Tf.Print () causing an error with gradients

Question

I tried using tf.Print debug statements to better understand the format of reported gradients and variables from compute_gradients (), but ran into an unexpected problem. The training procedure and the debugging procedure (gvdebug) are as follows:

def gvdebug(g, v):
    #g = tf.Print(g,[g],'G: ')
    #v = tf.Print(v,[v],'V: ')
    g2 = tf.zeros_like(g, dtype=tf.float32)
    v2 = tf.zeros_like(v, dtype=tf.float32)
    g2 = g
    v2 = v
    return g2,v2

# Define training operation
def training(loss, global_step, learning_rate=0.1):
    optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
    grads_and_vars = optimizer.compute_gradients(loss)
    gv2 = [gvdebug(gv[0], gv[1]) for gv in grads_and_vars]
    train_op = optimizer.apply_gradients(gv2, global_step=global_step)
    return train_op

This code works fine (but doesn't print), but if I uncomment the two tf.Print lines in gvdebug (), I get an error from apply_gradients: 'TypeError: Variable should be tf.Variable'. I thought tf.Print just went through tensors - what am I doing wrong?

Answer 1

TL DR

tf.Print gv[1], tf.Variable. , gradient gv[0].

compute_gradients gradients tf.Variable s.

grads_and_vars Tensor a tf.Variable. , .

v = tf.Print(v,[v],'V: ')