How to get gradients with respect to input and change input (rather than trainable vars) to minimize loss in TF 2?

Question

I want to use a trained model to change the input so it minimizes the loss (rather than changing the trainable variables) a la Deep Dreaming in Tensorflow 2.0 but I am not having success.

Say I have a basic NN as the one in the docs

class MyModel(Model):
  def __init__(self):
    super(MyModel, self).__init__()
    self.conv1 = Conv2D(32, 3, activation='relu')
    self.flatten = Flatten()
    self.d1 = Dense(128, activation='relu')
    self.d2 = Dense(10, activation='softmax')

  def call(self, x):
    x = self.conv1(x)
    x = self.flatten(x)
    x = self.d1(x)
    return self.d2(x)

model = MyModel()

Which I train using a simple tf.GradientTape function

@tf.function
def train_step(image, label):
  with tf.GradientTape() as tape:
    predictions = model(image)
    loss = loss_object(label, predictions)
  gradients = tape.gradient(loss, model.trainable_variables)
  optimizer.apply_gradients(zip(gradients, model.trainable_variables))

What's the idiomatic way to create a function that will instead calculate and apply the gradients to the input - images.

I assumed it will be as simple as

def train_step(image, label):
  with tf.GradientTape() as tape:
    predictions = model(image)
    loss = loss_object(label, predictions)
  gradients = tape.gradient(loss, image)
  optimizer.apply_gradients(zip(gradients, image))

However, that doesn't work.

Answer 1

tf.GradientTape.gradients can only differentiate wrt to a watched tensor. Variables are automatically watched on first access. In order to differentiate wrt an arbitrary tensor, you have to explicitly watch it:

>>> x = tf.constant([4.0])
>>> y = tf.constant([2.0])
>>> with tf.GradientTape() as tape:
...     tape.watch([x, y])
...     z = x * y
...     
>>> tape.gradient(z, [x, y])
[<tf.Tensor: id=9, shape=(1,), dtype=float32, numpy=array([ 2.], dtype=float32)>, 
 <tf.Tensor: id=10, shape=(1,), dtype=float32, numpy=array([ 4.], dtype=float32)>]