python - Keras自定义损失函数错误: 'AttributeError: ' function' object has no attribute 'get_shape'

Question

我必须编写自己的自定义损失函数，该函数可以采用 Keras 中除 y_true 和 y_pred 参数之外的不同输入。在阅读了一些解决方法后，我决定使用内部函数，如下所示:

from keras import backend as K

lambda_prn_regr = 0.6
lambda_prn_vis = 0.2
lambda_prn_class = 0.2

epsilon = 1e-4

# Person loss
def prn_loss_cls(y_true, y_pred):
    def prn_loss_cls_fixed_num(y_true, y_pred):
        # lambda * b_ce
        return lambda_prn_class * K.mean(K.binary_crossentropy(y_true, y_pred), axis=-1)
    return prn_loss_cls_fixed_num

# Regression loss
def prn_loss_regr(num_joints):
    def prn_loss_regr_fixed_num(y_true, y_pred):
        # lambda * sum(vis * (pose_pred - pose_true)^2) / sum(vis)
        return lambda_prn_regr * K.sum(y_true[:, :, :, :2*num_joints] * K.square(y_pred - y_true[:, :, :, 2*num_joints:])) / K.sum(y_true[:, :, :, :2*num_joints])
    return prn_loss_regr_fixed_num

# Visibility Loss
def prn_loss_vis(y_true, y_pred):
    def prn_loss_regr_fixed_num(y_true, y_pred):
        return lambda_prn_vis * K.mean(K.square(y_pred - y_true), axis=-1)
    return prn_loss_regr_fixed_num

三种不同的损失函数:每种都有权重，其中一种需要整数参数。

但是在执行 model.compile 函数时出现 AttributeError: 'function' object has no attribute 'get_shape' 错误。整个错误输出如下:

Traceback (most recent call last):
  File "train_mppn.py", line 97, in <module>
    model_prn.compile(optimizer=optimizer, loss=[losses.prn_loss_cls, losses.prn_loss_regr(C.num_joints), losses.prn_loss_vis(C.num_joints)])
  File "/usr/local/lib/python2.7/dist-packages/keras/engine/training.py", line 899, in compile
    sample_weight, mask)
  File "/usr/local/lib/python2.7/dist-packages/keras/engine/training.py", line 441, in weighted
    ndim = K.ndim(score_array)
  File "/usr/local/lib/python2.7/dist-packages/keras/backend/tensorflow_backend.py", line 439, in ndim
    dims = x.get_shape()._dims
AttributeError: 'function' object has no attribute 'get_shape'

编译部分:

model.compile(optimizer=optimizer, loss=[losses.prn_loss_cls, losses.prn_loss_regr(num_joints), losses.prn_loss_vis])

我找不到问题的根源。

Answer 1

您传递的函数不返回值，它们返回函数。

在 num_joints 情况下这样做是可以理解的(并且您实际上正在调用该函数)，但在其他情况下这很奇怪，特别是因为您没有在任何地方调用它们返回内部函数。

建议:

# Person loss
def prn_loss_cls(y_true, y_pred):
       return lambda_prn_class * K.mean(K.binary_crossentropy(y_true,y_pred), axis=-1)


# Visibility Loss
def prn_loss_vis(y_true, y_pred):
    return lambda_prn_vis * K.mean(K.square(y_pred - y_true), axis=-1)