Encoded image has ambiguous format about dali_backend HOT 3 CLOSED

Hello @KienPhung-AI-future !

The error you're facing suggests one of two possible problems:

1. You are already sending a decoded image to the DALI pipeline and then you try to decode it once again using fn.decoders.image(..., device="cpu").
2. Your input image is corrupted

Could you tell us more about the input data? Is it an image? What format is it in? Could you send the DALI pipeline you're using? This way we could help you more with the problem have.

from dali_backend.

Thank you for relying on me to specify my problem you can see here : https://github.com/KienPhung-AI-future/dali_triton

from dali_backend.

Hello @KienPhung-AI-future ,

If I understood your example correctly, you want to preprocess image with DALI and then process it using YOLO.

I don't know precisely what is the code of DALI pipeline (you attached only serialized file), but I assume you are using an image decoder there. Images kept in JPEG format on disk are encoded (compressed), so that they occupy less space. Therefore, when reading any JPEG, system conducts its decoding. That is what cv2.imread function does: reads the image file from disk and then decodes the image.

The same thing that cv2.imread does with one function, in DALI is split into several steps to leverage GPU for image decoding. When using DALI Backend, the typical approach to work with images is to create an entrypoint to the DALI pipeline with fn.external_source operator and then put fn.decoders.image operator to decode the image. Judging from the error you've pasted above, that what you did, cause this error comes from image decoder.

What I see in the example you've provided, you are feeding the DALI with already decoded images: https://github.com/KienPhung-AI-future/dali_triton/blob/de15b53b1134a91de100a1e210a03f4f3f6a6d6c/yolo/image_client_yolo_v4.py#L473. To fix it, you can do one of two things:

1. Keep the image decoding in the client as it is now (so call cv2.imread) and remove fn.decoders.image from the DALI pipeline
2. (better and recommended) Load binary image and send it to DALI pipeline to be decoded.

To see how option 2. shall be implemented, please refer to this example:

Should you have any more questions, don't hesitate to ask!

from dali_backend.