Skip to content

traing error on distill_hy_detail_expert #22

@weihualiuhupituzi

Description

@weihualiuhupituzi

Dear Author,

I have successfully completed the first-stage training of distill_hy_semantic_expert. However, when proceeding to distill the detail expert model using the aforementioned results, I encountered the following error:

File "/root/work/filestorage/weihua/DCM-main/fastvideo/distill/discriminator.py", line 45, in forward
[rank3]: x = x.view(-1, 49 * 80, c)
[rank3]: RuntimeError: shape '[-1, 3920, 3072]' is invalid for input of size 9768960

Upon reviewing the code, I traced the issue to the Discriminator function. The input tensor x has a shape of [1, 3180, 3072], which is not divisible by 49 × 80.

Questions:​​

  1. What does 49 × 80 represent in this context? Is it related to a fixed feature map size or a structural constraint?
  2. Would it be possible to adjust this value (or the input dimensions) to ensure alignment with the tensor shape?
    I’d appreciate any guidance on resolving this dimensionality mismatch. Thank you for your time and support!

Best regards,

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions