python - 如何在 Keras 中创建自定义目标函数？

Question

Keras中有很多目标函数here.

但是你如何创建自己的目标函数，我尝试创建一个非常基本的目标函数，但它给出了一个错误，我无法知道在运行时传递给该函数的参数的大小。

def loss(y_true,y_pred):
    loss = T.vector('float64')
    for i in range(1):
        flag = True
        for j in range(y_true.ndim):
            if(y_true[i][j] == y_pred[i][j]):
                flag = False
        if(flag):
            loss = loss + 1.0
    loss /= y_true.shape[0]
    print loss.type
    print y_true.shape[0]
    return loss

我得到 2 个相互矛盾的错误，

model.compile(loss=loss, optimizer=ada)
  File "/usr/local/lib/python2.7/dist-packages/Keras-0.0.1-py2.7.egg/keras/models.py", line 75, in compile
    updates = self.optimizer.get_updates(self.params, self.regularizers, self.constraints, train_loss)
  File "/usr/local/lib/python2.7/dist-packages/Keras-0.0.1-py2.7.egg/keras/optimizers.py", line 113, in get_updates
    grads = self.get_gradients(cost, params, regularizers)
  File "/usr/local/lib/python2.7/dist-packages/Keras-0.0.1-py2.7.egg/keras/optimizers.py", line 23, in get_gradients
    grads = T.grad(cost, params)
  File "/usr/local/lib/python2.7/dist-packages/theano/gradient.py", line 432, in grad
    raise TypeError("cost must be a scalar.")
TypeError: cost must be a scalar.

它说函数中返回的成本或损失必须是标量，但如果我将第 2 行从 loss = T.vector('float64')
到
loss = T.scalar('float64')

它显示了这个错误

 model.compile(loss=loss, optimizer=ada)
  File "/usr/local/lib/python2.7/dist-packages/Keras-0.0.1-py2.7.egg/keras/models.py", line 75, in compile
    updates = self.optimizer.get_updates(self.params, self.regularizers, self.constraints, train_loss)
  File "/usr/local/lib/python2.7/dist-packages/Keras-0.0.1-py2.7.egg/keras/optimizers.py", line 113, in get_updates
    grads = self.get_gradients(cost, params, regularizers)
  File "/usr/local/lib/python2.7/dist-packages/Keras-0.0.1-py2.7.egg/keras/optimizers.py", line 23, in get_gradients
    grads = T.grad(cost, params)
  File "/usr/local/lib/python2.7/dist-packages/theano/gradient.py", line 529, in grad
    handle_disconnected(elem)
  File "/usr/local/lib/python2.7/dist-packages/theano/gradient.py", line 516, in handle_disconnected
    raise DisconnectedInputError(message)
theano.gradient.DisconnectedInputError: grad method was asked to compute the gradient with respect to a variable that is not part of the computational graph of the cost, or is used only by a non-differentiable operator: <TensorType(float64, matrix)>

Answer 1

这是我编写新损失函数并在使用前测试它们的小片段:

import numpy as np

from keras import backend as K

_EPSILON = K.epsilon()

def _loss_tensor(y_true, y_pred):
    y_pred = K.clip(y_pred, _EPSILON, 1.0-_EPSILON)
    out = -(y_true * K.log(y_pred) + (1.0 - y_true) * K.log(1.0 - y_pred))
    return K.mean(out, axis=-1)

def _loss_np(y_true, y_pred):
    y_pred = np.clip(y_pred, _EPSILON, 1.0-_EPSILON)
    out = -(y_true * np.log(y_pred) + (1.0 - y_true) * np.log(1.0 - y_pred))
    return np.mean(out, axis=-1)

def check_loss(_shape):
    if _shape == '2d':
        shape = (6, 7)
    elif _shape == '3d':
        shape = (5, 6, 7)
    elif _shape == '4d':
        shape = (8, 5, 6, 7)
    elif _shape == '5d':
        shape = (9, 8, 5, 6, 7)

    y_a = np.random.random(shape)
    y_b = np.random.random(shape)

    out1 = K.eval(_loss_tensor(K.variable(y_a), K.variable(y_b)))
    out2 = _loss_np(y_a, y_b)

    assert out1.shape == out2.shape
    assert out1.shape == shape[:-1]
    print np.linalg.norm(out1)
    print np.linalg.norm(out2)
    print np.linalg.norm(out1-out2)


def test_loss():
    shape_list = ['2d', '3d', '4d', '5d']
    for _shape in shape_list:
        check_loss(_shape)
        print '======================'

if __name__ == '__main__':
    test_loss()

如您所见，我正在测试 binary_crossentropy 损失，并定义了 2 个单独的损失，一个是 numpy 版本 (_loss_np)，另一个是张量版本 (_loss_tensor) Theano 和 Tensorflow...但如果您依赖其中之一，您也可以通过 K.theano.tensor.function 或 K.tf.function 引用它们]

稍后我将比较输出形状和输出的 L2 范数(应该几乎相等)和差值的 L2 范数(应该接近 0)

一旦您对损失函数正常工作感到满意，您就可以将其用作:

model.compile(loss=_loss_tensor, optimizer=sgd)