RuntimeError: Missing out variants: {'aten::alias'} #1132
Assignees
Labels
need-user-input
The issue needs more information from the reporter before moving forward
triaged
This issue has been looked at a team member, and triaged and prioritized into an appropriate module
wontfix
This will not be worked on
Comments
|
Hmm alias shouldnt be appearing only alias_copy. cc @SS-JIA to take a look |
|
Do you have an example model that we can reproduce this issue on our end? |
|
@mergennachin |
|
@adonnini can you print out the program? I'm thinking maybe we should remove this node, just want to verify if it's a noop. |
|
I thought it would be easier and give you the information you are seeking if I sent you the link to the github repository I got the model from I added the executorch code to train.py after Please let me know if you need anything else |
|
I tried the following code: It failed with the following error: line 326 in train.py is |
|
re flatc: if you run |
|
Looking at your code it seems it should be Notice that you should call |
BTW I'm not able to follow your instruction to run |
|
assinging to you @larryliu0820 |
|
Answering in order of occurrence:
seems to work now. It produces a .pte file of around 200MB please note that m is an instance of my model (i.e. I called the constructor)
|
|
Update on code execution. After running successfully for four epochs, the execution failed with the error listed below. Please note that the line numbers of model.py and train.py listed in the traceback do not correspond to the line numbers in the model on github as I made some small changes to the code. |
|
Do you mind sharing your code? |
|
Below you will find a link to the github repository with my code and dataset. A couple of points to note:
Please let me know if you have any questions or encounter any problems |
|
@adonnini are you running training on exported model? If so, are the input sizes changing from epoch to epoch? |
|
@adonnini it seems like you are trying to export after every training epoch. One suspect I have is that you may be using different input shapes in each epoch. Can you provide a minimum repro? For example, we would really appreciate it if you can give a code snippet that only contains the model and the input, and the code to export it. |
|
@larryliu0820 did you try to run the code I sent you? If you did, did it fail as I reported? |
|
Okay before we go into specifics of this particular issue and debugging, let's step back a bit. Could you elaborate what kind of problem you are trying to solve and how executorch fits in your scenario? It looks like you are using executorch for training, which is not the intended use. As of today, we don't support training. ExecuTorch is an inference engine for on-device deployment. We expect developers to do training (either in eager mode or compiled mode). Once they have a trained model, we expect them to use torch.export and to_executorch to generate an ExecuTorch program artifact once, so that they can deploy inference for edge/embedded devices. If you are trying to speed up the training, we recommend to use torch.compile instead. |
|
@mergennachin Thanks for your help. I think I resolved my problem. I am all set for now. I will probably need assistance (questions and issue resolution) once I attempt to deploy for inference on Android devices. |
mergennachin
added
need-user-input
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
I was able to have
exir.capturerun the trace of my model (I think). However, now the code fails with the error listed below. Could yo please take a look and let me know what you think I am doing wrong and what I should do next?Thanks
The text was updated successfully, but these errors were encountered: