python - 如果手动使用 BCE 损失，所有梯度值计算为“无”

Question

我正在研究一个多输出模型，我需要在计算整体损失之前权衡所有输出损失。我有一个定制的model. fit() 训练循环来实现这一点。

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，因此我自定义了标准代码。现在，损失是按样本计算的，但是在计算梯度时，所有梯度值都计算为“无”。我也试着放了tape.watch(loss)，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}

更新 我按照@rvinas的建议做了一些更改 现在它正在计算梯度而没有任何错误，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}

标签： pythontensorflowkerasdeep-learning

Answer 1

NaN 梯度的出现是因为您正在使用 NumPy 操作（例如np.sum, np.reshape, ...），这会导致图形断开连接。相反，只需要使用 tensorflow 操作来实现逻辑。

---

例如，可以实现评论部分中描述的权重，如下所示：

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)