Скачать или смотреть Transfer Learning | Feature Extraction | Dogs vs Cats | Tensorflow-gpu | Achieved 99.5% Accuracy

This implementation uses advanced transfer learning and feature extraction and achived 99.5% accuracy in dogs vs cats kaggle problem.

https://github.com/jiahualihuanahuan/...

The write_gap function is here below for those of you are interested.

def write_gap(MODEL, image_size, preprocess):
width = image_size[0]
height = image_size[1]
input_tensor = tf.keras.Input((height, width, 3))
x = input_tensor

base_model = MODEL(input_tensor=x, weights='imagenet', include_top=False)
model = tf.keras.Model(base_model.input, tf.keras.layers.GlobalAveragePooling2D()(base_model.output))

gen = tf.keras.preprocessing.image.ImageDataGenerator(preprocessing_function=preprocess)
train_generator = gen.flow_from_directory(train_data_directory, image_size, shuffle=False,
batch_size=batch_size, class_mode='binary')
test_generator = gen.flow_from_directory(test_data_directory, image_size, shuffle=False,
batch_size=batch_size, class_mode='binary')

print('calculating training set')
train = model.predict(train_generator, len(train_generator))
print('calculating test set')
test = model.predict(test_generator, len(test_generator))
print('write file to disk')
with h5py.File(f"./dogs_vs_cats/tensorflow/features/gap_{model_name}.h5", 'w') as h:
h.create_dataset("train", data=train)
h.create_dataset("test", data=test)
h.create_dataset("label", data=train_generator.classes)