Setting random seed when finetuning

[mesotron]

Is there a simple way to set the random seed to a constant at the beginning of finetuning, so that finetuning the same model on the same corpus for the same number of iterations will not result in different parameters each time? I have tried calling "tf.random.set_random_seed(0)" just before "sess = gpt2.start_tf_sess()", but that doesn't seem to have worked.

If I have to dig into the source code to do this, any tips on where/how would be appreciated.