Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[onert/acl-cl] Fullyconnected hybrid model test fail #9928

Closed
hseok-oh opened this issue Oct 25, 2022 · 2 comments
Closed

[onert/acl-cl] Fullyconnected hybrid model test fail #9928

hseok-oh opened this issue Oct 25, 2022 · 2 comments
Labels
area/onert ONE runtime bug Something isn't working

Comments

@hseok-oh
Copy link
Contributor

After bug patch on #9926, fullyconnected/hybrid model test fail.

Fail release build, acl-cl backend on ubuntu 18.04
Passed on 20.04 (my local environment).

Run fullyconnected/hybrid
======================
[Execution] Stage start!
[Execution] Model is deserialized!
[Execution] Model compiled!
[Execution] Input data is defined!
[Execution] Done!
[Comparison] Stage start!
[Comparison] TFLite run done!
Output #0, Element Index : 256, ref: 0.0680161, act: 0.0694258 (diff: 0.00140966)
Output #0, Element Index : 258, ref: 1.85973, act: 1.85871 (diff: 0.00102198)
Output #0, Element Index : 292, ref: -2.89502, act: -2.89382 (diff: 0.00119829)
Output #0, Element Index : 350, ref: -0.616233, act: -0.617396 (diff: 0.00116301)
Output #0, Element Index : 381, ref: -1.46097, act: -1.46199 (diff: 0.00102198)
Output #0, Element Index : 403, ref: -2.49767, act: -2.49873 (diff: 0.00105739)
[Comparison] outputs is not equal!
[Comparison] Done!

Originally posted by @hseok-oh in #9926 (comment)

@hseok-oh hseok-oh added bug Something isn't working area/onert ONE runtime labels Oct 25, 2022
@hseok-oh
Copy link
Contributor Author

hseok-oh commented Oct 28, 2022

Maybe too large size FC operation makes quantization error.

[#0  (MAIN)]

#0  (MAIN) input tensors: [0]
    %0    : buffer (Empty)                   | FLOAT32 | Memory 158.4K | Shape [4, 10137] (b'decoder/one_hot')
#0  (MAIN) output tensors: [1]
    %1    : buffer (Empty)                   | FLOAT32 | Memory 4.0K   | Shape [4, 256] (b'decoder/target_embed/MatMul')

#0 FULLY_CONNECTED
  Fused Activation: NONE
  Input Tensors[%0, %3, %2]
    %0    : buffer (Empty)                   | FLOAT32 | Memory 158.4K | Shape [4, 10137] (b'decoder/one_hot')
    %3    : buffer (&2   : [[ ... ]])        | UINT8   | Memory 2.5M   | Shape [256, 10137] (b'target_embed/kernel/u/transpose')
    %2    : buffer (&1   : [ ... ])          | FLOAT32 | Memory 1.0K   | Shape [256] (b'decoder/target_embed/MatMul_bias')
  Output Tensors[%1]
    %1    : buffer (Empty)                   | FLOAT32 | Memory 4.0K   | Shape [4, 256] (b'decoder/target_embed/MatMul')

Number of all operator types          :    1
  FULLY_CONNECTED                     :    1
Number of all operators               :    1

Expected TOTAL  memory: 2.6M
Expected FILLED memory: 2.5M

But it is strange why there is not error on debug build.

@hseok-oh
Copy link
Contributor Author

hseok-oh commented Dec 13, 2024

It is resolved by ARMCompute version up. #14454

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
area/onert ONE runtime bug Something isn't working
Projects
None yet
Development

No branches or pull requests

1 participant