About inference

[dongdongdong233]

Hi,
Thanks for the excellent code, when I inference the pictures, My program output is：

KeyError: 'relative_path'

Then I found "relative_path" is a key of "data_i", but when I print data_i, I found there doesn't have the key named "relative_path" in the "data_i". So what is "relative_path"? Please reply!

[liuqk3]

Hi @dongdongdong233 ,

Thanks for your interests.

Sorry for this mistake. The relative_path is just used to save the completed results. You can simply modify the config file like the following example:

[dongdongdong233]

Hi, thanks so much for your replying! After I add relative_path to the config file, this problem have been solved！ But there is another output like: AttributeError: 'PatchVQGAN' object has no attribute 'generate_content' I found there is no such a function named 'generate_content', the 'generate_content' is written in masked_image_inpainting_transformer_in_feature.py, so I copied the generate_content method to the end of the PatchVQGANclass in patch_vqgan.py, but the program still reported the same error. Could you please help me to solve this problem? Looking forward to your reply！

[dongdongdong233]

您好，最近读了您的论文Reduce Information Loss in Transformers for Pluralistic Image Inpainting，对您的论文十分有兴趣，在网上下载了您发布的代码。训练过程十分顺利，在推理单个修复结果（即运行python scripts/inference.py --name OUTPUT/transformer_naturalscene/checkpoint/last.pth --func inference_complet_sample_in_feature_for_evaluation --gpu 0 --batch_size 1  此行代码时），系统报了许多错，例如：AttributeError: 'PatchVQGAN' object has no attribute 'emb_proj'等 根据debug的过程来看，发现系统在调用patch_vqgan.py文件中的PatchVQGAN类时，用到了很多masked_image_inpainting_transformer_in_feature.py文件中写到的属性和方法，而PatchVQGAN类中并未定义这些方法和属性，那么程序可不可以直接调用masked_image_inpainting_transformer_in_feature.py文件里的方法呢。 由于程序的代码量对我来说很大，我做了许多尝试后发现自己并不能在短时间内读懂代码并解决这个问题，故而冒昧向您求助，热切盼望您能在百忙之中抽空回复我的邮件，解答疑惑。

祝您科研顺利，生活愉快！

------------------ 原始邮件 ------------------ 发件人: "liuqk3/PUT" @.>; 发送时间: 2022年10月11日(星期二) 下午4:55 @.>; @.**@.>; 主题: Re: [liuqk3/PUT] About inference (Issue #12)

Hi @dongdongdong233 ,

Thanks for your interests.

Sorry for this mistake. The relative_path is just used to save the completed results. You can simply modify the config file like the following example:

— Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you were mentioned.Message ID: @.***>

[liuqk3]

Hi @dongdongdong233 ，

P-VQVAE is just and auto-encoder, which can not inpaint the images but only reconstruct the images. If you want to inpaint an image, you need to train a UQ-Trainsformer further (or use our pretrained models). After that, you can use the command python scripts/inference.py --name path/to/transformer/last.pth --func inference_complet_sample_in_feature_for_evaluation --gpu 0 --batch_size 1

[zborger]

您好，感谢您优秀的工作。我下载您的预训练模型进行推理测试遇到一些问题想请教您。 在inference之前，我首先将自己准备的6张ffqh测试图和mask分别放入新建的文件夹下。然后修改OUTPUT/model_name/configs/config.yaml 文件中的data_root,provided_mask_name 和 name，用来指向自己存放的测试图路径。但是在运行inference后没有出现测试图的结果，并显示Missing keys in created model:[]Unexpected keys in state dict: []Found masks length 0, set to None等提示。我想程序并没有接收我指定的测试图，想请教您如果要测试少量自己准备的图片，我是还需要更改.yaml中的其他地方吗？期待您的解答。

[liuqk3]

From the logs you provided, it seems that the images are not located correctly, which means 0 images need to be inpainted. You can have a debug about the class ImageListDataset in PUT/image_synthesis/data/image_list_dataset.py to see how the images are located in these codes and provid correct paths to your images and masks.

Thanks.

[zborger]

From the logs you provided, it seems that the images are not located correctly, which means 0 images need to be inpainted. You can have a debug about the class ImageListDataset in PUT/image_synthesis/data/image_list_dataset.py to see how the images are located in these codes and provid correct paths to your images and masks.

Thanks.

Thanks for your prompt reply，I‘ll try it.

[zhangbaijin]

你好，在测试的过程中，我发现测试的mask似乎并不是我指定的mask，我想问一下应该如何设定呢，而且测试的mask覆盖率普遍在百分之30之上？

[liuqk3]

@zhangbaijin ，

You can have a look at this 12 #issue comment. Or you can modify 'inference.py' to load masks from another directory.

[Lecxxx]

@liuqk3 Hi, Thanks for the excellent work, when I run the inference.py, My program output is：

Traceback (most recent call last): File "scripts/inference.py", line 859, in launch(inference_func_map[args.func], args.ngpus_per_node, args.num_node, args.node_rank, args.dist_url, args=(args,)) File "/data/zixuan/PUT/image_synthesis/distributed/launch.py", line 52, in launch fn(local_rank, args) File "scripts/inference.py", line 204, in inference_reconstruction rec = model.decode(token['token'], combine_rec_and_gt=False, token_shape=token.get('token_shape', None)) File "/data/zixuan/anaconda3/envs/ImgSyn/lib/python3.7/site-packages/torch/autograd/grad_mode.py", line 27, in decorate_context return func(args, **kwargs) File "/data/zixuan/PUT/image_synthesis/modeling/codecs/image_codec/patch_vqgan.py", line 1501, in decode rec = self.decoder(quant, self.mask_im_tmp, mask=self.mask_tmp) File "/data/zixuan/anaconda3/envs/ImgSyn/lib/python3.7/site-packages/torch/nn/modules/module.py", line 948, in getattr type(self).name, name)) AttributeError: 'PatchVQGAN' object has no attribute 'mask_im_tmp'

Could you please help me to solve this problem? Looking forward to your reply！

[Hgy12345]

I also encountered this problem.

[liuqk3]

@Lecxxx ,

I think the error is thrown by the reconstruction function. Currently, the pretrained P-VQVAE is not seperately provided. You can try the inpainting function.

[liuqk3]

@Lecxxx @Hgy12345 While inpainting with inference.py or inference_inpainting.py, the argument --func inference_inpainting is required. Please read readme.md carefully.