How can i disable the auto_slice_resolution

[claudio9russo7]

Hi, during the preparation of a dataset I encountered a problem. I use sahi + yolov5. I sliced the images for the training of the models but just give me an error about some images truncated. As I have used sahi in the past and had no problems I have used an older version where auto_slice_resolution is not present. With the same dataset, but processed with an older version of sahi (0.9.2) I managed to train the CNN without error. I think there is a problem with auto_slice_resolution.

The error is like:
Os Error: image file is truncated (114 bytes not processed)
The environment used is kaggle

[fcakyon]

Hello @claudio9russo7, auto_slice_resolution doesn not affect coco slice command.

Can you please write down the function or script you use?

Can you also paste here the full error traceback?

[claudio9russo7]

Sure:

Just make this for slicing the dataset, after this I convert the .json file into usable format for yolo and train the network on kaggle.

[fcakyon]

Thanks @claudio9russo7!

When do you get the error? When running slice_coco function?

Can you also post the full error traceback?

Are you having this error with sahi==0.9.4 or sahi==0.10.1?

[claudio9russo7]

no this part works perfectly, but when I go to train the network with the first dataset I had that error, sliced with the old version I had no errors and the training was successful.
Error with the latest version
About the error traceback i need to re-slicing because just deleted everything as i managed to fix with older version.

EDIT: Slicing in this moment after preparation i wil try to train

[fcakyon]

Which SAHI version are you trying currently? if sahi==0.9.4 fails, you can try with sahi==0.9.3.

If sahi==0.9.3 works and sahi==0.9.4 fails, then this PR might be the reason: #445

@devrimcavusoglu do you have any idea?

[claudio9russo7]

Im trying with sahi==0.10.4 to re-check the problem

[fcakyon]

If you could try with sahi==0.9.4 it would be better. Because we already know sahi==0.10.4 has a problem.

[claudio9russo7]

As you requested I tried with sahi==0.9.4 and no error for now during the train (last time the error appeared at the first epoch), so maybe some images get corrupted durin upload/download i dont'know

[fcakyon]

It would be great if this is only a file corruption during upload/download.

Then can you try everything again with sahi==0.10.4 and see if the first epoch finishes without any error?

[claudio9russo7]

Yep, no problem but i will try tomorrow.

[claudio9russo7]

Okkei i just used sahi==0.10.4, run the training and no problem. So, i think was only an error during the upload on wandb of the dataset.