optimizer_name, loss=loss_func, metrics=metrics)
        return model


def build_visionmodel(name,
                      num_classes, 
                      input_shape=(224,224,3),
                      freeze_layers=2, 
                      activation='softmax',
                      pt_fc=[],
                      pt_ps = []):

    if name in PREDEFINED_MODELS:
        model = build_predefined(name, num_classes,
                                  input_shape=input_shape,
                                  freeze_layers=freeze_layers,
                                  activation=activation,
                                  pt_fc = pt_fc,
                                  pt_ps = pt_ps)
    elif name == CNN:
        model = build_cnn(num_classes,
                           input_shape=input_shape,
                           activation=activation)
    elif name == WRN22:
        model = create_wide_residual_network(input_shape, nb_classes=num_classes, 
                                             N=3, k=6, dropout=0.00,
                                             activation=activation, verbose=0)
    else:
        raise ValueError('Unknown model: %s' % (name))
    U.vprint('%s model created.' % (name))

    return model



def build_cnn(num_classes, 
              input_shape=(28,28,1),
              activation='softmax'):
    model = Sequential()
    model.add(Conv2D(32, kernel_size=(3, 3),activation='relu',
                     kernel_initializer='he_normal',input_shape=input_shape))
    model.add(Conv2D(32, kernel_size=(3, 3),activation='relu',
                     kernel_initializer='he_normal'))
    model.add(MaxPooling2D((2, 2)))
    model.add(Dropout(0.20))
    model.add(Conv2D(64, (3, 3), activation='relu',padding='same',
                      kernel_initializer='he_normal'))
    model.add(Conv2D(64, (3, 3), activation='relu',padding='same',
                     kernel_initializer='he_normal'))
    model.add(MaxPooling2D(pool_size=(2, 2)))
    model.add(Dropout(0.25))
    model.add(Conv2D(128, (3, 3), activation='relu',padding='same',
                     kernel_initializer='he_normal'))
    model.add(Dropout(0.25))
    model.add(Flatten())
    model.add(Dense(128, activation='relu'))
    model.add(BatchNormalization())
    model.add(Dropout(0.25))
    model.add(Dense(num_classes, activation=activation))
    return model


def build_predefined(
                   name,
                   num_classes, 
                   input_shape=(224,224,3),
                   freeze_layers=None, 
                   activation='softmax',
                   pt_fc=[],
                   pt_ps=[]):
    """
    Builds a pre-defined architecture supported in Keras.

    Args:
        name (str): one of ktrain.vision.model.PREDEFINED_MODELS
        num_classes (int): # of classes
        input_shape (tuple): the input shape including channels
        freeze_layers (int): number of early layers to freeze.
                             Only takes effect if name in PRETRAINED_MODELS.
                             If None and name in PRETRAINED_MODELS,
                             all layers except the "custom head" 
                             fully-connected (Dense) layers are frozen.
        activation (str):    name of the Keras activation to use in final layer
        pt_fc (list of ints): number of hidden units in extra Dense layers
                                before final Dense layer of pretrained model
        pt_ps (list of floats): dropout probabilities to use before
                                      each extra Dense layer in pretrained model

    """

    # default parameters
    include_top = False
    input_tensor = None
    dropout = 0.5 # final dropout

    # setup pretrained
    weights = 'imagenet' if name in PRETRAINED_MODELS else None

    # setup the pretrained network
    if name in [RESNET50, PRETRAINED_RESNET50]:
        with warnings.catch_warnings():
            warnings.simplefilter('ignore')
            net = ResNet50(include_top=include_top, 
                           weights=weights,
                           input_tensor=input_tensor,
                           input_shape = input_shape)
    elif name in [MOBILENET, PRETRAINED_MOBILENET]:
        net = MobileNet(include_top=include_top, 
                        weights=weights,
                        input_tensor=input_tensor,
                        input_shape = input_shape)
    elif name in [INCEPTION, PRETRAINED_INCEPTION]:
        net = InceptionV3(include_top=include_top, 
                          weights=weights,
                          input_tensor=input_tensor,
                           input_shape = input_shape)
    else:
        raise ValueError('Unsupported model: %s' % (name))


    if freeze_layers is None:
        for layer in net.layers:
            layer.trainable = False

    x = net.output
    x = Flatten()(x)

    # xtra FCs in pretrained model
    if name in PRETRAINED_MODELS:
        if len(pt_fc) != len(pt_ps):
            raise ValueError('size off xtra_fc must match size of fc_dropouts')
        for i, fc in enumerate(pt_fc):
            p = pt_ps[i]
            fc_name = "fc%s" % (i)
            if p is not None:
                x = Dropout(p)(x)
            x = Dense(fc, activation='relu', 
                      kernel_initializer='he_normal', name=fc_name)(x)


    # final FC
    x = Dropout(dropout)(x)
    output_layer = Dense(num_classes, activation=activation, name=activation)(x)
    model = Model(inputs=net.input, outputs=output_layer)

    if freeze_layers is not None:
        # set certain earlier layers as non-trainable
        for layer in model.layers[:freeze_layers]:
            layer.trainable = False
        for layer in model.layers[freeze_layers:]:
            layer.trainable = True

    # set optimizer, loss, and metrics and return model
    return model