Bug in TF2 PGD implementation

[meta-inf]

Describe the bug
In future.tf2.attacks.projected_gradient_descent, the sanity check clip_min<= x <= clip_max is implemented incorrectly: here you are appending a tensor to asserts, and a subsequent call of np.all on asserts invoked the __bool__ cast of the tensor, which raises the "The truth value of an array with more than one element is ambiguous" error.

To Reproduce

Use the following code:

import tensorflow as tf
import numpy as np
from cleverhans.future.tf2.attacks import projected_gradient_descent

def model_fn(x):
  return tf.nn.softmax(x, axis=-1)

X = [[1., 2., 3.], [4., 5., 6.]]
print(projected_gradient_descent(
  model_fn, X, 1., 1., 5, np.inf, 0., 7.))

Expected behavior
Code shouldn't crash.

System configuration

• Ubuntu 18.04.3
• Python 3.6.0
• TensorFlow 2.0.0

[CNOCycle]

I have implemented a TF2 version for PGD attack in my fork project. The issues should be fixed.

https://github.com/CNOCycle/cleverhans/tree/feature/tf2.x