Keras 中的自定义损失函数（IoU 损失函数）和梯度误差？

Posted 2023-03-28

【中文标题】Keras 中的自定义损失函数（IoU 损失函数）和梯度误差？

【英文标题】：Customized Loss function in Keras (IoU loss function) and gradient error?

【发布时间】：2021-02-24 12:53:28

【问题描述】：

我是 ML 新手，我正在尝试实现我自己的损失函数（IoU 损失函数），但得到关于梯度的错误（“没有为任何变量提供梯度”）。 请注意，我正在尝试预测将代表矩阵中行号的数字（例如 y_pred = [1 5 3 9]）。在损失函数中，我应该计算正确预测的行数，然后将其除以 y_true 的元素总数。由于应该尽量减少损失，所以我在函数末尾放了 1-IoU。

这是我的功能，因为我不知道如何准确解释我的问题，所以更清楚：

**def** loss_IoU(y_true, y_pred):

  intersection = []
  #roundig the predicted values y_pred, since the values could be floats ex. [1.5 2 8.98 ...]
  Roundy_pred = tf.round(y_pred)   # (none, 6) with shape of none, and the size of the outputs 6.
  intersection = tf.math.count_nonzero( Roundy_pred == y_true, axis=1, keepdims=True)   # (none, 1) count the number of equal values for each row (i.e. for each predicted y and a true y)
  union = y_true.shape[1] # which is 6 here
  iou = intersection/union
  returned_iou = 1 - iou
  return returned_iou

这是我得到的错误：

    ValueError                                Traceback (most recent call last)
<ipython-input-176-631f68b50b34> in <module>()
      9 history = model.fit(AS_Training_Set, Label_Training_Set, 
     10                     steps_per_epoch=8, epochs=600, validation_data=
---> 11           (AS_Validation_Set, Label_Validation_Set))

10 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/func_graph.py in wrapper(*args, **kwargs)
    971           except Exception as e:  # pylint:disable=broad-except
    972             if hasattr(e, "ag_error_metadata"):
--> 973               raise e.ag_error_metadata.to_exception(e)
    974             else:
    975               raise

ValueError: in user code:

    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:806 train_function  *
        return step_function(self, iterator)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:796 step_function  **
        outputs = model.distribute_strategy.run(run_step, args=(data,))
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:1211 run
        return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:2585 call_for_each_replica
        return self._call_for_each_replica(fn, args, kwargs)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:2945 _call_for_each_replica
        return fn(*args, **kwargs)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:789 run_step  **
        outputs = model.train_step(data)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:757 train_step
        self.trainable_variables)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:2737 _minimize
        trainable_variables))
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:562 _aggregate_gradients
        filtered_grads_and_vars = _filter_grads(grads_and_vars)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py:1271 _filter_grads
        ([v.name for _, v in grads_and_vars],))

    ValueError: No gradients provided for any variable: ['conv2d_16/kernel:0', 'conv2d_16/bias:0', 'conv2d_17/kernel:0', 'conv2d_17/bias:0', 'batch_normalization_8/gamma:0', 'batch_normalization_8/beta:0', 'conv2d_18/kernel:0', 'conv2d_18/bias:0', 'conv2d_19/kernel:0', 'conv2d_19/bias:0', 'batch_normalization_9/gamma:0', 'batch_normalization_9/beta:0', 'conv2d_20/kernel:0', 'conv2d_20/bias:0', 'conv2d_21/kernel:0', 'conv2d_21/bias:0', 'batch_normalization_10/gamma:0', 'batch_normalization_10/beta:0', 'conv2d_22/kernel:0', 'conv2d_22/bias:0', 'conv2d_23/kernel:0', 'conv2d_23/bias:0', 'batch_normalization_11/gamma:0', 'batch_normalization_11/beta:0', 'dense_4/kernel:0', 'dense_4/bias:0', 'dense_5/kernel:0', 'dense_5/bias:0'].

谁能帮我解决这个错误，我尝试修复它但没有任何结果，我不知道问题是否与舍入函数有关。

从现在开始谢谢。

【参考方案1】：

对于 IoU 损失函数，我将这个用于 Pascal VOC 数据集。

def IoU_loss(y_true, y_pred):
    nb_classes = K.int_shape(y_pred)[-1]
    iou = []
    pred_pixels = K.argmax(y_pred, axis=-1)
    for i in range(0, nb_classes):  # exclude first label (background) and last label (void)
        true_labels = K.equal(y_true[:, :, 0], i)
        pred_labels = K.equal(pred_pixels, i)
        inter = tf.cast(true_labels & pred_labels, dtype=tf.int32)
        union = tf.cast(true_labels | pred_labels, dtype=tf.int32)
        legal_batches = K.sum(tf.cast(true_labels, dtype=tf.int32), axis=1) > 0
        ious = K.sum(inter, axis=1) / K.sum(union, axis=1)
        iou.append(K.mean(ious[legal_batches]))

    iou = tf.stack(iou)
    legal_labels = ~tf.math.is_nan(iou)
    iou = iou[legal_labels]
    return K.mean(iou)

它需要进行一些更改，但它也适用于您。

【讨论】：

非常感谢，会试一试，希望对我有用。