[Keras] 모델 불러오기

 

 

2020/09/23 - [Python/Deep Learning] - [Keras] 모델 저장하기

[Keras] 모델 저장하기

 

저장하는 방법엔 3가지 방법이 있습니다.

 

• ModelCheckpoint 
• model.save()
• to_json(), to_yaml() and save_weight

1,2번째 방법은  따로 가중치만 저장(weights)을 설정하지 않았다면 불러오는 방법이 같고 밑의 방법은 약간 다른 방법을 써야 합니다. 

---

1. load.model

from keras.models import load_model

model = load_model('save_model.h5')

 

h5 형식으로 저장된 모델을 불러올 수 있습니다. 이렇게 불러오는 모델은 컴파일까지 다된 상태로 불러오는 것이라 바로 사용할 수 있습니다.

 

2.from_json(), from_yaml() and load_weight

from keras.models import model_from_json
with open("./save/model_json.json", "r") as model_json :
	loaded_model = model_json.read()

loaded_model = model_from_json(loaded_model)

loaded_model.load_weights("./save/save_weight.h5")
print("Loaded model from disk")

loaded_model.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy'])

json 파일과 yaml파일은 from_json, yaml을 통해 모델 아키텍처를 불러오고 load_weight를 불러와 가중치를 더한뒤

컴파일하여 사용합니다. 예시는 json 파일을 불러오는 경우 입니다.

Model: "model_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 12, 12, 3)         0         
_________________________________________________________________
separable_conv2d_1 (Separabl (None, 10, 10, 256)       1051      
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 5, 5, 256)         0         
_________________________________________________________________
separable_conv2d_2 (Separabl (None, 5, 5, 512)         133888    
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 2, 2, 512)         0         
_________________________________________________________________
separable_conv2d_3 (Separabl (None, 2, 2, 1024)        527360    
_________________________________________________________________
global_max_pooling2d_1 (Glob (None, 1024)              0         
_________________________________________________________________
dense_1 (Dense)              (None, 256)               262400    
_________________________________________________________________
re_lu_1 (ReLU)               (None, 256)               0         
_________________________________________________________________
dense_2 (Dense)              (None, 64)                16448     
_________________________________________________________________
re_lu_2 (ReLU)               (None, 64)                0         
_________________________________________________________________
dense_3 (Dense)              (None, 27)                1755      
=================================================================
Total params: 942,902
Trainable params: 942,902
Non-trainable params: 0
_________________________________________________________________
Loaded model from disk
Model: "model_1"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 12, 12, 3)         0         
_________________________________________________________________
separable_conv2d_1 (Separabl (None, 10, 10, 256)       1051      
_________________________________________________________________
max_pooling2d_1 (MaxPooling2 (None, 5, 5, 256)         0         
_________________________________________________________________
separable_conv2d_2 (Separabl (None, 5, 5, 512)         133888    
_________________________________________________________________
max_pooling2d_2 (MaxPooling2 (None, 2, 2, 512)         0         
_________________________________________________________________
separable_conv2d_3 (Separabl (None, 2, 2, 1024)        527360    
_________________________________________________________________
global_max_pooling2d_1 (Glob (None, 1024)              0         
_________________________________________________________________
dense_1 (Dense)              (None, 256)               262400    
_________________________________________________________________
re_lu_1 (ReLU)               (None, 256)               0         
_________________________________________________________________
dense_2 (Dense)              (None, 64)                16448     
_________________________________________________________________
re_lu_2 (ReLU)               (None, 64)                0         
_________________________________________________________________
dense_3 (Dense)              (None, 27)                1755      
=================================================================
Total params: 942,902
Trainable params: 942,902
Non-trainable params: 0
_________________________________________________________________

두 가지 방법으로 저장된 모델을 model.summay()를 통해 비교해보면 똑같음을 알 수 있습니다.