GradientTape：获取nan的渐变

Question

我正在尝试计算 tensorflow 中的梯度，但返回 None。我已经将类型调整为tensorflow.python.framework.ops.EagerTensor，但是htis并没有解决问题。

这是目前为止的代码：

accuracy = tf.keras.metrics.CategoricalAccuracy('accuracy')
loss = tf.keras.metrics.CategoricalCrossentropy('loss')
  
for epoch in range(epochs):
    accuracy.reset_states()
    loss.reset_states()
    
    for batch in iterate_minibatches(X_train, y_train, batch_size):
        imgs = batch[0]
        labels = batch[1]
        with tf.GradientTape() as tape:
            preds = model(imgs)
            
            labels = tf.convert_to_tensor(labels, dtype=tf.float32)
            
            #print(loss(labels,preds))
            # Loss is crossentropy loss with regularization term for each parameter
            total_loss = loss(labels, preds) #+l2_penalty(model, theta_A) 

        grads = tape.gradient(total_loss, model.trainable_variables)
        model.optimizer.apply_gradients(zip(grads, model.trainable_variables))
       
        accuracy.update_state(labels, preds)
        loss.update_state(labels, preds)
        print("\rEpoch: {}, Batch: {}, Loss: {:.3f}, Accuracy: {:.3f}".format(
            epoch+1, batch+1, loss.result().numpy(), accuracy.result().numpy()), flush=True, end='')
        print("")
   
print("Task B accuracy after training trained model on Task B: {}".format(model.evaluate(task_B_test)))
print("Task A accuracy after training trained model on Task B: {}".format(model.evaluate(task_A_test)))

有谁知道为什么它没有转或我该如何解决这个问题？

编辑：我的错误消息如下所示：

AttributeError Traceback（最近一次调用最后一次） C:\Users\DC5DE~1.ALB\AppData\Local\Temp/ipykernel_13300/818221091.py 在 34 grads = tape.gradient(total_loss, model.trainable_variables) 35 ---> 36 model.optimizer.apply_gradients(zip(grads, model.trainable_variables)) 37 38 accuracy.update_state(labels, preds)

AttributeError: 'NoneType' 对象没有属性 'apply_gradients'

由于我不确定这是否与我将图像数据传递给 GradientTape 的方式有关，这里是我的小批量函数：

def iterate_minibatches(inputs, targets, batchsize, shuffle=False):
    assert inputs.shape[0] == targets.shape[0]
    if shuffle:
        indices = np.arange(inputs.shape[0])
        np.random.shuffle(indices)
    for start_idx in range(0, inputs.shape[0] - batchsize + 1, batchsize):
        if shuffle:
            excerpt = indices[start_idx:start_idx + batchsize]
        else:
            excerpt = slice(start_idx, start_idx + batchsize)
        yield inputs[excerpt], targets[excerpt]

另外：here 提到了一个类似的问题，但是没有任何可行的解决方案。

Answer 1

你把一些事情搞混了。您要么需要调用model.compile，要么定义自己的优化器并使用它。此外，您不应将指标与损失函数混为一谈。这是一个工作示例：

import tensorflow as tf

accuracy = tf.keras.metrics.CategoricalAccuracy('accuracy')
metric = tf.keras.metrics.CategoricalCrossentropy('metric_ categorical_crossentropy')
loss = tf.keras.losses.CategoricalCrossentropy(from_logits=True)
epochs = 2
model = tf.keras.Sequential([
    tf.keras.layers.Dense(units=3, input_shape=(1,))
]) 
optimizer = tf.keras.optimizers.Adam()

dataset = tf.data.Dataset.from_tensor_slices((tf.random.normal((50, 1)), tf.random.normal((50, 3)))).batch(5)
for epoch in range(epochs):
    accuracy.reset_states()
    metric.reset_states()
    
    for i, batch in enumerate(dataset):
        imgs = batch[0]
        labels = batch[1]
        print(imgs.shape, labels.shape)
        with tf.GradientTape() as tape:
            preds = model(imgs)
                      
            #print(loss(labels,preds))
            # Loss is crossentropy loss with regularization term for each parameter
            total_loss = loss(labels, preds) #+l2_penalty(model, theta_A) 

        grads = tape.gradient(total_loss, model.trainable_variables)
        optimizer.apply_gradients(zip(grads, model.trainable_variables))
       
        accuracy.update_state(labels, preds)
        metric.update_state(labels, preds)
        print("\rEpoch: {}, Batch: {}, Loss: {:.3f}, Accuracy: {:.3f}".format(
            epoch+1, i+1, metric.result().numpy(), accuracy.result().numpy()), flush=True, end='')
        print("")

Epoch: 1, Batch: 1, Loss: 4.209, Accuracy: 0.200
Epoch: 1, Batch: 2, Loss: 1.641, Accuracy: 0.400
Epoch: 1, Batch: 3, Loss: 1.294, Accuracy: 0.333
Epoch: 1, Batch: 4, Loss: 1.025, Accuracy: 0.300
Epoch: 1, Batch: 5, Loss: -0.110, Accuracy: 0.320
Epoch: 1, Batch: 6, Loss: 0.316, Accuracy: 0.267
Epoch: 1, Batch: 7, Loss: -0.118, Accuracy: 0.257
Epoch: 1, Batch: 8, Loss: -0.284, Accuracy: 0.225
Epoch: 1, Batch: 9, Loss: -0.249, Accuracy: 0.244
Epoch: 1, Batch: 10, Loss: -0.464, Accuracy: 0.260
Epoch: 2, Batch: 1, Loss: 4.468, Accuracy: 0.200
Epoch: 2, Batch: 2, Loss: 1.578, Accuracy: 0.400
Epoch: 2, Batch: 3, Loss: 1.012, Accuracy: 0.400
Epoch: 2, Batch: 4, Loss: 0.836, Accuracy: 0.350
Epoch: 2, Batch: 5, Loss: -0.294, Accuracy: 0.360
Epoch: 2, Batch: 6, Loss: 0.168, Accuracy: 0.300
Epoch: 2, Batch: 7, Loss: -0.201, Accuracy: 0.286
Epoch: 2, Batch: 8, Loss: -0.634, Accuracy: 0.250
Epoch: 2, Batch: 9, Loss: -0.552, Accuracy: 0.267
Epoch: 2, Batch: 10, Loss: -0.920, Accuracy: 0.280

Answer 2

您需要使用tf.keras.losses.CategoricalCrossentropy 进行损失计算，而不是使用tf.keras.metrics.CategoricalCrossentropy，后者的工作方式不同并且会停止梯度传播。