def create_model(conf):
"""
"""
if conf["model"] == 'EfficientNetB0':
from efficientnet.tfkeras import EfficientNetB0 as Network # 4.7M params
elif conf["model"] == 'EfficientNetB1':
from efficientnet.tfkeras import EfficientNetB1 as Network # 7.2M params
elif conf["model"] == 'EfficientNetB2':
from efficientnet.tfkeras import EfficientNetB2 as Network # 8.5M params
elif conf["model"] == 'EfficientNetB3':
from efficientnet.tfkeras import EfficientNetB3 as Network # 11.5M params
elif conf["model"] == 'EfficientNetB4':
from efficientnet.tfkeras import EfficientNetB4 as Network # 18.6M params
elif conf["model"] == 'EfficientNetB5':
from efficientnet.tfkeras import EfficientNetB5 as Network # 29.5M params
elif conf["model"] == 'EfficientNetB6':
from efficientnet.tfkeras import EfficientNetB6 as Network # 42.2M params
elif conf["model"] == 'EfficientNetB7':
from efficientnet.tfkeras import EfficientNetB7 as Network # 65.4M params
elif conf["model"] == 'ResNet50':
from tensorflow.keras.applications import ResNet50 as Network # 23,6M params
base_model = Network(
weights=conf["weights"], # "imagenet", None or "noisy-student"
include_top=False,
input_shape=conf["img_shape"])
# Unfreeze the layers. I.E we're just using the pre-trained
# weights as initial weigths and biases and train over them
base_model.trainable = True
# Define model
model = Sequential()
model.add(base_model)
model.add(layers.GlobalAveragePooling2D())
model.add(layers.Dropout(conf["dropout"]))
model.add(layers.Dense(512, activation='relu'))
model.add(layers.Dropout(conf["dropout"]))
model.add(
layers.Dense(conf["num_classes"], activation=conf["final_activation"]))
if conf['optimizer'] == 'Adam':
opt = tf.keras.optimizers.Adam(learning_rate=conf["learning_rate"])
elif conf['optimizer'] == 'SGD':
opt = tf.keras.optimizers.SGD(learning_rate=conf["learning_rate"])
model.compile(optimizer=opt,
loss='sparse_categorical_crossentropy',
metrics=['sparse_categorical_accuracy'])
if conf["verbosity"]:
model.summary()
return model
def build_efficientnet(self):
"""Builds efficient model with dense layers on top
Returns:
keras.models.Model
:return:
"""
from efficientnet.tfkeras import EfficientNetB0
strategy = tf.distribute.MirroredStrategy()
with strategy.scope():
conv_base = EfficientNetB0(weights='noisy-student',
include_top=False,
input_shape=(224, 224, 3))
outputconv_base = conv_base.output
t_flat = Flatten()(outputconv_base)
t_dense1 = Dense(1024, activation='relu')(t_flat)
t_dense2 = Dense(256, activation='relu')(t_dense1)
t_dense3 = Dense(128, activation='relu')(t_dense2)
t_do = Dropout(0.5)(t_dense3)
predictions = Dense(2, activation='softmax')(t_do)
model = Model(inputs=conv_base.input,
outputs=predictions,
name='model')
conv_base.trainable = False
opt = tf.keras.optimizers.Adam(learning_rate=0.0002)
model.compile(
loss=tf.keras.losses.BinaryCrossentropy(label_smoothing=0.01),
optimizer=opt,
metrics=['accuracy'])
return model
def SR_model(num_classes,
dropout,
mc_dropout,
input_dim,
training,
pooling='avg'):
inputs = Input(input_dim)
base_model = EfficientNetB0(include_top=False,
weights='imagenet',
input_tensor=inputs)
base_model.trainable = True
x = base_model.output
x = Dropout(dropout, name='top_dropout_1')(x, training=training)
if pooling == 'avg':
x = GlobalAveragePooling2D(name='avg_pool')(x)
elif pooling == 'max':
x = GlobalMaxPooling2D(name='max_pool')(x)
x = Dropout(dropout, name='top_dropout_2')(x, training=training)
x = Dense(512, activation='relu', name='dense_512')(x)
x = BatchNormalization()(x)
x = Dropout(dropout, name='top_dropout_3')(x, training=training)
x = Lambda(lambda x: K.dropout(x, level=mc_dropout))(x)
#classification head (f)
sr = Dense(num_classes, activation='softmax', name='dense_f')(x)
return Model(inputs=inputs, outputs=sr)
def __init__(self,
dropout=0.2,
mc_dropout=0.2,
num_classes=1,
training=True,
input_dim=(224, 224, 3),
pooling="avg"):
self.c = 0.75
self.lamda = 32
self.alpha = 0.5
self.dropout = dropout
self.mc_dropout = mc_dropout
self.pooling = pooling
self.input_dim = input_dim
self.training = training
self.num_classes = num_classes
#create model
inputs = Input(shape=self.input_dim)
base_model = EfficientNetB0(include_top=False,
weights='imagenet',
input_tensor=inputs)
base_model.trainable = True
x = base_model.output
x = Dropout(self.dropout, name='top_dropout_1')(x,
training=self.training)
if pooling == 'avg':
x = GlobalAveragePooling2D(name='avg_pool')(x)
elif pooling == 'max':
x = GlobalMaxPooling2D(name='max_pool')(x)
x = Dropout(self.dropout, name='top_dropout_2')(x,
training=self.training)
x = Dense(512, activation='relu', name='dense_512')(x)
x = BatchNormalization()(x)
x = Dropout(self.mc_dropout,
name='top_dropout_3')(x, training=self.training)
x = Lambda(lambda x: K.dropout(x, level=self.mc_dropout))(x)
#classification head (f)
f = Dense(self.num_classes, activation='softmax', name='f_head')(x)
#selection head (g)
g = Dense(512, activation='relu', name='dense_512_g')(x)
g = BatchNormalization()(g)
# this normalization is identical to initialization of batchnorm gamma to 1/10
g = Lambda(lambda a: a / 10)(g)
g = Dense(1, activation='sigmoid', name='g_head')(g)
# auxiliary head (h)
selective_output = Concatenate(axis=1, name="selective_head")([f, g])
auxillary_output = Dense(self.num_classes,
activation='softmax',
name='auxilary_head')(x)
self.model = Model(inputs=inputs,
outputs=[selective_output, auxillary_output])
def efficientnet(b,
weights='imagenet',
include_top=False,
input_shape=(None, None, 3)):
"""Loads the appropriate EfficientNet model with weights
:param b: The size of the EfficientNet model.
:type b: int
:param weights: The pretrained weights to load. Defaults to iamgenet.
:type weights: str
:param include_top: Include the pretrained softmax layer. Defaults to False
:type include_top: bool
:param input_shape: Shape of input images. Defaults to no hight or width, 3 channels.
:type input_shape: Tuple
:return: EfficientNet Model.
:rtype: tf.keras.models.Model
"""
if b == 0:
return EfficientNetB0(weights=weights,
include_top=include_top,
input_shape=input_shape)
elif b == 1:
return EfficientNetB1(weights=weights,
include_top=include_top,
input_shape=input_shape)
elif b == 2:
return EfficientNetB2(weights=weights,
include_top=include_top,
input_shape=input_shape)
elif b == 3:
return EfficientNetB3(weights=weights,
include_top=include_top,
input_shape=input_shape)
elif b == 4:
return EfficientNetB4(weights=weights,
include_top=include_top,
input_shape=input_shape)
elif b == 5:
return EfficientNetB5(weights=weights,
include_top=include_top,
input_shape=input_shape)
elif b == 6:
return EfficientNetB6(weights=weights,
include_top=include_top,
input_shape=input_shape)
elif b == 7:
return EfficientNetB7(weights=weights,
include_top=include_top,
input_shape=input_shape)
else:
raise Exception("Invalid size for EfficientNet")
def load_effnet(shape, efftype = 'B0', weights = 'imagenet', trainable = False):
if efftype == 'B0':
backbone = EfficientNetB0(weights = weights, include_top = False, input_shape = shape)
if efftype == 'B3':
backbone = EfficientNetB3(weights = weights, include_top = False, input_shape = shape)
if efftype == 'B5':
backbone = EfficientNetB5(weights = weights, include_top = False, input_shape = shape)
for layer in backbone.layers:
layer.trainable = trainable
return backbone
def __init__(self):
num_cls = 5
imag_size = 224
#------------------------------------start building model-----------------------------#
model = EfficientNetB0(weights = 'imagenet', input_shape = (imag_size,imag_size,3), include_top = False)
ENet_out = model.output
ENet_out = Flatten()(ENet_out)
Hidden1_in = Dense(1024, activation="relu")(ENet_out)
Hidden1_in = Dropout(0.5)(Hidden1_in)
predictions = Dense(units = num_cls, activation="softmax")(Hidden1_in)
self.model_f = Model(inputs = model.input, outputs = predictions)
self.model_f.compile(optimizers.Adam(lr=0.0001, beta_1=0.9, beta_2=0.999, epsilon=1e-08), loss='categorical_crossentropy', metrics=[metrics.mae, metrics.categorical_accuracy])
self.model_f.load_weights("efficientnetB0_weights.h5")
def build_model(self):
self.base_model = EfficientNetB0(weights='noisy-student',
include_top=False)
self.base_model.trainable = True
self.x = self.base_model.output
self.x = GlobalAveragePooling2D()(self.x)
self.x = Dense(512, activation='relu')(self.x)
self.x = Dropout(rate=0.2)(self.x)
self.x = Dense(128, activation='relu')(self.x)
self.x = Dropout(rate=0.2)(self.x)
self.preds = Dense(1, activation='sigmoid')(self.x)
self.model = Model(inputs=self.base_model.input, outputs=self.preds)
self.model.compile(optimizer='adam',
loss='binary_crossentropy',
metrics=['accuracy'])
def get_base_model(modelname):
if modelname == 'vgg16':
print('MODELO CARGADO: VGG16')
return tf.keras.applications.VGG16(input_shape=(img_size(), img_size(),
3),
include_top=False,
weights='imagenet')
elif modelname == 'mobilenet':
print('MODELO CARGADO: MOBILENET')
return tf.keras.applications.MobileNetV2(input_shape=(img_size(),
img_size(), 3),
include_top=False,
weights='imagenet')
elif modelname == 'efficientnet':
print('MODELO CARGADO: EFFICIENTNET')
return EfficientNetB0(input_shape=(img_size(), img_size(), 3),
include_top=False,
weights='imagenet')
def Eff_net():
# loading pretrained conv base model
conv_base = EfficientNetB0(weights="imagenet",
include_top=true,
input_shape=(200, 200, 3))
dropout_rate = 0.05
model = models.Sequential()
model.add(conv_base)
model.add(layers.Flatten(name="flatten"))
# model.add(layers.Dense(32, activation="relu"))
# model.add(layers.Dropout(dropout_rate))
# model.add(layers.Dense(16, activation="relu"))
# model.add(layers.Dropout(dropout_rate))
# model.add(layers.Dense(256, activation='relu', name="fc1"))
model.add(layers.Dense(1, activation="sigmoid", name="fc_out"))
conv_base.trainable = False
return model
def __init__(self, arch, pretrained=None):
super(EfficientNet, self).__init__()
self.net = EfficientNetB0(include_top=False,
weights=pretrained,
input_shape=(224, 224, 3),
classes=10)
# elif arch == 'efficientnet-b2':
# self.net = EfficientNetB2(include_top=False,
# weights=pretrained,
# input_shape=(260, 260, 3),
# classes=10)
self.feature_3 = self.net.get_layer('block4a_expand_activation').output
self.feature_2 = self.net.get_layer('block6a_expand_activation').output
self.feature_1 = self.net.get_layer('top_activation').output
self.feature_extractor = tf.keras.Model(
inputs=self.net.input,
outputs=[self.feature_3, self.feature_2, self.feature_1])
def create_efficientNet():
if image_model == 4:
baseModel = EfficientNetB4(weights='imagenet',
include_top=False,
input_shape=input_size)
elif image_model == 2:
baseModel = EfficientNetB2(weights='imagenet',
include_top=False,
input_shape=input_size)
elif image_model == 0:
baseModel = EfficientNetB0(weights='imagenet',
include_top=False,
input_shape=input_size)
probs = baseModel.layers.pop()
top_conv = probs.input
headModel = layers.Activation(swish, name='top_activation')(top_conv)
headModel = layers.GlobalAveragePooling2D(name='avg_pool')(headModel)
headModel = layers.Dropout(0.2, name='top_dropout')(headModel)
headModel = layers.Dense(num_classes, activation='softmax')(headModel)
model = Model(inputs=baseModel.input, outputs=headModel)
return model
def __init__(self):
self.num_cls = 5
self.imag_size = 224
model = EfficientNetB0(weights='imagenet',
input_shape=(self.imag_size, self.imag_size, 3),
include_top=False)
outputs = model.output
outputs = Flatten()(outputs)
Hidden1_inputs = Dense(1024, activation="relu")(outputs)
Hidden1_inputs = Dropout(0.5)(Hidden1_inputs)
predictions = Dense(units=self.num_cls,
activation="softmax")(Hidden1_inputs)
self.model_f = Model(inputs=model.input, outputs=predictions)
self.model_f.compile(
optimizers.Adam(lr=0.0001, beta_1=0.9, beta_2=0.999,
epsilon=1e-08),
loss='categorical_crossentropy',
metrics=[metrics.mae, metrics.categorical_accuracy])
def Eff_net_B0_simplest(Net_file_path, Trainable='All', load=True):
# loading pretrained conv base model
conv_base = EfficientNetB0(weights="imagenet",
include_top=False,
input_shape=(200, 200, 3))
dropout_rate = 0.2
model = models.Sequential()
model.add(conv_base)
model.add(layers.Flatten(name="flatten"))
model.add(tf.keras.layers.BatchNormalization())
model.add(layers.Dense(32, activation="relu", name="Dense1_add"))
# model.add(tf.keras.layers.BatchNormalization())
model.add(layers.Dropout(dropout_rate))
model.add(layers.Dense(16, activation="relu", name="Dense2_add"))
# model.add(tf.keras.layers.BatchNormalization())
model.add(layers.Dropout(dropout_rate))
model.add(layers.Dense(1, activation="sigmoid", name="fc_out"))
if load:
model.load_weights(Net_file_path)
if Trainable == 'Classifier':
conv_base.trainable = False
elif Trainable == 'Class_FirstConv':
conv_base.trainable = True
set_trainable = False
for layer in conv_base.layers:
if layer.name == 'multiply_16':
set_trainable = True
if set_trainable:
layer.trainable = True
else:
layer.trainable = False
elif Trainable == 'Class_SecondConv':
conv_base.trainable = True
set_trainable = False
for layer in conv_base.layers:
if layer.name == 'multiply_14':
set_trainable = True
if set_trainable:
layer.trainable = True
else:
layer.trainable = False
elif Trainable == 'Class_ThirdConv':
conv_base.trainable = True
set_trainable = False
for layer in conv_base.layers:
if layer.name == 'multiply_12':
set_trainable = True
if set_trainable:
layer.trainable = True
else:
layer.trainable = False
return model
print(f'Train of Label : {x_test.shape}')
print(f'Test of Label : {y_test.shape}')
x_test = x_test[:5138]
y_test = y_test[:5138]
print(x_train.shape)
print(y_train.shape)
print(x_test.shape)
print(y_test.shape)
# model
BATCH_SIZE = 64
EPOCHS = 10
model = tf.compat.v1.keras.models.Sequential([
EfficientNetB0(include_top=False, input_shape=(224, 224, 3),
pooling='avg'),
tf.compat.v1.keras.layers.Dense(20),
tf.compat.v1.keras.layers.BatchNormalization(),
tf.compat.v1.keras.layers.Activation('softmax')
])
# fitting
model.compile(loss='sparse_categorical_crossentropy',
optimizer=tf.keras.optimizers.Adam(lr=0.01),
metrics=['accuracy'])
hist = model.fit(x_train,
y_train,
epochs=EPOCHS,
batch_size=BATCH_SIZE,
validation_split=0.2)
def run_inference(tfrecords_folder, output_folder, batch_size=1000):
Path(output_folder).mkdir(parents=True, exist_ok=True)
model = EfficientNetB0(weights="imagenet", include_top=False, pooling="avg")
tfrecords_to_write_embeddings(tfrecords_folder, output_folder, model, batch_size)
def run_inference_from_files(image_folder, output_folder, num_shards=10, batch_size=1000):
model = EfficientNetB0(weights="imagenet", include_top=False, pooling="avg")
list_ds = list_files(image_folder)
compute_save_embeddings(list_ds, output_folder, num_shards, model, batch_size)
#from efficientnet.tfkeras import center_crop_and_resize, preprocess_input
from tensorflow.keras import models
from tensorflow.keras import layers
from keras.datasets import cifar100
from tensorflow.keras.preprocessing.image import ImageDataGenerator
from tensorflow.keras import optimizers
from tensorflow.keras.utils import plot_model
from IPython.display import Image
from keras.utils import to_categorical
import matplotlib.pyplot as plt
from skimage.transform import resize
import numpy as np
import cv2
from keras import backend as K
baseModel = EfficientNetB0(weights="imagenet")
#print(baseModel.summary())
dropout_rate = 0.2
batch_size = 64
epochs = 25
baseModel = EfficientNetB0(weights="imagenet", include_top=False)
#print(baseModel.summary())
model = models.Sequential()
model.add(baseModel)
model.add(layers.GlobalMaxPooling2D(name="gmp"))
if dropout_rate > 0:
model.add(layers.Dropout(dropout_rate, name="dropout_out"))
f.write("Model : " + model_name)
f.close()
else:
pass
train_ds = make_tf_dataset(train_images, train_labels)
valid_ds = make_tf_dataset(valid_images, valid_labels)
train_ds = train_ds.repeat().batch(BATCH_SIZE)
train_ds = train_ds.prefetch(AUTOTUNE)
valid_ds = valid_ds.repeat().batch(BATCH_SIZE)
valid_ds = valid_ds.prefetch(AUTOTUNE)
base_model = EfficientNetB0(input_shape=(IMG_SIZE, IMG_SIZE, 3),
weights="imagenet",
include_top=False)
avg = tf.keras.layers.GlobalAveragePooling2D()(base_model.output)
output = tf.keras.layers.Dense(train_labels_len, activation="sigmoid")(avg)
model = tf.keras.Model(inputs=base_model.input, outputs=output)
for layer in base_model.layers:
layer.trainable = True
# optimizer = tf.keras.optimizers.SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
# model.compile(loss="categorical_crossentropy", optimizer=optimizer, metrics=["accuracy"])
model.compile(optimizer='adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
model.summary()
# train_generator = CutMixImageDataGenerator(
# generator1=train_generator1,
# generator2=train_generator2,
# img_size=(224,224),
# batch_size=64
# )
test_generator = test_datagen.flow_from_directory(
'/home/john/Study/pytorch_retinaface/facedetect/data/test',
target_size=(224, 224),
batch_size=64,
class_mode='binary')
# '/home/john/Study/mask/mask_model_0925.h5'
# 2. 모델
model = Sequential()
model.add(EfficientNetB0(include_top=False, pooling='avg'))
model.add(Dropout(0.1, name='hidden1'))
#model.add(GlobalAveragePooling2D(name='hidden2'))
model.add(Dense(1, activation='sigmoid', name='s1'))
model.summary()
# 3. 훈련
# tb_hist = TensorBoard(log_dir='./graph', histogram_freq=0, write_graph=True, write_images=True, update_freq='batch')
# parallel_model = multi_gpu_model(model, gpus=['/xla_gpu:1', '/xla_gpu:2'])
model.compile(optimizer='adam',
loss=tf.keras.losses.BinaryCrossentropy(label_smoothing=0.9),
metrics=['acc'])
hist = model.fit_generator(train_generator,
steps_per_epoch=100,
epochs=100,
validation_data=test_generator,
import os
import sys
import numpy as np
from skimage.io import imread
import matplotlib.pyplot as plt
sys.path.append('..')
# if you use tensorflow.keras:
from efficientnet.tfkeras import EfficientNetB0
from efficientnet.tfkeras import center_crop_and_resize, preprocess_input
from tensorflow.keras.applications.imagenet_utils import decode_predictions
# test image
test_b_2 = "../predict/pics/ROI.png"
image = imread(test_b_2)
plt.figure(figsize=(10, 10))
plt.imshow(image)
plt.show()
# loading pretrained model
model = EfficientNetB0(weights='imagenet')
# preprocess input
image_size = model.input_shape[1]
x = center_crop_and_resize(image, image_size=image_size)
x = preprocess_input(x)
print(x.shape)
x = np.expand_dims(x, 0)
# make prediction and decode
y = model.predict(x)
print(decode_predictions(y))
rotation_range=45,
width_shift_range=0.2,
height_shift_range=0.2)
# Validation data generator
datagen_val = ImageDataGenerator(rescale=1. / 255)
##################################
## MODEL #########################
##################################
input_shape = (IMG_ROWS, IMG_COLS, 3)
#### Efficient ####
effnet = EfficientNetB0(weights='imagenet',
include_top=False,
input_shape=input_shape)
effnet.trainable = True
x = effnet.output
x = Flatten()(x)
x = Dense(2048, activation='relu')(x)
x = Dropout(0.5)(x)
x = Dense(1024, activation='relu')(x)
x = Dropout(0.5)(x)
predictions = Dense(NUM_CLASSES, activation='sigmoid')(x)
model = Model(inputs=effnet.input, outputs=predictions)
###################
def build_model(base_model, input_shape, metrics, loss, loss_weights,
**kwargs):
if base_model == 'resnet50':
from tensorflow.keras.applications import ResNet50
base_model = ResNet50(include_top=False,
weights='imagenet',
input_shape=input_shape)
if base_model == 'densenet121':
from tensorflow.keras.applications import DenseNet121
base_model = DenseNet121(include_top=False,
weights='imagenet',
input_shape=input_shape)
if base_model == 'efficientnetb0':
from efficientnet.tfkeras import EfficientNetB0
base_model = EfficientNetB0(include_top=False,
weights='imagenet',
input_shape=input_shape)
if base_model == 'efficientnetb1':
from efficientnet.tfkeras import EfficientNetB1
base_model = EfficientNetB1(include_top=False,
weights='imagenet',
input_shape=input_shape)
if base_model == 'efficientnetb2':
from efficientnet.tfkeras import EfficientNetB2
base_model = EfficientNetB2(include_top=False,
weights='imagenet',
input_shape=input_shape)
if base_model == 'efficientnetb3':
from efficientnet.tfkeras import EfficientNetB3
base_model = EfficientNetB3(include_top=False,
weights='imagenet',
input_shape=input_shape)
if base_model == 'efficientnetb4':
from efficientnet.tfkeras import EfficientNetB4
base_model = EfficientNetB4(include_top=False,
weights='imagenet',
input_shape=input_shape)
if base_model == 'efficientnetb5':
from efficientnet.tfkeras import EfficientNetB5
base_model = EfficientNetB5(include_top=False,
weights='imagenet',
input_shape=input_shape)
x_in = Input(shape=input_shape)
x = Conv2D(3, (3, 3), padding='same')(x_in)
x = base_model(x)
x = GlobalAvgPool2D()(x)
x = BatchNormalization()(x)
x = Dropout(0.5)(x)
out_grapheme = Dense(168, activation='softmax', name='root')(x)
out_vowel = Dense(11, activation='softmax', name='vowel')(x)
out_consonant = Dense(7, activation='softmax', name='consonant')(x)
model = Model(inputs=x_in,
outputs=[out_grapheme, out_vowel, out_consonant])
model.compile(Adam(lr=0.0001),
metrics=metrics,
loss=loss,
loss_weights=loss_weights)
return model
from efficientnet.tfkeras import EfficientNetB0
import numpy as np
image = np.random.randint(0, 255, (1, 224, 224, 3)) / 255.
label = np.array([1])
model = EfficientNetB0()
model.compile(loss='sparse_categorical_crossentropy',
optimizer='adam',
metrics=['acc'])
model.fit(image, label, epochs=1)
pred = model.predict(image)
print(pred)
"dogs.csv")
dogs_photos_path = os.path.join(embedding_utils.DATASET_PATH, "dog_data_small",
"dogs_photos.csv")
dog_images_path = os.path.join(embedding_utils.DATASET_PATH,
"resized_dogs_small")
embeddings_path = os.path.join(embedding_utils.DATASET_PATH,
"embeddings_small")
# Variable to hold the dogs embeddings
embeddings = None
stacked_embeddings = None
embeddings_id_to_name = None
faiss_index = None
dogs = None
efficientnet_model = EfficientNetB0(weights='imagenet',
include_top=False,
pooling="avg")
def create_faiss_index(stacked_embeddings):
d = 1280
index = faiss.IndexFlatIP(d)
index.add(stacked_embeddings)
return index
def faiss_search(index, id_to_name, emb, k=5):
# Peform actual search
D, I = index.search(np.expand_dims(emb, 0), k)
return list(zip(D[0], [id_to_name[x] for x in I[0]]))
headModel = baseModel.output
headModel = Dense(51, activation='softmax')(headModel)
elif network == "NASNetMobile":
print(network)
train, test, lb2, labelsTest = preprocessing(network)
baseModel = NASNetMobile(weights=pretraining,
include_top=False,
input_tensor=Input(shape=(224, 224, 3)),
pooling="avg")
headModel = baseModel.output
headModel = Dense(51, activation='softmax')(headModel)
elif network == "EfficientNetB0":
print(network)
train, test, lb2, labelsTest = preprocessing_EfficcientNet()
baseModel = EfficientNetB0(weights=pretraining,
include_top=False,
input_tensor=Input(shape=(224, 224, 3)),
pooling="avg")
headModel = baseModel.output
headModel = Dropout(0.2)(headModel)
headModel = Dense(51, activation='softmax')(headModel)
else:
print(network)
train, test, lb2, labelsTest = preprocessing_EfficcientNet()
baseModel = EfficientNetB3(weights=pretraining,
include_top=False,
input_tensor=Input(shape=(224, 224, 3)),
pooling="avg")
headModel = baseModel.output
headModel = Dropout(0.2)(headModel)
headModel = Dense(51, activation='softmax')(headModel)
test_datagen = ImageDataGenerator(rescale=1 /
255 #rescale the tensor values to [0,1]
)
test_generator = test_datagen.flow_from_directory(directory=test_path,
classes=['real', 'fake'],
target_size=(input_size,
input_size),
color_mode="rgb",
class_mode=None,
batch_size=1,
shuffle=False)
# Train a CNN classifier
efficient_net = EfficientNetB0(weights='imagenet',
input_shape=(input_size, input_size, 3),
include_top=False,
pooling='max')
model = Sequential()
model.add(efficient_net)
model.add(Dense(units=512, activation='relu'))
model.add(Dropout(0.5))
model.add(Dense(units=128, activation='relu'))
model.add(Dense(units=1, activation='sigmoid'))
model.summary()
# Compile model
model.compile(optimizer=Adam(lr=0.0001),
loss='binary_crossentropy',
metrics=['accuracy'])
def train_model(CONFIG):
checkpoint_path = CONFIG['checkpoint_path'].format(
experiment_i=CONFIG['experiment_i'])
checkpoint_dir = os.path.dirname(checkpoint_path)
# Create a callback that saves the model's weights every 1 epoch
cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_path,
verbose=1,
save_weights_only=True,
save_best_only=True)
#early_stopping_callback = tf.keras.callbacks.EarlyStopping(
# monitor='val_loss', min_delta=0, patience=5, verbose=0, mode='min',
# baseline=None, restore_best_weights=True
#)
tensorboard_callback = tf.keras.callbacks.TensorBoard(
log_dir=CONFIG['log_dir'].format(experiment_i=CONFIG['experiment_i']),
update_freq='epoch')
callbacks = [tensorboard_callback, cp_callback]
image_dir = CONFIG['image_dir']
base_model = EfficientNetB0(weights='noisy-student', include_top=False)
base_model.trainable = True
x = base_model.output
x = GlobalAveragePooling2D()(x)
x = Dense(1000, activation='relu')(x)
x = Dropout(rate=0.2)(x)
preds = Dense(200, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=preds)
train_datagen = ImageDataGenerator(rotation_range=20,
zoom_range=0.15,
width_shift_range=0.2,
height_shift_range=0.2,
shear_range=0.15,
horizontal_flip=True,
fill_mode="nearest")
val_datagen = ImageDataGenerator()
train_generator = train_datagen.flow_from_directory(
directory=os.path.join(image_dir, 'train'),
batch_size=32,
target_size=(224, 224),
seed=3,
shuffle=True,
class_mode="categorical")
val_generator = val_datagen.flow_from_directory(directory=os.path.join(
image_dir, 'val'),
batch_size=32,
target_size=(224, 224),
seed=3,
shuffle=False,
class_mode="categorical")
model.compile(optimizer='Adam',
loss='categorical_crossentropy',
metrics=['accuracy'])
step_size_train = train_generator.n // train_generator.batch_size
step_size_val = val_generator.n // val_generator.batch_size
model.fit(train_generator,
validation_data=val_generator,
validation_steps=step_size_val,
steps_per_epoch=step_size_train,
epochs=CONFIG['epochs'],
callbacks=callbacks)
path_to_save = CONFIG['saved_model'].format(
experiment_i=CONFIG['experiment_i'])
print(f"Saving model to {path_to_save}...")
model.save(path_to_save)
print("Done")
return model
