How can I define the initial weights used by an autoencoder?

Question

When I run my autoencoder several times with a fixed set of parameters I have some different results on accuracy and other metrics. I suppose this is because initially the neural network chooses random weights. I want the results are always the same. Can I modify the random choice of weights? How can I make my result deterministic and not random?

can you add more information about what language/package/libraries you are using? — venkata krishnan, Jun 28 '19 at 07:48
could you maybe give the code also? Please add the infos to the question, don't leave them in the comments — Frayal, Jun 28 '19 at 07:49
See the excelent answer [here](https://stackoverflow.com/questions/32419510/how-to-get-reproducible-results-in-keras). Voting to close as duplicate. — dedObed, Jun 28 '19 at 07:58
this solution doensn't work for me. I run my code ( I can't attach this :() for 3 times and I have different solutions for accuracy and other metrics — Serena89, Jun 28 '19 at 08:59

score 2 · Answer 1 · answered Jun 28 '19 at 08:37

You need to set the seed

I usually use a this simple function bellow

def keras_seeding(seednum):
    np.random.seed(seednum)
    from tensorflow import set_random_seed
    set_random_seed(seednum)
    random.seed(seednum)
    os.environ['PYTHONHASHSEED'] = str(seednum)

seednum=1123
keras_seeding(seednum)

score 0 · Answer 2 · answered Jun 28 '19 at 07:53

If you are using Keras, at every layer you can mention the weight intializer. You can read more about the various types of initializers available in this link.

Example : 

model = Sequential()
model.add(Conv2D(64,(3,3),kernel_initializer="he_normal"))
model.add(Dense(10,kernel_initializer='ones'))

I have just mentioned the sample initializers here. Look for the one best suited for your case.

How can I define the initial weights used by an autoencoder?

2 Answers2