[tensorflow] about seed

At present, all my code is based on tensorflow1.15, The way of setting seeds in the experiment is

tf.compat.v1.set_random_seed(args.seed)
np.random.seed(args.seed)
random.seed(args.seed)

After this setting , identical seed The experimental results are basically the same （ The initial parameters of the network model are constant , The result of each step is fixed ）. however , If the tf Of graph Make some unimportant changes , Such as ： Add one more useless placeholder, Add a useless full connection layer and so on , The result of each step will be similar to Previously unmodified There are differences in time .

for instance , Algorithm algo1.py and algo2.py, The logic of the latter is only one more useless than the former placeholder, Even if the two algorithms seed identical , The experimental results of each step will also be different .