Avoid unnecessary retracing

[malihass]

The calc_loss function should not be a tf.function since its behavior changes depending on whether discriminator or generator, or both are trained.
With the tf.function on, function retracing occurs (we see messages like WARNING:tensorflow:5 out of the last 18 calls to <function Sup3rGan.calc_loss at 0x7facf82f8290> triggered tf.function retracing...).

With tf.function off, no retracing occurs. I could also see a small speed up when running on my machine.

To reproduce results, run the following snippet in tests/

from test_train_gan import *


import time
time_start = time.time()
test_train_st_weight_update(n_epoch=5, log=True)
time_end = time.time()
print("Total Time = %g" % (time_end - time_start))

[grantbuster]

Fine by me! What kind of speedup are you seeing?

[malihass]

I ran the snippet I showed 5 times to get statistics more rigorously about the execution time.
With the PR : 124.05 +/- 9.17 s
Without the PR: 133.199 +/- 7.47359s

[grantbuster]

Gotcha, thanks! I keep hoping for a weird error i made that results in a 10x speedup lol.