Save model checkpoint

[choidam]

공식문서 참고하기

https://www.tensorflow.org/tutorials/keras/save_and_load?hl=tr

[choidam]

sample model & save checkpoints

# save checkpoint
checkpoint_path = "training_2/cp-{epoch:04d}.ckpt"
checkpoint_dir = os.path.dirname(checkpoint_path)

cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_path,
                                                 save_weights_only=True,
                                                 period=5,
                                                 verbose=1)

newmodel = create_model()
newmodel.save_weights(checkpoint_path.format(epoch=0))

newmodel.fit(dataset, 
           epochs=40,
           shuffle=True,
           validation_data=val_dataset,
           callbacks=[LearningRateScheduler(scheduler, verbose=1), callbacks, cp_callback],
           workers=4)

[choidam]

check directory

latest = tf.train.latest_checkpoint(checkpoint_dir)
latest

load weights to new model

# Create a new model instance
model2 = create_model()

# Load the previously saved weights
model2.load_weights(latest)

train model2

model2.fit(dataset, 
           epochs=40,
          #  shuffle=True,
           validation_data=val_dataset,
           callbacks=[LearningRateScheduler(scheduler, verbose=1), callbacks]),
           workers=4)

확실히 epoch1부터 정확도는 높지만 훈련도가 전혀 증가하지 않는다 ... ㅠㅠ 문제가 뭘까 .. 처음부터 쓰면 안 되는 방법이었나 ...?...