jianzongwu / betrayed-by-captions Goto Github PK

Hi everyone,

Is there anyone can reproduce accuracy scores in the paper?
I have tried to use 2 Nvidia Geforce 3090 but the scores were dropped by 5%.

How to achieve paper's accuracy?

Thanks.

Hi, can you tell me how to evaluate the performance of caption generation according to your code?

open_set里提供的coco_panoptic_open.py文件不支持bbox。

        allowed_metrics = ['PQ']
        for metric in metrics:
            if metric not in allowed_metrics:
                raise KeyError(f'metric {metric} is not supported')

The demo notebook fails with a cuda error, possibly relating to incorrect matrix shapes during multiplications. I'm using the exact same torch/cuda/mmcv/mmdet versions as listed in the README, on a Quaddro RTX 8000 GPU.

The stack trace I get after running the cell

result = inference_detector(model, img, with_caption=True, logging=True)[0]

is

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
Cell In[8], [line 2](vscode-notebook-cell:?execution_count=8&line=2)
      [1](vscode-notebook-cell:?execution_count=8&line=1) # Predict segmentation results, as well as image captions
----> [2](vscode-notebook-cell:?execution_count=8&line=2) result = inference_detector(model, img, with_caption=True, logging=True)[0]
...
--> [360](~/miniforge3/envs/cgg/lib/python3.8/site-packages/torch/functional.py:360) return _VF.einsum(equation, operands)

RuntimeError: CUDA error: CUBLAS_STATUS_INVALID_VALUE when calling `cublasSgemmStridedBatched( handle, opa, opb, m, n, k, &alpha, a, lda, stridea, b, ldb, strideb, &beta, c, ldc, stridec, num_batches)

Have you encountered this before? And any way to fix this?

Great Job!
When will your group release the code?

I want to use this model as a baseline to generate pseudo labels for my task.

Hi, thanks for releasing your great work. However, when I run "python tools/train.py" to try to reproduce, an error appears as follows:
ImportError: cannot import name 'build_dataloader' from 'open_set.datasets'
It seems that the build_dataloader.py function is missing, could you kindly provide the missing file? Thanks a lot.