How can I measure the complexity of the model?

Research & Models
, ,
#1

I need to measure time and memory complexity for a keras model or captioning model using keras … how can I start ?
Thanks

#3

Check:

https://tensorflow-prod.ospodiscourse.com/t/how-to-find-out-keras-model-memory-size/5249

#5

Thanks a lot for replying … excuse me confirm just i got the point or not please …
i read the links and got to measure the time and memory consumption, i need to add those lines to my model

run_metadata = tf.RunMetadata()
with tf.Session() as sess:
_ = sess.run(train_op,
options=tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE),
run_metadata=run_metadata)

after the model finished , i should run tf.profiler.profile ?
what if i finished the training stages and got the final file with h5 at epoch 20 … is there another way to measure time and memory instead start training again

#6

You can approximate memory with some calc of input, number parameter, dtype etc… but to really profile your model you need to run it.

#8

Thanks a lot … is there any tutorial for these calculations ? In what title I should search for this , please

#10

You can check Ability to calculate projected memory usage for a given model · Issue #36327 · tensorflow/tensorflow · GitHub

But I suggest you to estimate It at runtime with the previous solution.

1 Like
#11

thanks a lot but excuse me what will be the difference between the runtime and the other way ?

#12

Cause at runtime it is what it is going to really consume the library on the machine e.g. kernel memory requirements and any other allocation, data transfer etc…

1 Like