[Bug]: BSRN model compatibility issues -- GPU inference exception

[wanan]

Hi,

I’m trying to use the super-resolution network model bsrn as the openvino inference model.

The environment used is Ubuntu 20.04 and OpenVINO 2023.0, the training set is DIV2K.

First, download GitHub resources, train to obtain the Pytorch network. (loot at train.png)

Second, modify the script to convert to the onnx model.

Third, use the mo command to switch to xml + bin. (loot at mo.png)

Finally, use the demo (image_processing_demo) of the open_model_zoo to infer. (loot at used.png)

There is a significant difference in inference performance between CPU and GPU device selection, poor inference performance using GPU. Choosing different models during the training process may result in black images or significant distortion, and the specific reasons are not yet clear - Only for GPU.

Training set: https://data.vision.ee.ethz.ch/cvl/DIV2K/

Paper address: https://arxiv.org/abs/2205.05996

Code address: https://github.com/xiaom233/BSRN

./image_processing_demo -m models/model_best.xml -i images/0829x3.png -d CPU -at sr

(loot at cpu.png)

./image_processing_demo -m models/model_best.xml -i images/0829x3.png -d GPU -at sr

(loot at gpu.png)

Processed the training dataset and used intel AVS scaling algorithm and jpeg encoding and decoding to obtain small images, the inference effect using GPU is as follows, and the CPU result is normal.

(loot at gpu2.png)

[Zulkifli_Intel]

Hi Wanan,

Thank you for reaching out to us,

Please share with us which GPU you are using, ONNX model, and the optimized model for replication purposes. Also, to obtain the small images, what is the command was used?

Sincerely,

Zulkifli 

[Zulkifli_Intel]

We will continue to support the case on GitHub Thread #19355. Therefore, this case will no longer be monitored.