Qing

> Could not find a version that satisfies the requirement lama-cleaner (from versions: ) No matching distribution found for lama-cleaner What is your version of python? python version should >=...

In some environments, the python binary directory may not exist in the shell's `PATH`. Check [Python Installing Packages Guides](https://packaging.python.org/en/latest/tutorials/installing-packages/#installing-to-the-user-site) to see how to add the binary directory to the shell's...

Implemented a simple version [0.23.0](https://github.com/Sanster/lama-cleaner/releases/tag/0.23.0): press alt/opt and move the mouse left/right to change brush size

1. If you are currently running on a cpu, you can use the gpu 2. Try to use `crop strategy`,

Another example: https://github.com/Sanster/lama-cleaner https://user-images.githubusercontent.com/3998421/192548839-9d0cbfc5-6bf7-4a5a-b556-446878696bdf.mp4

In test_loss `loss.backward()` is performed, this mean grad is calculated. Then in train_loss, grad calculated by test_loss is used by opt.step()

I implement Decoder part of MASTER-pytorch in FasterTransformer, compare to origin pytorch it's **much faster**. [Example code](https://github.com/Sanster/FasterTransformer/tree/v4.0-master-example/sample/pytorch/master_example)。It should be easy to add to TableMASTER with the minor changes

@RogerYu123 我用预训练的模型在 pubtabnet 的数据上跑，结果也会有偏移，你这边正常么？ 

> > @RogerYu123 我用预训练的模型在 pubtabnet 的数据上跑，结果也会有偏移，你这边正常么？  > > It is abnormal. maybe you can check your visualization code. You're right, my code is buggy。

Sorry, currently no multi-GPU support