harisreedhar / face-upscalers-onnx Goto Github PK

How do I use the codeformer.onnx model.

from pth to onnx?
I try it ,but the resout of pth and onnx is different/
can u share your code?

I'm planning to create an app in Unity3D using barracuda. the model size seems larger than expected. Can you help me to create a low size models like tflite?

I have tested gfpganv1.4.onnx on my custom image. However, it leads to a quality decrease (compared with gfpganv1.4.pth).

Thus, I want to know the reasons and figure out how to achieve the same quality as gfpganv1.4.pth. My test code and results are attached below.

# gfpgan v1.4 pytorch
from gfpgan import GFPGANer

def load_sr():
    run_params = GFPGANer(
        model_path="checkpoints/GFPGANv1.4.pth",
        upscale=1,
        arch="clean",
        channel_multiplier=2,
        bg_upsampler=None,
    )
    return run_params

def upscale(image, properties):
    _, _, output = properties.enhance(
        image, has_aligned=False, only_center_face=False, paste_back=True
    )
    return output

# inference with onnx
from GFPGAN.GFPGAN import GFPGAN

def load_sr_onnx():
    gfpganv14 = GFPGAN(model_path="GFPGANv1.4.onnx", device="cuda")
    return gfpganv14

def upscale_onnx(image, model):
    return model.enhance(image)

I need that, because my GPU is AMD, and only works with NCNN, and with CPU only takes a LOT time to do an upscale.
I've tried to convert any of this models with chaiNNer from ONNX or PTH to NCNN without success.
Previously was able to convert some models from PTH or ONNX.
I don't know nothing about to code, maybe a simple bash script sometimes.

@harisreedhar, thanks so much for sharing the models and code!

for sharing code and converted models.