ValueError：没有为任何变量提供渐变：['Variable:0']

Question

I am trying to prepare an example notebook for showing how to create adversarial examples in TensorFlow 2.x following this tutorial .

I have been able to port some of it but I am now stuck with a weird problem of ValueError: No gradients provided for any variable: ['Variable:0']. and can't seem to figure out why.

Here's how I load the image for which I am trying to create an adversary:

def preprocess_image(image_pixels):
    # image_pixels = preprocess_input(image_pixels)
    image_pixels = cv2.resize(image_pixels, (224, 224))
    image_pixels = np.expand_dims(image_pixels, axis=0)

    return image_pixels

# Load and preprocess image but a but without any preprocess_input
sample_image = show_image("pig.jpg")
preprocessed_image = preprocess_image(sample_image)

# Initialize the perturbation quantity
image_tensor = tf.convert_to_tensor(preprocessed_image, dtype=tf.float32)
delta = tf.Variable(tf.zeros_like(image_tensor), trainable=True)

Here's how my training loop looks like -

for t in range(30):
    inp = preprocess_input(image_tensor + delta)
    with tf.GradientTape() as tape:
        tape.watch(delta)
        pred = resnet50(inp)
        loss = - scc_loss(
                tf.convert_to_tensor([341]),
                pred
            )
        if t % 5 == 0:
            print(t, loss.numpy())

        # Get the gradients
        gradients = tape.gradient(loss, delta)

        # Update the weights
        optimizer.apply_gradients([(gradients, delta)])

        # Clip so that the delta values are within [0,1]
        delta.assign(clip_eps(delta))

where, preprocess_input is tf.keras.applications.resnet50 and my model is a pre-trained ResNet50. The optimizer is defined as optimizer = tf.keras.optimizers.SGD(learning_rate=1e-1) . Error is listed below:

0 -0.00041249825
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-284-174284e3aea4> in <module>()
     15 
     16         # Update the weights
---> 17         optimizer.apply_gradients([(gradients, delta)])
     18 
     19         # Clip so that the delta values are within [0,1]

1 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py in apply_gradients(self, grads_and_vars, name, experimental_aggregate_gradients)
    472       ValueError: If none of the variables have gradients.
    473     """
--> 474     grads_and_vars = _filter_grads(grads_and_vars)
    475     var_list = [v for (_, v) in grads_and_vars]
    476 

/usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/optimizer_v2/optimizer_v2.py in _filter_grads(grads_and_vars)
   1201   if not filtered:
   1202     raise ValueError("No gradients provided for any variable: %s." %
-> 1203                      ([v.name for _, v in grads_and_vars],))
   1204   if vars_with_empty_grads:
   1205     logging.warning(

ValueError: No gradients provided for any variable: ['Variable:0'].

Here's where the entire Colab notebook can be found. Appreciate any help/pointers.

Answer 1

Your delta (the trainable variable) is not participating in any operation inside the tape. It can't be watched if it is not doing anything.

To have a better performance, since you're using eager execution, use training=False when you use a model that is not supposed to be updated (the tape will watch every trainable kernel in the models unless you do this, any batch normalization and dropout will also behave different) -- Since you said you're using a ResNet for preprocessing, I assume this ResNet should use training=False .

for t in range(30):
    with tf.GradientTape() as tape:
        tape.watch(delta)
        inp = preprocess_input(image_tensor + delta)
        pred = resnet50(inp)
        loss = - scc_loss(
                tf.convert_to_tensor([341]),
                pred
            )
        if t % 5 == 0:
            print(t, loss.numpy())

#these things should not be inside the tape:
    # Get the gradients
    gradients = tape.gradient(loss, delta)

    # Update the weights
    optimizer.apply_gradients([(gradients, delta)])

    # Clip so that the delta values are within [0,1]
    delta.assign(clip_eps(delta))