当我包含优化程序时，TF sess.run（）创建CustomOp的新对象

Question

Following the example from here [ https://www.tensorflow.org/extend/adding_an_op] , I created my own operator, let's call it MyConv .

Now, if I make something like this:

# placeholders for inputs and labels
input = tf.placeholder(...)
label = tf.placeholder(...)

# MyConv operator on inputs... returns 1D tensor
conv_out = my_module.my_conv(input, ...)
# final classifier 
output = tf.layers.dense(inputs=conv_out, units=10)

# eveything we need for backprop
probs = tf.losses.sparse_softmax_cross_entropy(logits=output, labels=label)
loss = tf.reduce_mean(probs)
adam_opt = tf.train.AdamOptimizer(learning_rate=0.01)
train_op = adam_opt.minimize(loss)

And run the model N-1 times with a simple input forwarding, and only the Nth time with the backprop:

num_of_chunks = 100

init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    for c in range(num_of_chunks):
        single_input = np.array(...)
        # for num_of_chunks-1 times, just push the single_input through the network
        if c != num_of_chunks - 1:
            result = sess.run([output], feed_dict={input:single_input})
        # only the last time, do the backprop and run the optimizer
        else:
            result, loss_out, train_out = sess.run([output, loss, train_op], feed_dict={input:single_input, label:np.array(...)})

An interesting thing occurs. TF sess.run() method will instantiate one object of MyConv class and use it num_of_chunks-1 times, but for the last time (when it executes this else path) it will create a new object of class MyConv and use that new object (MyConv() constructor has been invoked one more time). It looks like, just because in the else path I have introduced the backprop, sess.run() creates new objects of MyConv operator, and that's not really a behavior I have expected (I store some information there as a private class member, and I would really like to keep one MyConv object while running the session).

So, the question is: do you know why would sess.run() method create a new object of an operator just because I have switched from 'only-forward mode' to 'forward and backprop mode'?

Also, one note. If I remove this if-else statement and let the backprop to be executed for every chunk, all is good, only one object of MyConv is instantiated and used.