Poor optimization

[stalker3331111]

Hello, I have a question how can I optimize the program as much as possible, otherwise I have 15 pages of manga translated for almost an hour

[tak2hu]

Do you have a GPU? If not you could run this repo using google colab (havent tried it but its prob possible) for free gpu

[aqssxlzc]

How much could GPU improving the performance? In my computer it cost 30s per page, I am using AMD 3600 CPU, but my GPU is AMD that means I could not use CUDA.

[tak2hu]

How much could GPU improving the performance?

Because it accelerates the deep learning models of detecting/segmenting text, ocr, and inpainting (and translating if using sugoi translator).

See how fast this website is https://cotrans.touhou.ai/ as it uses the same deep learning models (this BallonTranslator project is based on that manga-image-translator project)

In my computer it cost 30s per page, I am using AMD 3600 CPU

Same, I’m using Intel Core i5 7300U, it is what it is

but my GPU is AMD that means I could not use CUDA.

Pretty sure there is a HIP option when selecting device. See https://pytorch.org/docs/stable/notes/hip.html but I don’t care about HIP because I don’t have the hardware

run this repo using google colab

Google colab is the only place I know that offer free gpu, just see these scripts to make a script using this project

So how this app works (simplified):

• it takes an image (jpeg, png, webp) as input
• input image gets passed to TextDetector function
• TextDetector outputs 2 things: image mask and TextBlock
• Image mask is an input for Inpainter function
• TextBlock is to store text boxes, bounding boxes, and a whole lot of plethora
• Inpainter function takes image mask and original image to output redrawn/cleaned image
• OCR function takes TextBlock and original image to output TextBlock with text that was scanned
• And when you edit text, it edits a TextBlock

Here is a picture to describe each model:

Imho the current models are not optimized for CPU inference (except LamaMPE for small masks, like using the healing brush), a more optimized or lightweight model should be implemented, a pull request is probably welcome by this project’s maintainer.

[aqssxlzc]

thanks for your explain. After a few experiment, I found HIP could not working in my Video Card RX6700XT, Because it need ROCm, which is maybe not support this device yet, and ROCm is not available on windows.

[qianzhou123]

谢谢你的解释。经过几次实验，我发现 HIP 在我的视频卡RX6700XT中无法工作，因为它需要 ROCm，它可能还不支持此设备，并且 ROCm 在 Windows 上不可用。 现在ROC上ROC可以在Windows↑