Full explanation by clicking here!

import numpy as np
import keras
from keras.models import load_model
import keras.backend as K
import matplotlib.pyplot as plt
import os

features = 50*11

x0 = np.ones((1, 11, 50), dtype=float)*100
model = load_model('model2.h5')

eps = 1
target = np.array([0, 1])

#If you want to find the 0 gradient, just plug this loss inside the derivative.
loss = keras.losses.categorical_crossentropy(target, model.output)

# Gradient wrt to last dense layer! Why?
# Because the softmax flattens everything to zero (try and see, just swap the definitions)

session = K.get_session()
d_model_d_x = K.gradients(model.layers[2].output, model.input)
# d_model_d_x = K.gradients(model.output, model.input)

prediction = model.predict(x0)
while np.argmax(prediction) != 1:

    # Thank you Keras + Tensorflow!
    # That [0][0] is just ugly, but it is needed to obtain the value as an array.
    eval_grad = session.run(d_model_d_x, feed_dict={model.input: x0})
    eval_grad = eval_grad[0][0]

    # We don't need to cap the sign, hence, we need to craft the flag point.
    # The FSGM resude uniformly the image in general, because it flattens the gradient to barely +/-1.
    # We need pixels to be modified according to the value of the gradient in their position.
    fsgm = eps * eval_grad

    # The gradient always points to maximum ascent direction, but we need to minimize.
    # Hence, we swap the sign of the gradient.
    x0 = x0 - fsgm

    # We do not need to clip, as we don't need to save the result as a regular image.
    prediction = model.predict(x0)
    print(prediction)

# I will rescale the output for visualization purpouses.
plt.imshow(x0[0] / np.max(x0[0]) - np.min(x0[0]))
plt.show()