Inception v3 网络的标签平滑

Question

In inception v3 paper they discussed about label smoothing but in (keras code) here there is nothing mention about label smoothing. How to write code for label smoothing for inception v3 network. It is suggested that it should be included in loss function. i am using categorical crossentropy loss function. So how to make custom categorical crossentropy for label smoothing.

Answer 1

You can pass a label_smoothing parameter to the function CategoricalCrossentropy for label smoothing.

import tensorflow as tf

# when compiling the model
model.compile('sgd', loss=tf.keras.losses.CategoricalCrossentropy(label_smoothing=0.1))