xuanandsix / gfpgan-onnxruntime-demo Goto Github PK

nice work!

i run at 4090，the infer time is: 0.6s，a little long. Is this the normal effect?

Running Ubuntu 22.04 LTS I get the following error:

Traceback (most recent call last): File "/home/user/GFPGAN-onnxruntime-demo/torch2onnx.py", line 67, in <module> ort_session = onnxruntime.InferenceSession(onnx_model_path) File "/home/user/miniconda3/envs/richard-roop/lib/python3.10/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 396, in __init__ raise e File "/home/user/miniconda3/envs/richard-roop/lib/python3.10/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 383, in __init__ self._create_inference_session(providers, provider_options, disabled_optimizers) File "/home/user/miniconda3/envs/richard-roop/lib/python3.10/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 415, in _create_inference_session raise ValueError( ValueError: This ORT build has ['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider'] enabled. Since ORT 1.9, you are required to explicitly set the providers parameter when instantiating InferenceSession. For example, onnxruntime.InferenceSession(..., providers=['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider'], ...)

To fix the issue I modify line 67 in torch2onnx.py:

From:
ort_session = onnxruntime.InferenceSession(onnx_model_path)

To:
ort_session = onnxruntime.InferenceSession(onnx_model_path, providers=['TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider'])

I have noted a quality decrease when comparing the results to GFPGAN 1.4 on Replicate.

Any reasons why that could be happening?

I have used your script and converted gfpgan pth model into onnx model but I am getting horrible results. Like output image is same as input image (in terms of dimensions/size) and the output gets worst after inference. In short, it is not doing super resolution on the image and also making current input image bad.

@xuanandsix So, what could be the possible reason behind this problem ? ( Btw, I have done all the steps properly that you have mentioned in the readme file.)

Input Image:

Output Image:

great work 😊

Hi, is there any chance I can do batch inference with ONNX model, I notice the example is do single imge inference. Does the onnx model generated from torch2onnx.py support batch inference?

Such a great work. I am looking for quicker inference for GFPGAN. Is there any chance we can load the GFPGAN.onnx and run it on TensorRT? Looking forward to your suggestion.

Thanks.

我按照上述步骤成功将1.4和1.3版本的模型导出成onnx模型了，经过我测试，似乎只在输入的图像仅包含头部的时候效果最好。我有个想法，将一张全身照的图片的头部截下来，调整尺寸到512后运行得到生成的图像，再将图像调整回原来的尺寸后放会原来裁剪的位置，但这种方式会导致裁剪出的图像边缘有条纹，结果如下所示：请问这种问题该如何解决？

I will release a PR which supports the batching of the model via ONNX.
I have a protoype working.

在GPU环境下，使用onnxruntime的时间和PyTorch基本是一致的。