if i run tensorflow on my machine or google codelab , why both are different

Question

Training works best if the examples are in random order. Use tf.data.Dataset.shuffle to randomize entries, setting buffer_size to a value larger than the number of examples (120 in this case). To train the model faster, the dataset's batch size is set to 32 examples to train at once.

train_dataset = tf.data.TextLineDataset(train_dataset_fp)
train_dataset = train_dataset.skip(1)             # skip the first header row
train_dataset = train_dataset.map(parse_csv)      # parse each row
train_dataset = train_dataset.shuffle(buffer_size=1000)  # randomize
train_dataset = train_dataset.batch(32)

features, label = iter(train_dataset).next()
>>> print("example features:", features[0])
example features: tf.Tensor([5.1 3.7 1.5 0.4], shape=(4,), dtype=float32)
>>> print("example label:", label[0])
example label: tf.Tensor(0, shape=(), dtype=int32)

tensorflow output

example features: tf.Tensor([6.  2.7 5.1 1.6], shape=(4,), dtype=float32)
example label: tf.Tensor(1, shape=(), dtype=int32)

Comments

[YouheiSakurai]

What’s your problem? You mean some code working on google codelab but not on your machine, don’t you?

[AliHassan]

no , both are running but results little different from one and another

[AliHassan]

GOOGLE CODLAB P TENSORFLOW 1.9 RC PRODUCE 98% TRAINING and 97% test . While my machine produce 70% for training and 53% test

[YouheiSakurai]

Garbage question-in, garbage (or no) answer-out. Describe your question once again carefully if you want to hear something different from my “don’t know”. Your question sounds like worse than students’.

[YouheiSakurai]

Clues (and contexts) are basically missing in your question, that enables others to identify by what, for what, how and so on you are blocked. If you don’t have idea of how to do it, please let me know so and I will give you guidance.

[AliHassan]

let suppose we have two machines , machine a produce 97% results and machine b produce 96% for same example . reason why these results are little different from one and another

[YouheiSakurai]

Difference makes difference. Also, randomness should be considered if everything looks the same.

[AliHassan]

thanks

Answer 1

The results should differ even with the same machine if you do not fix the random seed.
If you are using tensorflow on the colab with GPU, you might not be able to fix the seed.

https://github.com/tensorflow/tensorflow/issues/2732

Did cross validate your model to comfirm the variance?

Comments

[AliHassan]

thankyou sir
no

[YouheiSakurai]

Should? I supposed “may differ” depending on how complex the cost function is. (Don’t take this comment wrongly, I am just curious to know it precisely and exactly.)

[mkgrei]

> train_dataset = train_dataset.shuffle(buffer_size=1000)

I thought the dataset is always shuffled.
And if it is shuffled and the random seed is not fixed, then the dataset should differ.

https://www.tensorflow.org/api_docs/python/tf/set_random_seed

[YouheiSakurai]

Ah, question was asking why first items of features and labels are different (between machines). Thanks, agree. That should differ.

[AliHassan]

okay thanks , i randomly assigned batch number , but this issue resolved. thank you for help