csxmli2016
csxmli2016
> Thank you very much for your reply. I am new to the field of image editing with diffusion models and hope to learn from and improve upon your work....
> @csxmli2016 Thank you very much. ①Can you explain why the training time in the second stage is three times longer than in the first stage? Is it mainly due...
> @csxmli2016 Thank you for your continued responses. After reviewing the information and scripts you provided regarding checkpoints, I did not find where you explicitly save `global_step` in the checkpoint....
> hi!I am a new learner,how do I Emotion Editing.Thanks! You can change the 'att' to the emotion-related attributes (see this https://github.com/csxmli2016/w-plus-adapter/issues/2#issuecomment-1904082119)
> Hi @csxmli2016 > > Thank you for creating such a good project! > > I want to train this model for English data. I also have some custom LR...
> (marconet) C:\Users\L\Pictures\MARCONet>python test_sr.py -i "C:\Users\L\Downloads\bsrgan\inputs" --real_ocr ################################################################ Input Path : C:\Users\L\Downloads\bsrgan\inputs Save Path : C:\Users\L\Downloads\bsrgan\inputs_11-10_16-59_MARCONet The format of text label : using predicted text label OCR Module : using...
> 首先感谢作者杰出的工作,另外有一点点不明白的地方,希望作者解答一下。在checkpoints/download_github.py下载后,似乎可以得到5个模型权重,分别是net_new_bbox.pth,net_prior_generation.pth,net_real_world_ocr.pth,net_sr.pth和net_transformer_encoder.pth。然后在train.yml文件中,似乎只使用net_prior_generation.pth,net_transformer_encoder.pth和net_sr.pth。 > > 1. 关于train.yml文件中的net_d.pth和net_srd.pth来自哪里? > 2. 关于train.yml文件中,pretrain_network_ocr是否也可以使用预训练权重,模型来自哪里?是net_real_world_ocr.pth吗? > 3. download_github.py下载的net_real_world_ocr.pth和net_new_bbox.pth用在何处? 1. 关于训练代码,可以进入https://github.com/csxmli2016/MARCONet/tree/main/Train 里有详细的介绍,按照这里的```python scripts/download.py```可以下载所需的所有文件。你提的checkpoints/download_github.py是下载测试用的模型,不是训练的。 2. pretrain_network_ocr可以单独训练,不用跟sr模型一起,这样可以让网络更快的去微调先验,并嵌入SR过程中。 3. 用到测试时的这里了https://github.com/csxmli2016/MARCONet/blob/58582fe5801b4ff3b5cdaf6e96c8feb3a426b68a/test_sr.py#L68 和 https://github.com/csxmli2016/MARCONet/blob/58582fe5801b4ff3b5cdaf6e96c8feb3a426b68a/test_sr.py#L61
> 你好,再请问一下,TrainData/BGSample/DF2K_Patch给了数张裁剪图像的例子,请问下背景裁剪后图像的数量对最终模型效果的影响大吗?有这个疑问,是因为我想train一下,是否需要下载DF2K或者DIV2K数据集裁出更多的背景。 影响不大,但是也不要让模型过拟合到这种背景纹理了。
> Restoring syn_lq03.png. The predicted text: 的地址就是不知道要把它添加在 Traceback (most recent call last): File "test_sr.py", line 266, in main prior_cha, prior_fea64, prior_fea32 = modelTSPGAN(styles=w0.repeat(labels.size(0), 1), labels=labels, noise=None) File "/home/anmy/miniconda3/envs/marconet/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1553,...
> _No description provided._ https://github.com/csxmli2016/MARCONet/tree/main/Train#a-simple-demo-for-synthesizing-the-training-data 看这里合成的数据。对于识别和字符位置检测,可以离线合成一些进行训练。对于SR过程,可以在线合成数据