如果手动使用 BCE 损失，则所有梯度值计算为“无”

我正在研究一个多输出模型，在计算总体损失之前，我需要权衡所有输出损失。我有一个定制的model. fit() 训练循环 https://www.tensorflow.org/guide/keras/customizing_what_happens_in_fit为了达成这个。

由于我需要计算所有四个输出的样本损失并在应用权重后融合这些样本损失，因此我定制了标准代码。现在，损失是按样本计算的，但在计算梯度时，所有梯度值都计算为“无”。我试着把tape.watch(loss)也，但它不起作用。请帮我解决这个问题。

class CustomModel(keras.Model):
    def train_step(self, data):
        print(tf.executing_eagerly())
        # Unpack the data. Its structure depends on your model and
        # on what you pass to `fit()`.
        x, y = data
        alpha = 0.1
        loss = 0
        y_pred_all = []

        with tf.GradientTape() as tape:
            bce = tf.keras.losses.BinaryCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
            for spl in range(1 if np.shape(x)[0] == None else np.shape(x)[0]):
                tape.watch(loss)
                tape.watch(loss_mean)
                tape.watch(loss_element)
                x_spl = np.reshape(x[spl], (1, np.shape(x)[1], np.shape(x)[2], np.shape(x)[3]))
                y_pred = self(x_spl, training=True)  # Forward pass
                y_pred_all.append(y_pred)
                loss_element = bce(y[spl], y_pred)
                loss_mean = [np.mean(loss_element[0]), np.mean(loss_element[1]), np.mean(loss_element[2]), np.mean(loss_element[3])]
                id = np.argmin(loss_mean)
                for i, ele in enumerate(loss_mean):
                    if i == id:
                        loss_mean[i] *= 1
                    else:
                        loss_mean[i] *= alpha

                loss = loss + np.sum(loss_mean)

        # Compute gradients
        trainable_vars = self.trainable_variables

        gradients = tape.gradient(loss, trainable_vars)

        # Update weights
        self.optimizer.apply_gradients(zip(gradients, trainable_vars))

        # Update metrics (includes the metric that tracks the loss)
        self.compiled_metrics.update_state(y, y_pred_all)
        # Return a dict mapping metric names to current value
        return {m.name: m.result() for m in self.metrics}

UPDATE我按照建议做了一些更改@rvinas现在它正在计算梯度，没有任何错误，但我不确定我所做的更改是否正确：

class CustomModel(keras.Model):
    def train_step(self, data):
        # print(tf.executing_eagerly())
        # Unpack the data. Its structure depends on your model and
        # on what you pass to `fit()`.
        x, y = data
        alpha = 0.1
        loss = tf.Variable(0, dtype='float32')
        y_pred_all = []

        with tf.GradientTape() as tape:
            bce = tf.keras.losses.BinaryCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
            for spl in tf.range(1 if tf.shape(x)[0] == None else tf.shape(x)[0]):
                loss_mean=tf.convert_to_tensor([])
                x_spl =  tf.reshape(x[spl], (1, tf.shape(x)[1], tf.shape(x)[2], tf.shape(x)[3]))
                y_pred = self(x_spl, training=True)  # Forward pass
                y_pred_all.append(y_pred)
                loss_element = bce(y[spl], y_pred)
                loss_mean = [tf.reduce_mean(loss_element[0]), tf.reduce_mean(loss_element[1]), tf.reduce_mean(loss_element[2]), tf.reduce_mean(loss_element[3])]

                id = tf.argmin(loss_mean)
                for i, ele in enumerate(loss_mean):
                    if i == id:
                        loss_mean[i] = tf.multiply(loss_mean[i], 1)
                    else:
                        loss_mean[i] = tf.multiply(loss_mean[i], alpha)

                loss = tf.add(loss, tf.add(tf.add(tf.add(loss_mean[0],loss_mean[1]), loss_mean[2]), loss_mean[3]))

        # Compute gradients
        trainable_vars = self.trainable_variables

        gradients = tape.gradient(loss, trainable_vars)

        # Update weights
        self.optimizer.apply_gradients(zip(gradients, trainable_vars))

        # Update metrics (includes the metric that tracks the loss)
        self.compiled_metrics.update_state(y, y_pred_all)
        # Return a dict mapping metric names to current value
        return {m.name: m.result() for m in self.metrics}

出现 NaN 梯度是因为您正在使用 NumPy 运算（例如np.sum, np.reshape, ...)，这会导致图表断开连接。相反，我们只需要使用张量流运算来实现逻辑。

例如，可以按如下方式实现评论部分中描述的权重：

bce = tf.keras.losses.BinaryCrossentropy(reduction=tf.keras.losses.Reduction.NONE)
with tf.GradientTape() as tape:
    # Compute element-wise losses
    y_pred = self(x, training=True)
    losses = bce(y, y_pred)  # Shape=(bs, 4)

    # Find maximum loss for each sample
    idx_max = tf.argmax(losses, axis=-1)  # Shape=(bs,)
    idx_max_onehot = tf.one_hot(idx_max, depth=y.shape[-1])  # Shape=(bs, 4)

    # Create weights tensor
    weight_max = 1
    weight_others = 0.1
    weights = idx_max_onehot * weight_max + (1 - idx_max_onehot) * weight_others

    # Aggregate losses
    losses = tf.reduce_sum(weights * losses, axis=-1)
    loss = tf.reduce_mean(losses)