我在尝试在Keras中实现一个自定义损失函数，但无法使其正常工作。

我已经在numpy和keras.backend中实现了它：

def log_rmse_np(y_true, y_pred):    d_i = np.log(y_pred) -  np.log(y_true)    loss1 = (np.sum(np.square(d_i))/np.size(d_i))    loss2 = ((np.square(np.sum(d_i)))/(2 * np.square(np.size(d_i))))    loss = loss1 - loss2    print('np_loss =  %s - %s = %s'%(loss1, loss2, loss))    return lossdef log_rmse(y_true, y_pred):    d_i = (K.log(y_pred) -  K.log(y_true))    loss1 = K.mean(K.square(d_i))    loss2 = K.square(K.sum(K.flatten(d_i),axis=-1))/(K.cast_to_floatx(2) * K.square(K.cast_to_floatx(K.int_shape(K.flatten(d_i))[0])))    loss = loss1 - loss2    return loss

当我使用以下函数测试并比较损失时，一切似乎都正常工作。

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(log_rmse(K.variable(y_a), K.variable(y_b)))    out2 = log_rmse_np(y_a, y_b)    print('shapes:', str(out1.shape), str(out2.shape))    print('types: ', type(out1), type(out2))    print('log_rmse:    ', np.linalg.norm(out1))    print('log_rmse_np: ', np.linalg.norm(out2))    print('difference:  ', np.linalg.norm(out1-out2))    assert out1.shape == out2.shape    #assert out1.shape == shape[-1]def test_loss():    shape_list = ['2d', '3d', '4d', '5d']    for _shape in shape_list:        check_loss(_shape)        print ('======================')test_loss()

上述代码打印如下内容：

np_loss =  1.34490449177 - 0.000229461787517 = 1.34467502998shapes: () ()types:  <class 'numpy.float32'> <class 'numpy.float64'>log_rmse:     1.34468log_rmse_np:  1.34467502998difference:   3.41081509703e-08======================np_loss =  1.68258448859 - 7.67580654591e-05 = 1.68250773052shapes: () ()types:  <class 'numpy.float32'> <class 'numpy.float64'>log_rmse:     1.68251log_rmse_np:  1.68250773052difference:   1.42057615005e-07======================np_loss =  1.99736933814 - 0.00386228512295 = 1.99350705302shapes: () ()types:  <class 'numpy.float32'> <class 'numpy.float64'>log_rmse:     1.99351log_rmse_np:  1.99350705302difference:   2.53924863358e-08======================np_loss =  1.95178217182 - 1.60006871892e-05 = 1.95176617114shapes: () ()types:  <class 'numpy.float32'> <class 'numpy.float64'>log_rmse:     1.95177log_rmse_np:  1.95176617114difference:   3.78277884572e-08======================

当我使用这个损失函数编译并训练我的模型时，从未出现异常，并且当我使用’adam’优化器运行模型时，一切正常工作。然而，使用这个损失函数时，Keras总是显示NaN损失：

Epoch 1/10000 17/256 [>.............................] - ETA: 124s - loss: nan

我现在有点卡住了… 我做错了什么吗？

我在Ubuntu 16.04上使用Tensorflow 1.4

更新：

根据Marcin Możejko的建议，我更新了代码，但遗憾的是训练损失仍然是NaN：

def get_log_rmse(normalization_constant):    def log_rmse(y_true, y_pred):        d_i = (K.log(y_pred) -  K.log(y_true))        loss1 = K.mean(K.square(d_i))        loss2 = K.square(K.sum(K.flatten(d_i),axis=-1))/K.cast_to_floatx(2 * normalization_constant ** 2)        loss = loss1 - loss2        return loss    return log_rmse

然后通过以下方式编译模型：

model.compile(optimizer='adam', loss=get_log_rmse(batch_size))

更新2：

模型摘要如下所示：

Layer (type)                 Output Shape              Param #   =================================================================input_2 (InputLayer)         (None, 160, 256, 3)       0         _________________________________________________________________block1_conv1 (Conv2D)        (None, 160, 256, 64)      1792      _________________________________________________________________block1_conv2 (Conv2D)        (None, 160, 256, 64)      36928     _________________________________________________________________block1_pool (MaxPooling2D)   (None, 80, 128, 64)       0         _________________________________________________________________block2_conv1 (Conv2D)        (None, 80, 128, 128)      73856     _________________________________________________________________block2_conv2 (Conv2D)        (None, 80, 128, 128)      147584    _________________________________________________________________block2_pool (MaxPooling2D)   (None, 40, 64, 128)       0         _________________________________________________________________block3_conv1 (Conv2D)        (None, 40, 64, 256)       295168    _________________________________________________________________block3_conv2 (Conv2D)        (None, 40, 64, 256)       590080    _________________________________________________________________block3_conv3 (Conv2D)        (None, 40, 64, 256)       590080    _________________________________________________________________block3_pool (MaxPooling2D)   (None, 20, 32, 256)       0         _________________________________________________________________block4_conv1 (Conv2D)        (None, 20, 32, 512)       1180160   _________________________________________________________________block4_conv2 (Conv2D)        (None, 20, 32, 512)       2359808   _________________________________________________________________block4_conv3 (Conv2D)        (None, 20, 32, 512)       2359808   _________________________________________________________________block4_pool (MaxPooling2D)   (None, 10, 16, 512)       0         _________________________________________________________________conv2d_1 (Conv2D)            (None, 10, 16, 1)         513       =================================================================Total params: 7,245,777Trainable params: 7,245,777Non-trainable params: 0_________________________________________________________________

更新3：

示例y_true如下：

回答：