Qing
Your logs look strange, what command did you execute? `pip install lama-cleaner` should work, tested on following environment macOS 13.0 (Intel) - python 3.8.11 + pip 21.1.1 - python 3.7.8...
> Add more functions on mask You can turn on `Manual Inpainting Mode`, it has the ability to rollback mask changes when pressing `ctrl+z` When Manual Inpainting Mode is on,...
~/.cache/huggingface
You can use `lama` model to remove the watermark, this model does not need a prompt. `Empty Response` may be caused by other errors, please check the message in the...
> I tried, but there are some operators missing for the conversion. It will not be possible without the help of the official team of CoreMLTools which model have you...
@nttstar 你的炼出来的精度很高啊 :+1: 。请教个问题，我看 SimpleNet 中有这样两行代码，这是出于什么考虑呢？ ```python3 max = mx.sym.Pooling(data=net, name='pool-0_m', pool_type='max', kernel=(2, 2), stride=(2, 2)) avg = mx.sym.Pooling(data=net, name='pool-0_a', pool_type='avg', kernel=(2, 2), stride=(2, 2)) net = max - avg...
@nttstar 多谢。 我也试着用 SimpleNet 和原版的 crnn 结构在 caffe_ocr 的数据集上炼了下，练到 98% 左右就停了，然后又用 [text_renderer](https://github.com/Sanster/text_renderer) 生成了 1w 的数据测了下，精度只有 40% 多。如果用 [text_renderer](https://github.com/Sanster/text_renderer) 生成的数据(500w) 来练（验证集到97%），在 caffe_ocr 的测试集上也只有 67%，感觉网络范化性能不太行啊... :disappointed:
@piaobuliao 你 `@` 错人... 确实，两个数据集生成的规律不一样，字符频率也不一样。
@xiaoqiao728295 hi，你具体实现的是什么网络机构？可以参考下我这个项目 [tf_crnn](https://github.com/Sanster/tf_crnn) 目前有 resnet, squeezenet, mobienet 等结构
有需要的可以参考下: https://github.com/Sanster/text_renderer