Keras model 로딩해서 layer 변경하기

on device learning for android 

안드로이드에서 on device model 을 base (head)모델과 fully connected layer 로 나누기 위해 

Keras SavedModel 로딩해서 뒤에 몇개 layer 를 잘라내었다. 

 

import os
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
import tensorflow as tf
import tensorflow.python.keras.models
import tensorflow_hub as hub


def create_model():
    # 모델 define
    model = tf.keras.models.Sequential([
        # The first convolution
        tf.keras.layers.Conv2D(16, (3, 3), activation='relu', input_shape=(300, 300, 3)),
        tf.keras.layers.MaxPool2D(2, 2),
        # The second convolution
        tf.keras.layers.Conv2D(32, (3, 3), activation='relu'),
        tf.keras.layers.MaxPool2D(2, 2),
        # The third convolution
        tf.keras.layers.Conv2D(64, (3, 3), activation='relu'),
        tf.keras.layers.MaxPool2D(2, 2),
        # The fourth convolution
        tf.keras.layers.Conv2D(64, (3, 3), activation='relu'),
        tf.keras.layers.MaxPool2D(2, 2),
        # The fifth convolution
        tf.keras.layers.Conv2D(64, (3, 3), activation='relu'),
        tf.keras.layers.MaxPool2D(2, 2),
        # Flatten
        tf.keras.layers.Flatten(),
        # 512 Neuron (Hidden layer)
        tf.keras.layers.Dense(512, activation='relu'),
        # 1 Output neuron
        tf.keras.layers.Dense(1, activation='sigmoid')
    ])


    return model;


loaded = tf.keras.models.load_model("./training_2")
loaded.summary()

loadedModel = tf.keras.models.load_model("./training_2")
base_input  = loadedModel.input
base_output = loadedModel.layers[8].output

newModel = tf.keras.models.Model(inputs= base_input, outputs = base_output)
print(newModel.summary())


print("Done")

 

 

원본 모델 

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 298, 298, 16)      448       
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 149, 149, 16)      0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 147, 147, 32)      4640      
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 73, 73, 32)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 71, 71, 64)        18496     
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 35, 35, 64)        0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 33, 33, 64)        36928     
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 16, 16, 64)        0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 14, 14, 64)        36928     
_________________________________________________________________
max_pooling2d_4 (MaxPooling2 (None, 7, 7, 64)          0         
_________________________________________________________________
flatten (Flatten)            (None, 3136)              0         
_________________________________________________________________
dense (Dense)                (None, 512)               1606144   
_________________________________________________________________
dense_1 (Dense)              (None, 1)                 513       
=================================================================
Total params: 1,704,097
Trainable params: 1,704,097
Non-trainable params: 0
____________________________________

 

수정한 모델 

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_input (InputLayer)    [(None, 300, 300, 3)]     0         
_________________________________________________________________
conv2d (Conv2D)              (None, 298, 298, 16)      448       
_________________________________________________________________
max_pooling2d (MaxPooling2D) (None, 149, 149, 16)      0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 147, 147, 32)      4640      
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 73, 73, 32)        0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 71, 71, 64)        18496     
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 35, 35, 64)        0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 33, 33, 64)        36928     
_________________________________________________________________
max_pooling2d_3 (MaxPooling2 (None, 16, 16, 64)        0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 14, 14, 64)        36928     
=================================================================