search

menyifang / ADGAN

The Implementation of paper "Controllable Person Image Synthesis with Attribute-Decomposed GAN" CVPR 2020 (Oral); Pose and Appearance Attributes Transfer;
476 stars 90 forks source link

Reproduce using your pretrained model #4

Open noelcodes opened 4 years ago

noelcodes commented 4 years ago

I am trying to reproduce your work via pre-trained model (provided in your google link). Seem to be missing something. I ran this:

bash ./scripts/test.sh

but stuck here:

FileNotFoundError: [Errno 2] No such file or directory: 'your_path/deepfashion/semantic_merge3/000/01.jp/_/001..npy'

Can you provide the ***.npy files?

For detailed log error look here: https://colab.research.google.com/drive/19OPWmpnwgXdQuV06N3xRmew5wP6DuU4K#scrollTo=xl8EZycCiwPJ&uniqifier=1

Maybe I'm heading to the wrong direction. Please advise clearly how to reproduce your work without training?

menyifang commented 4 years ago

You should prepare the data step by step and modify the data path in test.sh (replace 'your_path' with your real path):

  1. download person images and split them into folders 'train' and 'test' following dataset split files 'tarin.lst' and 'test.lst'. (note: person image in each folder is named as 'fashionMENDenimid0000056501_1front.jpg')
  2. generate keypoints files and segmentation files. We have provided our segmentation files 'semantic_merge3.tar'(***.npy you mentioned) in google drive for convenience.
  3. modify the data path for your environment.
BadourAlBahar commented 4 years ago

How do we generate the segmentation files? Which human parser do you use?

menyifang commented 4 years ago

Hi @BadourAlBahar, we generated the segmentation files with pre-trained model provided in this project, and also tried more accuracy model, like Graphonomy.

BadourAlBahar commented 4 years ago

Thank you very much for the resources. In here, there are 20 classes for the human parts, how did you get the 8 segmented parts?

Edit: you mention in the supplementary material that you "integrate the original labels into 8 main categories".. How do you do this integration?

mkemka commented 4 years ago
  1. fashionMENDenimid0000056501_1front.jpg

Thank you so much for sharing your project - I am learning a lot about fashion. Can I ask why the files are joined like this instead of being in a folder structure? Is it a design choice? I have the source files but they are in folders.

menyifang commented 4 years ago

Edit: you mention in the supplementary material that you "integrate the original labels into 8 main categories".. How do you do this integration?

You can refer to this answer #10.

menyifang commented 4 years ago
  1. fashionMENDenimid0000056501_1front.jpg

Thank you so much for sharing your project - I am learning a lot about fashion. Can I ask why the files are joined like this instead of being in a folder structure? Is it a design choice? I have the source files but they are in folders.

This code is built on PATN, and we directly prepare the data like them.

shilongshen commented 4 years ago
  1. fashionMENDenimid0000056501_1front.jpg

Thank you so much for sharing your project - I am learning a lot about fashion. Can I ask why the files are joined like this instead of being in a folder structure? Is it a design choice? I have the source files but they are in folders.

This code is built on PATN, and we directly prepare the data like them.

Hello, can you provide the deep fasion dataset in-shop clothes retrival benchmark dataset that is not in the folder but used in the actual experiment?

menyifang commented 4 years ago

@shilongshen, you can download deep fashion dataset in-shop clothes retrival benchmark here

yyhnmn commented 4 years ago

Hi, @menyifang, I am trying to use your pretrained model to test the result. Which directory should I put the pretrained model file in?