如何使用 tensorflow 损失 function 与 keras Z20F35E630DAF44DBDFA4C3F68F539？

Question

I am using Tensorflow 1.14, and I have designed a model using Keras. I want to use tf.nn.sparse_softmax_cross_entropy_with_logits when I compile my model, but I do not see any equivalent loss function in Keras. Is there any way I can use this with my model?

My current code to compile:

model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])

Thank you for any help.

Answer 1

You can use the tf.losses.sparse_categorical_crossentropy with from_logits set to True and wrap it in a function

import tensorflow as tf

def my_tf_loss_fn(y_true, y_pred):
    return tf.losses.sparse_categorical_crossentropy(y_true, y_pred, from_logits=True)

model = tf.keras.applications.ResNet50()
model.compile(loss=my_tf_loss_fn, optimizer='adam')

But if you insist on using tf.nn.sparse_softmax_cross_entropy_with_logits I cannot think of a clean way to do this, but this works

import tensorflow as tf

def my_tf_loss_fn(y_true, y_pred):
    return tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y_true, logits=y_pred)

model = tf.keras.applications.ResNet50()
dummy_tensor = tf.placeholder(dtype=tf.int32, shape=[None])
model.compile(loss=my_tf_loss_fn, optimizer='adam', target_tensors=dummy_tensor)