Possible memory leak in executeAsync.

[annxingyuan]

I'm using the latest tfjs (0.14.1) and calling executeAsync on the converted universal sentence encoder lite model. When I log tf.memory().numTensors before and after executeAsync, I see that 11 new tensors are created. We can't wrap executeAsync in a tidy because it's asynchronous. Would you mind looking into the possibility that there is a memory leak in executeAsync?

I created a codepen that demonstrates the issue: https://codepen.io/annxingyuan/pen/LqyGzM

Thanks!

[nsthorat]

Also confirmed this is happening in 0.15.0.

[tafsiri]

Re-opening as I am experiencing a very similar issue on version 1.0.4 of tfjs.

Here is a reproduction with universal sentence encoder. https://codepen.io/tafsiri/pen/BEOENm?editors=0012 Digging into the code a little bit on my local machine suggests that it is the executeAsync call. I also get the same number of tensors leaked as the original issue.

[tafsiri]

cc @pyu10055

[Nk911]

Facing the similar issue on tfjs 1.2.10 , After the executeAsync tf.memory().numTensors increases by 54.
Any update on when will the fix be available? @nsthorat @pyu10055

[nsthorat]

cc @pyu10055

[alvinsunyixiao]

I observed potential memory leak in executeAsync in the latest release. Can we consider re-open this issue?

Since my model is more of a differential equation simulation step, there are a lot of control flow ops in the model. It is hard to create a minimal reproducible code snippet, but my code repo, https://github.com/alvinsunyixiao/itp-websim/tree/alvin/mem, can be easily tested by running yarn && yarn start in the root.

The python code for the TF model is defined here. This script also conveniently saves the model into TFSavedModel format. The inference where the memory leak was observed is in this function. I have tried to log the number of used tensors through tf.memory().numTensors in a variety of locations, and it seems like the number of tensors grows (nondeterministically) every time executeAsync is called.