Not updating weights when tried to mix with Keras

[v-i-s-h]

While trying to update the parameters of a model which is created using Keras and train_op in tensorflow, the parameters are not getting updated.
Below is a minimum working example

# Test for mixing Keras and TF for MLP

from __future__ import print_function

import tensorflow.keras as keras
from tensorflow.keras.models import Model
from tensorflow.keras.layers import Input, Dense
from tensorflow.keras.optimizers import SGD
import tensorflow.keras.backend as K

import tensorflow as tf

import Optimizers as Opt
# from spsa import SimultaneousPerturbationOptimizer  # Replace with this

import numpy as np

K.clear_session()
tf.reset_default_graph()

# ---------------- Create a dataset -------------------------
n_train = 1000
n_test = 100

y_train = np.random.randint(2, size=(n_train,2))
y_test = np.random.randint(2, size=(n_test,2))

x_train = 0.5*np.random.randn(n_train, 2) + y_train
x_test = 0.5*np.random.randn(n_test,2) + y_test

# --------------- Create model using Keras -------------------

with tf.name_scope("Model"):
  inputs = Input(shape=(2,))
  l1 = Dense(4, activation='relu')(inputs)
  l2 = Dense(4, activation='relu')(l1)
  predictions = Dense(2, activation='softmax')(l2)

model = Model(inputs=inputs, outputs=predictions)

model.summary()


# --------------- Create place holder and loss with TF------
with tf.name_scope("Model"):
  y_ph_tf = tf.placeholder(tf.float32, shape=[None, 2])
  loss = tf.losses.softmax_cross_entropy(y_ph_tf, predictions)

params = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, "Model")

# trainer = tf.train.GradientDescentOptimizer(learning_rate=0.5)   # This works
trainer = Opt.SimultaneousPerturbationOptimizer(a=0.05,c=0.05,alpha=0.99,gamma=0.40)   # This don't
_train = trainer.minimize(loss)

sess = K.get_session()

sess.run(tf.global_variables_initializer())

print(">>>>>> PRE")
print(sess.run(params[:2]))   # Print 

for i in range(20):
  l, _ = sess.run([loss,_train], feed_dict={inputs:x_train,y_ph_tf:y_train})
  
print(">>>>>> POST")
print(sess.run(params[:2]))

This produces output as below

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
input_1 (InputLayer)         (None, 2)                 0         
_________________________________________________________________
dense (Dense)                (None, 4)                 12        
_________________________________________________________________
dense_1 (Dense)              (None, 4)                 20        
_________________________________________________________________
dense_2 (Dense)              (None, 2)                 10        
=================================================================
Total params: 42
Trainable params: 42
Non-trainable params: 0
_________________________________________________________________
SPSA:   a = 0.050  c = 0.050  alpha = 0.990  gamma = 0.400
42 parameters
WARNING:tensorflow:VARIABLES collection name is deprecated, please use GLOBAL_VARIABLES instead; VARIABLES will be removed after 2017-03-02.
>>>>>> PRE
[array([[ 3.8956022e-01, -3.0612946e-01,  3.6983490e-02, -2.1815300e-04],
       [ 6.0597062e-01,  1.7913413e-01, -6.0641050e-01, -2.9886746e-01]],
      dtype=float32), array([0., 0., 0., 0.], dtype=float32)]
>>>>>> POST
[array([[ 3.8956022e-01, -3.0612946e-01,  3.6983490e-02, -2.1815300e-04],
       [ 6.0597062e-01,  1.7913413e-01, -6.0641050e-01, -2.9886746e-01]],
      dtype=float32), array([0., 0., 0., 0.], dtype=float32)]

At "POST"-update, there is no change in value of parameters.

Using tf-1.9.