def build_model(self):
if (self.tl):
base_model = efn.EfficientNetB0(weights="imagenet",input_shape=(self.input_shape[0],self.input_shape[1],3), include_top=False)
else:
base_model = efn.EfficientNetB0(weights=None,
input_shape=(self.input_shape[0], self.input_shape[1], 3),
include_top=False)
output = keras.layers.GlobalAveragePooling2D()(base_model.output)
output = keras.layers.Dense(32, activation='relu')(output)
# Nueva capa de salida
output = keras.layers.Dense(self.classes, activation='softmax')(output) # cambiar cantidad de clases
model = Model(inputs=base_model.input, outputs=output)
# Entrenar con nuevos datos
model.compile(optimizer='Adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])
return model
def load_efficientnet():
model = efn.EfficientNetB0(weights='imagenet',
include_top=False,
input_tensor=Input(shape=(config.IMAGE_HEIGHT,
config.IMAGE_WIDTH,
3)))
return model
def effcient0_unet(input_shape, filters=512, out_dim=4, act="sigmoid"):
model = efn.EfficientNetB0(include_top=False,
input_shape=input_shape,
weights=None)
blockid = [
'block2a_expand_activation', 'block3a_expand_activation',
'block4a_expand_activation', 'block6a_expand_activation',
'block7a_project_bn'
]
block_outputs = []
for i in range(len(blockid)):
blockout = model.get_layer(blockid[i]).output
block_outputs.append(blockout)
input = model.input
inputc = Conv2D(16, (3, 3), padding="same", activation="relu")(input)
x = block_outputs.pop()
for i in range(len(blockid) - 1):
x = transconv_block(x,
int(filters / (2**i)),
3,
2,
skip=block_outputs.pop())
x = transconv_block(x, 16, 3, 2, skip=inputc)
output_layer = Conv2D(out_dim, (1, 1), padding="same", activation=act)(x)
model = Model(input, output_layer)
return model
def get_model(model='b2', shape=(320,320)):
K.clear_session()
h,w = shape
if model == 'b0':
base_model = efn.EfficientNetB0(weights='imagenet', include_top=False, pooling='avg', input_shape=(h, w, 3))
elif model == 'b1':
base_model = efn.EfficientNetB1(weights='imagenet', include_top=False, pooling='avg', input_shape=(h, w, 3))
elif model == 'b2':
base_model = efn.EfficientNetB2(weights='imagenet', include_top=False, pooling='avg', input_shape=(h, w, 3))
elif model == 'b3':
base_model = efn.EfficientNetB3(weights='imagenet', include_top=False, pooling='avg', input_shape=(h, w, 3))
elif model == 'b4':
base_model = efn.EfficientNetB4(weights='imagenet', include_top=False, pooling='avg', input_shape=(h, w, 3))
elif model == 'b5':
base_model = efn.EfficientNetB5(weights='imagenet', include_top=False, pooling='avg', input_shape=(h, w, 3))
elif model == 'b6':
base_model = efn.EfficientNetB6(weights='imagenet', include_top=False, pooling='avg', input_shape=(h, w, 3))
else:
base_model = efn.EfficientNetB7(weights='imagenet', include_top=False, pooling='avg', input_shape=(h, w, 3))
x = base_model.output
y_pred = Dense(4, activation='sigmoid')(x)
return Model(inputs=base_model.input, outputs=y_pred)
def model_fn(FLAGS, objective, optimizer, metrics):
# base_model = efn.EfficientNetB3(include_top=False,
# input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
# classes=FLAGS.num_classes, )
# # input_size = 380
model = efn.EfficientNetB0(
include_top=False,
input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
classes=FLAGS.num_classes,
)
# # input_size = 456
# base_model = efn.EfficientNetB5(include_top=False,
# input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
# classes=FLAGS.num_classes, )
# # input_size = 528
# base_model = efn.EfficientNetB6(include_top=False,
# input_shape=(FLAGS.input_size, FLAGS.input_size, 3),
# classes=FLAGS.num_classes, )
x = model.output
# 插入双线性池化操作
# x = bilinear_pooling(x）
# x = GlobalAveragePooling2D()(x)
# x = Dropout(0.2)(x)
predictions = Dense(FLAGS.num_classes, activation='softmax')(
x) # activation="linear",activation='softmax'
model = Model(input=model.input, output=predictions)
p_model = multi_gpu_model(model, gpus=2)
p_model.compile(loss=objective, optimizer=optimizer, metrics=metrics)
p_model.summary()
return model, p_model
def build_basicRegressionCNN(self):
"""Builds basic convolution-only model with MSE loss function
-you dont need dense connections - thats what inception thought us
Returns:
keras.models.Model
:return:
"""
print("build_basicCNN input_shape:" + str(self.input_shape))
rgb_efficientNetB0 = efn.EfficientNetB0(include_top=False,
weights='imagenet',
input_shape=self.input_shape,
classes=1)
z = rgb_efficientNetB0.output
z = GlobalMaxPooling2D()(z)
z = Dense(1, activation='linear')(z)
model = Model(inputs=rgb_efficientNetB0.input, outputs=z)
optimizer = optimizers.Adam(lr=0.001, decay=0)
# optimizer = optimizers.Adadelta(lr=1.0, rho=0.95, epsilon=None, decay=0.0)
model.compile(optimizer=optimizer,
loss="MSE",
metrics=["MSE", "accuracy"])
print("BasicCNN model built as child model.\n Model summary:")
print(model.summary())
return model
def get_efficientnet_model(
model_name='efficientnetb0',
input_shape=(224, 224, 3),
input_tensor=None,
include_top=True,
classes=1000,
weights='imagenet',
):
layer_names = [
'block3a_expand_activation', #C2
'block4a_expand_activation', #C3
'block6a_expand_activation', #C4
'top_activation' #C5
]
Args = {
'input_shape': input_shape,
'weights': weights,
'include_top': include_top,
'input_tensor': input_tensor
}
if model_name == 'efficientnetb0':
backbone = efn.EfficientNetB0(**Args)
elif model_name == 'efficientnetb1':
backbone = efn.EfficientNetB1(**Args)
elif model_name == 'efficientnetb2':
backbone = efn.EfficientNetB2(**Args)
elif model_name == 'efficientnetb3':
backbone = efn.EfficientNetB3(**Args)
elif model_name == 'efficientnetb4':
backbone = efn.EfficientNetB4(**Args)
elif model_name == 'efficientnetb5':
backbone = efn.EfficientNetB5(**Args)
elif model_name == 'efficientnetb6':
backbone = efn.EfficientNetB6(**Args)
elif model_name == 'efficientnetb7':
backbone = efn.EfficientNetB7(**Args)
else:
raise ValueError('No such model {}'.format(model_name))
several_layers = []
several_layers.append(backbone.get_layer(layer_names[0]).output)
several_layers.append(backbone.get_layer(layer_names[1]).output)
several_layers.append(backbone.get_layer(layer_names[2]).output)
several_layers.append(backbone.get_layer(layer_names[3]).output)
model = keras.models.Model(inputs=[backbone.input], outputs=several_layers)
return model
def effnet_retinanet(num_classes, backbone='EfficientNetB0', inputs=None, modifier=None, **kwargs):
""" Constructs a retinanet model using a resnet backbone.
Args
num_classes: Number of classes to predict.
backbone: Which backbone to use (one of ('resnet50', 'resnet101', 'resnet152')).
inputs: The inputs to the network (defaults to a Tensor of shape (None, None, 3)).
modifier: A function handler which can modify the backbone before using it in retinanet (this can be used to freeze backbone layers for example).
Returns
RetinaNet model with a ResNet backbone.
"""
# choose default input
if inputs is None:
if keras.backend.image_data_format() == 'channels_first':
inputs = keras.layers.Input(shape=(3, None, None))
else:
# inputs = keras.layers.Input(shape=(224, 224, 3))
inputs = keras.layers.Input(shape=(None, None, 3))
# get last conv layer from the end of each block [28x28, 14x14, 7x7]
if backbone == 'EfficientNetB0':
model = efn.EfficientNetB0(input_tensor=inputs, include_top=False, weights=None)
elif backbone == 'EfficientNetB1':
model = efn.EfficientNetB1(input_tensor=inputs, include_top=False, weights=None)
elif backbone == 'EfficientNetB2':
model = efn.EfficientNetB2(input_tensor=inputs, include_top=False, weights=None)
elif backbone == 'EfficientNetB3':
model = efn.EfficientNetB3(input_tensor=inputs, include_top=False, weights=None)
elif backbone == 'EfficientNetB4':
model = efn.EfficientNetB4(input_tensor=inputs, include_top=False, weights=None)
elif backbone == 'EfficientNetB5':
model = efn.EfficientNetB5(input_tensor=inputs, include_top=False, weights=None)
elif backbone == 'EfficientNetB6':
model = efn.EfficientNetB6(input_tensor=inputs, include_top=False, weights=None)
elif backbone == 'EfficientNetB7':
model = efn.EfficientNetB7(input_tensor=inputs, include_top=False, weights=None)
else:
raise ValueError('Backbone (\'{}\') is invalid.'.format(backbone))
layer_outputs = ['block4a_expand_activation', 'block6a_expand_activation', 'top_activation']
layer_outputs = [
model.get_layer(name=layer_outputs[0]).output, # 28x28
model.get_layer(name=layer_outputs[1]).output, # 14x14
model.get_layer(name=layer_outputs[2]).output, # 7x7
]
# create the densenet backbone
model = keras.models.Model(inputs=inputs, outputs=layer_outputs, name=model.name)
# invoke modifier if given
if modifier:
model = modifier(model)
# create the full model
return retinanet.retinanet(inputs=inputs, num_classes=num_classes, backbone_layers=model.outputs, **kwargs)
def __init__(self):
self.earlystop_callback = EarlyStopping(monitor='accuracy',
min_delta=0.001,
patience=2)
self.base_model = efn.EfficientNetB0(input_shape=(224, 224, 3),
include_top=False,
weights='imagenet')
for layer in self.base_model.layers:
layer.trainable = False
def create_effnet_model(self):
base_model = efn.EfficientNetB0(weights='imagenet',
include_top=False,
pooling='avg',
input_shape=(self.input_dims[0],
self.input_dims[0], 3))
x = base_model.output
x = keras.layers.Dropout(0.05)(x)
out = keras.layers.Dense(6, activation="sigmoid",
name='dense_output')(x)
return keras.models.Model(inputs=base_model.input, outputs=out)
def construct_mlp(input_size, num_classes, num_frames,
dropout_size=0.5, ef_mode=4, l2_reg=1e-5):
"""
Construct a MLP model for urban sound tagging.
Parameters
----------
num_frames
input_size
num_classes
dropout_size
ef_mode
l2_reg
Returns
-------
model
"""
# Add hidden layers
from keras.layers import Flatten, Conv1D, Conv2D, GlobalMaxPooling1D, GlobalAveragePooling1D, LSTM, Concatenate, GlobalAveragePooling2D, LeakyReLU
import efficientnet.keras as efn
if ef_mode == 0:
base_model = efn.EfficientNetB0(weights='noisy-student', include_top=False, pooling='avg')
elif ef_mode == 1:
base_model = efn.EfficientNetB1(weights='noisy-student', include_top=False, pooling='avg')
elif ef_mode == 2:
base_model = efn.EfficientNetB2(weights='noisy-student', include_top=False, pooling='avg')
elif ef_mode == 3:
base_model = efn.EfficientNetB3(weights='noisy-student', include_top=False, pooling='avg')
elif ef_mode == 4:
base_model = efn.EfficientNetB4(weights='noisy-student', include_top=False, pooling='avg') #imagenet or weights='noisy-student'
elif ef_mode == 5:
base_model = efn.EfficientNetB5(weights='noisy-student', include_top=False, pooling='avg')
elif ef_mode == 6:
base_model = efn.EfficientNetB6(weights='noisy-student', include_top=False, pooling='avg')
elif ef_mode == 7:
base_model = efn.EfficientNetB7(weights='noisy-student', include_top=False, pooling='avg')
input1 = Input(shape=input_size, dtype='float32', name='input')
input2 = Input(shape=(num_frames,85), dtype='float32', name='input2') #1621
y = TimeDistributed(base_model)(input1)
y = TimeDistributed(Dropout(dropout_size))(y)
y = Concatenate()([y, input2])
y = TimeDistributed(Dense(num_classes, activation='sigmoid', kernel_regularizer=regularizers.l2(l2_reg)))(y)
y = AutoPool1D(axis=1, name='output')(y)
m = Model(inputs=[input1, input2], outputs=y)
m.summary()
m.name = 'urban_sound_classifier'
return m
def test_custom(self):
from efficientnet import keras as efn
keras.backend.set_learning_phase(0)
base_model = efn.EfficientNetB0(input_shape=(600, 600, 3),
weights=None)
backbone = keras.Model(base_model.input,
base_model.get_layer("top_activation").output)
res = run_image(backbone,
self.model_files,
img_path,
target_size=(600, 600),
rtol=1e-1)
self.assertTrue(*res)
def get_model():
"Returns a classifier model"
efficient_net = efn.EfficientNetB0(
include_top=False,
weights="imagenet",
input_shape=(input_image_size, input_image_size, 3),
)
x = efficient_net.output
x = GlobalAveragePooling2D()(x)
predictions = Dense(6, activation='sigmoid')(x)
model = Model(inputs=efficient_net.input, outputs=predictions)
return model
def __init__(self, num_classes, **kwargs):
super(Model, self).__init__(**kwargs)
if FLAGS.arch == 'resnet':
self.resnet_model = resnet.resnet(
resnet_depth=FLAGS.resnet_depth,
width_multiplier=FLAGS.width_multiplier,
cifar_stem=FLAGS.image_size <= 32)
elif FLAGS.arch == 'efficientnet':
import efficientnet.keras as efn
self.resnet_model = efn.EfficientNetB0(weights=None, include_top=False)
self._projection_head = ProjectionHead()
if FLAGS.train_mode == 'finetune' or FLAGS.lineareval_while_pretraining:
self.supervised_head = SupervisedHead(num_classes)
def get_b0():
backbone = efn.EfficientNetB0(input_shape=(96, 96, 3), include_top=False, pooling=None, classes=None, weights='imagenet')
for layer in backbone.layers:
layer.trainable = True
avg = GlobalAveragePooling2D()(backbone.output)
grapheme_root_head = Dense(168, activation='softmax', name='grapheme_root')(avg)
vowel_diacritic_head = Dense(11, activation='softmax', name='vowel_diacritic')(avg)
consonant_diacritic_head = Dense(7, activation='softmax', name='consonant_diacritic')(avg)
return Model(backbone.input, outputs=[grapheme_root_head, vowel_diacritic_head, consonant_diacritic_head])
def build_efficientnet(height,
width,
depth,
include_top=False,
weights='imagenet'):
efficientnet_model = efn.EfficientNetB0(weights=weights,
include_top=include_top,
input_shape=(height, width, depth))
model = Sequential()
model.add(efficientnet_model)
model.add(GlobalAveragePooling2D())
model.add(Dropout(0.5))
model.add(Dense(2))
model.add(Activation("softmax"))
return model
def build(height, width, depth, num_classes):
input_shape = (height, width, depth)
chanDim = -1
model = efn.EfficientNetB0(weights='imagenet',
include_top=False,
input_shape=input_shape)
x = Flatten()(model.output)
x = Dense(500)(x)
x = Activation("relu")(x)
#x = Dense(100)(x)
#x = Activation("relu")(x)
x = Dense(num_classes)(x)
output = Activation("softmax")(x)
model = Model(model.input, output, name='efficientnetb0')
return model
def cnn_model(model_name, img_size, weights):
"""
Model definition using Xception net architecture
"""
input_size = (img_size, img_size, 3)
if model_name == "efn0":
baseModel = efn.EfficientNetB0(weights="imagenet", include_top=False,
input_shape=input_size)
elif model_name == "efn_noisy":
baseModel = efn.EfficientNetB5(weights="noisy-student", include_top=False,
input_shape=input_size)
headModel = baseModel.output
headModel = GlobalAveragePooling2D()(headModel)
headModel = Dense(1024, activation="relu", kernel_initializer="he_uniform", name="fc1")(
headModel
)
headModel = Dropout(0.4)(headModel)
predictions = Dense(
200,
activation="softmax",
kernel_initializer="he_uniform")(
headModel
)
model = Model(inputs=baseModel.input, outputs=predictions)
model.load_weights(weights)
model_fc = Model(
inputs=baseModel.input,
outputs=model.get_layer("fc1").output
)
for layer in baseModel.layers:
layer.trainable = False
optimizer = Nadam(
lr=0.002, beta_1=0.9, beta_2=0.999, epsilon=1e-08, schedule_decay=0.004
)
model.compile(
loss="categorical_crossentropy",
# loss=joint_loss,
optimizer=optimizer,
metrics=["accuracy"]
)
return model_fc
def build_model(input_shape):
input_tensor = Input(shape=input_shape)
base = efn.EfficientNetB0(weights='imagenet', include_top=False)
for l in base.layers:
l.trainable = True
conv = base.output
x = GlobalAveragePooling2D(name='pool1')(conv)
x = BatchNormalization()(x)
x = Dense(512, name='fc1')(x)
x = Activation('relu', name='relu1')(x)
x = BatchNormalization()(x)
#x = Dropout(0.5)(x)
x = Dense(4, name='fc3')(x)
x = Activation('sigmoid', name='sigmoid')(x)
model = Model(inputs=[base.input], outputs=[x])
model.compile(optimizer=RectifiedAdam(lr=LR), loss='binary_crossentropy')
return model
def get_model_effnet(img_shape, img_input, weights, effnet_version):
if effnet_version == 'B0':
effnet = efn.EfficientNetB0(include_top=False, input_tensor=img_input, weights=weights, pooling=None, input_shape=img_shape)
elif effnet_version == 'B1':
effnet = efn.EfficientNetB1(include_top=False, input_tensor=img_input, weights=weights, pooling=None, input_shape=img_shape)
elif effnet_version == 'B2':
effnet = efn.EfficientNetB2(include_top=False, input_tensor=img_input, weights=weights, pooling=None, input_shape=img_shape)
elif effnet_version == 'B3':
effnet = efn.EfficientNetB3(include_top=False, input_tensor=img_input, weights=weights, pooling=None, input_shape=img_shape)
elif effnet_version == 'B4':
effnet = efn.EfficientNetB4(include_top=False, input_tensor=img_input, weights=weights, pooling=None, input_shape=img_shape)
elif effnet_version == 'B5':
effnet = efn.EfficientNetB5(include_top=False, input_tensor=img_input, weights=weights, pooling=None, input_shape=img_shape)
elif effnet_version == 'B6':
effnet = efn.EfficientNetB6(include_top=False, input_tensor=img_input, weights=weights, pooling=None, input_shape=img_shape)
else:
effnet = efn.EfficientNetB7(include_top=False, input_tensor=img_input, weights=weights, pooling=None, input_shape=img_shape)
return effnet
def create_base_model(base_model_name, pretrained=True, IMAGE_SIZE=[300, 300]):
if pretrained is False:
weights = None
else:
weights = "imagenet"
if base_model_name == 'B0':
base = efn.EfficientNetB0(weights=weights,
include_top=False,
input_shape=[*IMAGE_SIZE, 3])
elif base_model_name == 'B1':
base = efn.EfficientNetB1(weights=weights,
include_top=False,
input_shape=[*IMAGE_SIZE, 3])
elif base_model_name == 'B2':
base = efn.EfficientNetB2(weights=weights,
include_top=False,
input_shape=[*IMAGE_SIZE, 3])
elif base_model_name == 'B3':
base = efn.EfficientNetB3(weights=weights,
include_top=False,
input_shape=[*IMAGE_SIZE, 3])
elif base_model_name == 'B4':
base = efn.EfficientNetB4(weights=weights,
include_top=False,
input_shape=[*IMAGE_SIZE, 3])
elif base_model_name == 'B5':
base = efn.EfficientNetB5(weights=weights,
include_top=False,
input_shape=[*IMAGE_SIZE, 3])
elif base_model_name == 'B6':
base = efn.EfficientNetB6(weights=weights,
include_top=False,
input_shape=[*IMAGE_SIZE, 3])
elif base_model_name == 'B7':
base = efn.EfficientNetB7(weights=weights,
include_top=False,
input_shape=[*IMAGE_SIZE, 3])
base = remove_dropout(base)
base.trainable = True
return base
def make_model():
"""
Creates an EfficientNet model with parameters specified `num_classes` outputs
and `img_width` x `img_height` input shape.
Args:
Returns: an EfficientNet model with specified global parameters.
"""
# create the base pre-trained model
base_model = efn.EfficientNetB0(input_shape=(img_width, img_height, 3),
include_top=False)
# add a global spatial average pooling layer
x = base_model.output
x = GlobalAveragePooling2D()(x)
predictions = Dense(num_classes, activation="softmax")(x)
model = Model(inputs=base_model.input, outputs=predictions)
model.compile(optimizer="adam",
loss="categorical_crossentropy",
metrics=["accuracy"])
return base_model, model
"models": keras.models,
"utils": keras.utils
}
if args.network == "resnet101":
zero = ResNet101(include_top=False,
weights=None if args.random_init else "imagenet",
input_tensor=input_tensor,
**model_kwargs)
model = ResNet101(include_top=False,
weights=None if args.random_init else "imagenet",
input_tensor=input_tensor,
**model_kwargs)
elif args.network == "enb0":
zero = en.EfficientNetB0(include_top=False,
weights=None if args.random_init else "imagenet",
input_tensor=input_tensor)
model = en.EfficientNetB0(include_top=False,
weights=None if args.random_init else "imagenet",
input_tensor=input_tensor)
else:
raise Error("Unknown network: " + args.network)
num_frozen = int(float(args.freeze_frac) * len(model.layers))
for l in model.layers[:num_frozen]:
if type(l) != keras.layers.normalization.BatchNormalization:
l.trainable = False
zero_features = zero.outputs[0]
model_features = model.outputs[0]
class SwishActivation(Activation):
def __init__(self, activation, **kwargs):
super(SwishActivation, self).__init__(activation, **kwargs)
self.__name__ = 'swish_act'
def swish_act(x, beta=1):
return (x * sigmoid(beta * x))
get_custom_objects().update({'swish_act': SwishActivation(swish_act)})
#______________________CLASSIFIER, Distance Models_____________________________
model = enet.EfficientNetB0(include_top=False,
input_shape=(70, 70, 3),
pooling='avg',
weights='efficientnet_b0.h5')
x = model.output
x = BatchNormalization()(x)
x = Dropout(0.4)(x)
x = Dense(512)(x)
x = BatchNormalization()(x)
x = Activation('swish_act')(x)
x = Dropout(0.5)(x)
x = Dense(256)(x)
x = Activation('swish_act')(x)
x = Dense(128)(x)
x = Activation('swish_act')(x)
x = Dense(64)(x)
x = Activation('swish_act')(x)
x = Dense(32)(x)
print('loading model...')
import efficientnet.keras as efn
from keras.applications.imagenet_utils import decode_predictions
from efficientnet.keras import center_crop_and_resize, preprocess_input
import numpy as np
import keras.backend.tensorflow_backend as tb
model = efn.EfficientNetB0(weights='noisy-student')
image_size = model.input_shape[1]
print('model loading complete!')
def predict(image):
x = center_crop_and_resize(image, image_size=image_size)
image_modified = x
x = preprocess_input(x)
x = np.expand_dims(x, 0)
# make prediction and decode
tb._SYMBOLIC_SCOPE.value = True
y = model.predict(x)
res = decode_predictions(y)
# collect result and the modified image
data = {'res': res[0][0], 'new_image': image_modified}
return data
def do_train():
base_dir = '/gpfs/gpfs0/deep/data/Otoliths_cod/codotoliths/'
dirs = set()
df_cod = pd.DataFrame(columns=['age', 'path'])
max_dataset_size = 1029 #6330
new_shape = (380, 380, 3)
IMG_SHAPE = (380, 380)
rb_imgs = np.empty(shape=(max_dataset_size, ) + new_shape)
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
tensorboard_path = './tensorboard_cod_age_softmax2'
checkpoint_path = './checkpoints_cod_age_softmax2/cod_oto_efficientnetB4.{epoch:03d}-{val_loss:.2f}.hdf5'
add_count = 0
for filename in Path(base_dir).glob('**/*.JPG'):
filepath = str(filename)
dirname = os.path.dirname(filepath)
if (dirname not in dirs):
dirs.add(dirname)
begin_age = filepath.find('age')
age = filepath[begin_age + 3:begin_age + 5]
age = int(age)
pil_img = load_img(filepath,
target_size=IMG_SHAPE,
grayscale=False)
array_img = img_to_array(pil_img, data_format='channels_last')
rb_imgs[add_count] = array_img
df_cod = df_cod.append({
'age': age,
'path': filepath
},
ignore_index=True)
add_count += 1
#print(add_count)
a_batch_size = 12
age = df_cod.age.values
early_stopper = EarlyStopping(patience=20)
train_datagen = ImageDataGenerator(
zca_whitening=True,
width_shift_range=0.,
height_shift_range=0., #20,
zoom_range=0.,
rotation_range=360,
horizontal_flip=False,
vertical_flip=True,
rescale=1. / 255)
#EXPERIMENT_FOLDER_PATH = os.path.join(parent_dir_of_file, f"reports/experiments/{EXPERIMENT_NAME}")
DEFAULT_CONFIG = {
"model": "basiccnn",
"method": "bayesian_optimization",
"train_set_size": 1029,
"opt_samples": 3,
"opt_last_n_epochs": 3,
"opt_initial_points": 10,
"child_epochs": 50,
"child_first_train_epochs": 0,
"child_batch_size": 64,
"pre_aug_weights_path": "pre_aug_weights.h5"
}
deepaug = DeepAugment(rb_imgs, age, config=DEFAULT_CONFIG)
train_idx, val_idx, test_idx = train_validate_test_split(
range(0, len(rb_imgs)))
train_rb_imgs = np.empty(shape=(len(train_idx), ) + new_shape)
train_age = []
for i in range(0, len(train_idx)):
train_rb_imgs[i] = rb_imgs[train_idx[i]]
train_age.append(age[train_idx[i]])
val_rb_imgs = np.empty(shape=(len(val_idx), ) + new_shape)
val_age = []
for i in range(0, len(val_idx)):
val_rb_imgs[i] = rb_imgs[val_idx[i]]
val_age.append(age[val_idx[i]])
test_rb_imgs = np.empty(shape=(len(test_idx), ) + new_shape)
test_age = []
for i in range(0, len(test_idx)):
test_rb_imgs[i] = rb_imgs[test_idx[i]]
test_age.append(age[test_idx[i]])
train_age = np.vstack(train_age)
val_age = np.vstack(val_age)
test_age = np.vstack(test_age)
val_rb_imgs = np.multiply(val_rb_imgs, 1. / 255)
test_rb_imgs = np.multiply(test_rb_imgs, 1. / 255)
train_generator = train_datagen.flow(train_rb_imgs,
train_age,
batch_size=a_batch_size)
efn.EfficientNetB0(weights='imagenet')
rgb_efficientNetB4 = EfficientNetB4(include_top=False,
weights='imagenet',
input_shape=new_shape,
classes=2)
z = dense1_linear_output(rgb_efficientNetB4)
scales = Model(inputs=rgb_efficientNetB4.input, outputs=z)
learning_rate = 0.0001
adam = optimizers.Adam(lr=learning_rate)
for layer in scales.layers:
layer.trainable = True
scales.compile(loss='mse',
optimizer=adam,
metrics=['accuracy', 'mse', 'mape'])
tensorboard, checkpointer = get_checkpoint_tensorboard(
tensorboard_path, checkpoint_path)
classWeight = None
history_callback = scales.fit_generator(
train_generator,
steps_per_epoch=1600,
epochs=150,
callbacks=[early_stopper, tensorboard, checkpointer],
validation_data=(val_rb_imgs, val_age),
class_weight=classWeight)
test_metrics = scales.evaluate(x=test_rb_imgs, y=test_age)
print("test metric:" + str(scales.metrics_names))
print("test metrics:" + str(test_metrics))
print("precision, recall, f1")
y_pred_test = scales.predict(test_rb_imgs, verbose=1)
y_pred_test_bool = np.argmax(y_pred_test, axis=1)
y_true_bool = np.argmax(test_age, axis=1)
#np.argmax inverse of to_categorical
argmax_test = np.argmax(test_age, axis=1)
unique, counts = np.unique(argmax_test, return_counts=True)
print("test ocurrence of each class:" + str(dict(zip(unique, counts))))
print("cslassification_report")
print(classification_report(y_true_bool, y_pred_test_bool))
print("confusion matrix")
print(str(confusion_matrix(y_true_bool, y_pred_test_bool)))
def create_model_efficientnet(existing='', is_halffeatures=True):
if len(existing) == 0:
print('Loading base model (efficientNetB0)..')
# Encoder Layers
base_model = efn.EfficientNetB0(weights='imagenet',
include_top=False,
input_shape=(None, None, 3))
print('Base model loaded.')
# Starting point for decoder
base_model_output_shape = base_model.layers[
-1].output.shape #15, 20, 1280
# Layer freezing?
# mid_start = base_model.get_layer('activation_22')
# for i in range(base_model.layers.index(mid_start)):
# base_model.layers[i].trainable = False
for layer in base_model.layers:
layer.trainable = True
# Starting number of decoder filters
if is_halffeatures:
decode_filters = int(int(base_model_output_shape[-1]) / 2)
else:
decode_filters = int(base_model_output_shape[-1])
# Define upsampling layer
def upproject(tensor, filters, name, concat_with):
up_i = BilinearUpSampling2D((2, 2),
name=name + '_upsampling2d')(tensor)
up_i = Concatenate(name=name + '_concat')(
[up_i,
base_model.get_layer(concat_with).output]) # Skip connection
up_i = Conv2D(filters=filters,
kernel_size=3,
strides=1,
padding='same',
name=name + '_convA')(up_i)
up_i = LeakyReLU(alpha=0.2)(up_i)
up_i = Conv2D(filters=filters,
kernel_size=3,
strides=1,
padding='same',
name=name + '_convB')(up_i)
up_i = LeakyReLU(alpha=0.2)(up_i)
return up_i
# Decoder Layers
decoder = Conv2D(filters=decode_filters,
kernel_size=1,
padding='same',
input_shape=base_model_output_shape,
name='conv2')(base_model.output)
decoder = upproject(decoder,
int(decode_filters / 2),
'up1',
concat_with='block4a_dwconv') #30, 40, 256
decoder = upproject(decoder,
int(decode_filters / 4),
'up2',
concat_with='block3a_dwconv') # 60, 80, 128
decoder = upproject(decoder,
int(decode_filters / 8),
'up3',
concat_with='block2a_dwconv') #120, 160, 64
decoder = upproject(decoder,
int(decode_filters / 16),
'up4',
concat_with='stem_conv') #240, 320, 64
if False:
decoder = upproject(decoder,
int(decode_filters / 32),
'up5',
concat_with='input_1')
# Extract depths (final layer)
conv3 = Conv2D(filters=1,
kernel_size=3,
strides=1,
padding='same',
name='conv3')(decoder)
# Create the model
model = Model(inputs=base_model.input, outputs=conv3)
else:
# Load model from file
if not existing.endswith('.h5'):
sys.exit(
'Please provide a correct model file when using [existing] argument.'
)
custom_objects = {
'BilinearUpSampling2D': BilinearUpSampling2D,
'depth_loss_function': depth_loss_function
}
model = load_model(existing, custom_objects=custom_objects)
print('\nExisting model loaded.\n')
print('Model created.')
return model
msg = "Accuracy on test set: {0:.1%} ({1} / {2})"
print(msg.format(acc, correct_sum, num_test))
if show_confusion_matrix:
print("Confusion Matrix:")
plot_test_confusion_matrix(cls_pred=cls_pred,
labels_test=labels_test,
classes=classes)
return float(acc)
#%%
model = efn.EfficientNetB0(include_top=False,
input_shape=(128, 128, 1),
weights=None)
x = model.output
x = GlobalAveragePooling2D()(x)
x = Dropout(0.7)(x)
predictions = Dense(3, activation="softmax")(x)
model_final = Model(inputs=model.input, outputs=predictions)
model_final.summary()
#%%
#Hyper Parameters
def UNetPlusPlus(img_rows,
img_cols,
color_type=3,
num_class=1,
deep_supervision=True):
nb_filter = [32, 64, 128, 256, 512]
import efficientnet.keras as efn
from keras.models import Model
from keras.layers.convolutional import Conv2D
from keras.layers import LeakyReLU, Add, Input, MaxPool2D, UpSampling2D, concatenate, Conv2DTranspose, \
BatchNormalization, Dropout
base_model = efn.EfficientNetB0(weights='imagenet',
include_top=False,
input_shape=(224, 224, 3))
input_model = base_model.input
# Handle Dimension Ordering for different backends
global bn_axis
bn_axis = 3
conv6_1 = base_model.get_layer('top_activation').output
conv5_1 = base_model.get_layer('block6a_expand_activation').output
conv4_1 = base_model.get_layer('block4a_expand_activation').output
conv3_1 = base_model.get_layer('block3a_expand_activation').output
conv2_1 = base_model.get_layer('block2a_expand_activation').output
print(conv6_1.shape) # 7*7
print(conv5_1.shape) # 14*14
print(conv4_1.shape) # 28*28
print(conv3_1.shape) # 56*56
print(conv2_1.shape) # 112*112
conv1_1 = standard_unit(input_model, stage='11', nb_filter=nb_filter[0])
# pool1 = MaxPooling2D((2, 2), strides=(2, 2), name='pool1')(conv1_1)
# conv2_1 = standard_unit(pool1, stage='21', nb_filter=nb_filter[1])
# pool2 = MaxPooling2D((2, 2), strides=(2, 2), name='pool2')(conv2_1)
up1_2 = Conv2DTranspose(nb_filter[0], (2, 2),
strides=(2, 2),
name='up12',
padding='same')(conv2_1)
conv1_2 = concatenate([up1_2, conv1_1], name='merge12', axis=bn_axis)
conv1_2 = standard_unit(conv1_2, stage='12', nb_filter=nb_filter[0])
# conv3_1 = standard_unit(pool2, stage='31', nb_filter=nb_filter[2])
# pool3 = MaxPooling2D((2, 2), strides=(2, 2), name='pool3')(conv3_1)
up2_2 = Conv2DTranspose(nb_filter[1], (2, 2),
strides=(2, 2),
name='up22',
padding='same')(conv3_1)
conv2_2 = concatenate([up2_2, conv2_1], name='merge22', axis=bn_axis)
conv2_2 = standard_unit(conv2_2, stage='22', nb_filter=nb_filter[1])
up1_3 = Conv2DTranspose(nb_filter[0], (2, 2),
strides=(2, 2),
name='up13',
padding='same')(conv2_2)
conv1_3 = concatenate([up1_3, conv1_1, conv1_2],
name='merge13',
axis=bn_axis)
conv1_3 = standard_unit(conv1_3, stage='13', nb_filter=nb_filter[0])
# conv4_1 = standard_unit(pool3, stage='41', nb_filter=nb_filter[3])
# pool4 = MaxPooling2D((2, 2), strides=(2, 2), name='pool4')(conv4_1)
up3_2 = Conv2DTranspose(nb_filter[2], (2, 2),
strides=(2, 2),
name='up32',
padding='same')(conv4_1)
conv3_2 = concatenate([up3_2, conv3_1], name='merge32', axis=bn_axis)
conv3_2 = standard_unit(conv3_2, stage='32', nb_filter=nb_filter[2])
up2_3 = Conv2DTranspose(nb_filter[1], (2, 2),
strides=(2, 2),
name='up23',
padding='same')(conv3_2)
conv2_3 = concatenate([up2_3, conv2_1, conv2_2],
name='merge23',
axis=bn_axis)
conv2_3 = standard_unit(conv2_3, stage='23', nb_filter=nb_filter[1])
up1_4 = Conv2DTranspose(nb_filter[0], (2, 2),
strides=(2, 2),
name='up14',
padding='same')(conv2_3)
conv1_4 = concatenate([up1_4, conv1_1, conv1_2, conv1_3],
name='merge14',
axis=bn_axis)
conv1_4 = standard_unit(conv1_4, stage='14', nb_filter=nb_filter[0])
# conv5_1 = standard_unit(pool4, stage='51', nb_filter=nb_filter[4])
conv5_1_upsampling_16 = Conv2DTranspose(nb_filter[0], (2, 2),
strides=(16, 16),
name='conv5_1_upsampling_16',
padding='same')(
conv5_1) # changes made
up4_2 = Conv2DTranspose(nb_filter[3], (2, 2),
strides=(2, 2),
name='up42',
padding='same')(conv5_1)
conv4_2 = concatenate([up4_2, conv4_1], name='merge42', axis=bn_axis)
conv4_2 = standard_unit(conv4_2, stage='42', nb_filter=nb_filter[3])
conv4_2_upsampling_8 = Conv2DTranspose(nb_filter[0], (2, 2),
strides=(8, 8),
name='conv4_2_upsampling_8',
padding='same')(
conv4_2) # changes made
up3_3 = Conv2DTranspose(nb_filter[2], (2, 2),
strides=(2, 2),
name='up33',
padding='same')(conv4_2)
conv3_3 = concatenate([up3_3, conv3_1, conv3_2],
name='merge33',
axis=bn_axis)
conv3_3 = standard_unit(conv3_3, stage='33', nb_filter=nb_filter[2])
conv3_3_upsampling_4 = Conv2DTranspose(nb_filter[0], (2, 2),
strides=(4, 4),
name='conv3_3_upsampling_4',
padding='same')(
conv3_3) # changes made
up2_4 = Conv2DTranspose(nb_filter[1], (2, 2),
strides=(2, 2),
name='up24',
padding='same')(conv3_3)
conv2_4 = concatenate([up2_4, conv2_1, conv2_2, conv2_3],
name='merge24',
axis=bn_axis)
conv2_4 = standard_unit(conv2_4, stage='24', nb_filter=nb_filter[1])
conv2_4_upsampling_2 = Conv2DTranspose(nb_filter[0], (2, 2),
strides=(2, 2),
name='conv2_4_upsampling_2',
padding='same')(
conv2_4) # changes made
up1_5 = Conv2DTranspose(nb_filter[0], (2, 2),
strides=(2, 2),
name='up15',
padding='same')(conv2_4)
conv1_5 = concatenate([up1_5, conv1_1, conv1_2, conv1_3, conv1_4],
name='merge15',
axis=bn_axis)
conv1_5 = standard_unit(conv1_5, stage='15', nb_filter=nb_filter[0])
nestnet_output_1 = Conv2D(num_class, (1, 1),
activation='sigmoid',
name='output_1',
kernel_initializer='he_normal',
padding='same',
kernel_regularizer=l2(1e-4))(conv1_2)
nestnet_output_2 = Conv2D(num_class, (1, 1),
activation='sigmoid',
name='output_2',
kernel_initializer='he_normal',
padding='same',
kernel_regularizer=l2(1e-4))(conv1_3)
nestnet_output_3 = Conv2D(num_class, (1, 1),
activation='sigmoid',
name='output_3',
kernel_initializer='he_normal',
padding='same',
kernel_regularizer=l2(1e-4))(conv1_4)
nestnet_output_4 = Conv2D(num_class, (1, 1),
activation='sigmoid',
name='output_4',
kernel_initializer='he_normal',
padding='same',
kernel_regularizer=l2(1e-4))(conv1_5)
# changes code
nested_conv2_4_upsampling_2 = Conv2D(
num_class, (1, 1),
activation='sigmoid',
name='output_5',
kernel_initializer='he_normal',
padding='same',
kernel_regularizer=l2(1e-4))(conv2_4_upsampling_2)
nested_conv3_3_upsampling_4 = Conv2D(
num_class, (1, 1),
activation='sigmoid',
name='output_6',
kernel_initializer='he_normal',
padding='same',
kernel_regularizer=l2(1e-4))(conv3_3_upsampling_4)
nested_conv4_2_upsampling_8 = Conv2D(
num_class, (1, 1),
activation='sigmoid',
name='output_7',
kernel_initializer='he_normal',
padding='same',
kernel_regularizer=l2(1e-4))(conv4_2_upsampling_8)
nested_conv5_1_upsampling_16 = Conv2D(
num_class, (1, 1),
activation='sigmoid',
name='output_8',
kernel_initializer='he_normal',
padding='same',
kernel_regularizer=l2(1e-4))(conv5_1_upsampling_16)
nestnet_output_all = keras.layers.Average()([
nestnet_output_1, nestnet_output_2, nestnet_output_3, nestnet_output_4,
nested_conv2_4_upsampling_2, nested_conv3_3_upsampling_4,
nested_conv4_2_upsampling_8, nested_conv5_1_upsampling_16
])
model = Model(inputs=[input_model], outputs=[nestnet_output_all])
# if deep_supervision:
# model = Model(input=img_input, output=[nestnet_output_1,
# nestnet_output_2,
# nestnet_output_3,
# nestnet_output_4,
# nested_conv2_4_upsampling_2,
# nested_conv3_3_upsampling_4,
# nested_conv4_2_upsampling_8,
# nested_conv5_1_upsampling_16
# ])
# else:
# model = Model(input=img_input, output=[nestnet_output_4])
return model
def model(input_form="all", aux_size=0, hyperparameters=dict()):
print("using the following hyperparameters: {}".format(hyperparameters))
if input_form == "features":
return features_model(aux_size, hyperparameters)
parameters = INPUT_FORM_PARAMETERS[input_form]
inputs = list()
outputs = list()
#retreiving the hyperparameters
DROPOUT = hyperparameters.get("dropout", 0.5)
OPTIMIZER = hyperparameters.get("optimizer", "sgd-0001-0.9")
DEEP_DENSE_TOP = hyperparameters.get("deep-dense-top", True)
CONVNET_FREEZE_PERCENT = hyperparameters.get("convnet-freeze-percent", 0.0)
#skip for now
'''
if parameters["t2"]:
convnet = applications.ResNet50(
weights="imagenet",
include_top=False,
input_shape=(config.IMAGE_SIZE, config.IMAGE_SIZE, 3),
)
for layer in convnet.layers:
layer.name = "{}_t2".format(layer.name)
apply_layer_freeze(convnet, CONVNET_FREEZE_PERCENT)
out = convnet.output
out = Flatten()(out)
inputs.append(convnet.input)
outputs.append(out)
'''
if parameters["t1"]:
# init ResNet
convnet = efn.EfficientNetB0(
weights="imagenet",
include_top=False,
input_shape=(config.IMAGE_SIZE, config.IMAGE_SIZE, 3),
)
apply_layer_freeze(convnet, CONVNET_FREEZE_PERCENT)
out = convnet.output
out = Flatten()(out)
inputs.append(convnet.input)
outputs.append(out)
if len(outputs) > 1:
out = concatenate(outputs)
else:
out = outputs[0]
out = Dense(256, activation="relu", kernel_regularizer=l1_l2(l1=0.01, l2=0.01))(out)
out = BatchNormalization()(out)
if DEEP_DENSE_TOP:
out = Dropout(DROPOUT)(out)
out = Dense(128, activation="relu", kernel_regularizer=l1_l2(l1=0.01, l2=0.01))(out)
out = BatchNormalization()(out)
out = Dropout(DROPOUT)(out)
out = Dense(64, activation="relu", kernel_regularizer=l1_l2(l1=0.01, l2=0.01))(out)
out = BatchNormalization()(out)
out = Dropout(DROPOUT)(out)
out = Dense(32, activation="relu", kernel_regularizer=l1_l2(l1=0.01, l2=0.01))(out)
out = BatchNormalization()(out)
out = Dropout(DROPOUT)(out)
if parameters["features"]:
aux_input = Input(shape=(aux_size,), name='aux_input')
inputs.append(aux_input)
out = concatenate([out, aux_input])
out = Dense(16, activation="relu", kernel_regularizer=l1_l2(l1=0.01, l2=0.01))(out)
out = BatchNormalization()(out)
predictions = Dense(1, activation="sigmoid", kernel_regularizer=l1_l2(l1=0.01, l2=0.01))(out)
# creating the final model
if len(inputs) > 1:
model = Model(inputs=inputs, outputs=predictions)
else:
model = Model(inputs=inputs[0], outputs=predictions)
# compile the model
model.compile(
loss="binary_crossentropy",
optimizer=OPTIMIZERS[OPTIMIZER](),
metrics=["accuracy"])
return model