summelon

@Derdin-datascience Not work under mirrored strategy

Working on converting model which has FFT layer from Tensorflow 2.X -> OpenVINO IR. To those who need a workaround to successfully convert fft/ifft op., this [white paper](https://www.intel.com/content/dam/www/public/us/en/ai/documents/openvino-and-devcloud-accelerate-cs-image-reconstruction-algorithms-for-mri.pdf) may be...

I guess you guys are using `pytorch-lightning==1.2.1`. Downgrading to 1.1.8 may fix this problem.

I encountered the same bug when I tried to load the segmentation dataset which has a high resolution. When I was saving 4kx4k gt & prediction mask to the sample...

Same here. I converted ViT-H encoder to float16 and used the float32 decoder. The prediction is totally wrong. Do you guys have any work around on this?

+1. Unseen domain + hierarchical objects are a big challenge for current foundation model , e.g., SAM, DINO etc. AFAIF, not only in this repo, the finetuning for such tasks...

Thanks for your solution! I encountered the same problem under TensorRT 10.0.1.6 / 8.6.3. It's worth noting that I tried on TensorRT 8.6.1, and trt did not report any conflict....

Thanks for your reply! So the ver.8.6.3 is the latest stable ver. for vision model before major ver. 10. Do you know which pull request is coping with the GEMM...

Hi, @lix19937. I tried `polygraphy run decoder.onnx --trt --onnxrt --input-shapes image_embeddings:[1,256,64,64]` w/ or w/o `--builder-optimization-level 5`. The difference did not change and still significant on 10.0.1

> This problem is caused by IABCEMdetr loss function. In this function, the presence_gamma's default value is 0. If the model correctly predicts entire presences, then the backward would be...