CODEWITHSUNDEEP
python-3.xtensorflowneural-networkconv-neural-networkrecurrent-neural-network
Mitra Lanka
Mitra Lanka

Reputation: 69

Result changes every time I run Neural Network code

I got the results by running the code provided in this link Neural Network – Predicting Values of Multiple Variables. I was able to compute losses accuracy etc. However, every time I run this code, I get a new result. Is it possible to get the same (consistent) result?

Upvotes: 4

Views: 4870

Answers (3)

Palguna Gopireddy
Palguna Gopireddy

Reputation: 81

You have to set the seed for different things like numpy, cuda etc.. Then your program gives same result. Use below function

def set_random_seed(random_seed):
    torch.manual_seed(random_seed)
    torch.cuda.manual_seed(random_seed)
    torch.cuda.manual_seed_all(random_seed)  # if use multi-GPU

    torch.backends.cudnn.deterministic = True
    torch.backends.cudnn.benchmark = False

    np.random.seed(random_seed)
    random.seed(random_seed)

Then call with a seed set_random_seed(1234) This will give you same result no matter on which machine you run. You can change the seed if you want. Different seed results in different results.

Upvotes: 1

Celius Stingher
Celius Stingher

Reputation: 18367

The code is full of random.randint() everywhere! Furthermore, the weights are most of the time randomly set aswell, and the batch_size also has an influence (although pretty minor) in the result.

  1. Y_train, X_test, X_train are generated randomly
  2. Using adam as optimizer, means you'll be performing stochastic gradient descent. With a random beginning point of the iterations in order to converge.
  3. A batch_size of 8 means you will run batches consisting of 8 randomly selected samples.

Solution:

  1. Set a random seed in your code to have always the random values generated with np.random.seed()
  2. Doesn't generate much of an issue although minor deviations
  3. Same as 2.

If I find a way to have consistente sampling methods for the batch_size/epoch issue I will edit my answer.

Upvotes: 4

Matt Hall
Matt Hall

Reputation: 8112

There are lots of random arrays in there. Use np.random.seed() to get the same ones each time. For example:

np.random.seed(42)
for _ in range(3):    
    print(np.random.random(3))

Every time you run this code, you'll get the same result. On my machine:

[0.37454012 0.95071431 0.73199394]
[0.59865848 0.15601864 0.15599452]
[0.05808361 0.86617615 0.60111501]

Note that lots of other bits of the machine learning pipeline use randomization too. For example:

  • Splitting into train, validation and test datasets with train_test_split().
  • Setting initial weights in a neural network.
  • Optimization pathways.

Most ML functions allow you to pass a seed as an argument. Have a look in the documentation. Depending on what you are doing, and which libraries you're using, you may or may not be able to make the entire pipeline reproducible.

You might also like this article or this one about getting reproducible results with Keras.

Upvotes: 2

Related Questions