losses.py

import tensorflow as tf

# def SoftmaxLoss():
#     """softmax loss"""
#     def softmax_loss(y_true, y_pred):
#         # y_true: sparse target
#         # y_pred: logist
#         y_true = tf.cast(tf.reshape(y_true, [-1]), tf.int32)
#         ce = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y_true,
#                                                             logits=y_pred)
#         return tf.reduce_mean(ce)
#     return softmax_loss

def softmax_loss(y_true, y_pred):
    # y_true: sparse target
    # y_pred: logist
    y_true = tf.cast(tf.reshape(y_true, [-1]), tf.int32)
    ce = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=y_true,
                                                        logits=y_pred)
#     tf.print("target: ", y_true)
#     tf.print("logit: ", y_pred)
#     tf.print("softmax_loss: ", tf.reduce_mean(ce))
    return tf.reduce_mean(ce)