switchablenorms / deepfashion_try_on Goto Github PK

Can you please provide testing file/ inference file for us.
Seems like test.py is basically a train.py
When i try to run test.py, training gets started.

When tried to use custom images with the code, it gives list out of index error.

Very nice work! Thanks for sharing. Could you please explain how you pre-process data (e.g. how you derived the data_colormask, data_mask, data_edge, data_pose etc.)? Many thanks!

Could you please share your training scripts? There are so many options in train_options.py and base_options.py, and I am not sure if the default settings will give the best results. Thank you very much.

hey，does this problem with heatmap will influence the model result?

How to generate a result image after changing clothes based on a portrait picture and a clothes picture.Thanks!

I thought that using your pretrained models mean I can already utilize it for inference. However, i checked clothes warping module alone and it produces results with many artefacts. Is it normal?

Your work is really amazing and interesting, so I wonder if you can publish the training code or share it with me([email protected]). Thank you~

hi，thank you for your work：
why the result data need add one and divide two

cv_img=(combine.permute(1,2,0).detach().cpu().numpy()+1)/2

Hi there!

Thanks for providing this awesome Repository and make it openSource.I am facing some problems while running TRAIN.PY

Here is the error I am getting :
AssertionError: /dockerdata/benchmark_datasets/try_on_training/train_label is not a valid directory

when I run:
$ python test.py --dataroot /Data_preprocessing --phase test
I got an error:
$ AssertionError: /Data_preprocessing/test_label is not a valid directory
can some body tell me why? I just want to use the test.py to test your test data. And I download your test dataset and put it in the folder named Data_preprocessing. why occur such a strange error???

Is there an option to only use a percentage of training data available while training? Thanks.

I am trying the regenerate the test result. I have tested that my pose model is working fine.
But I have tested human parsing with both CIHP_PGN model and also with another model trained on LIP dataset.
Below is the result I am getting and I think the edges in the segmentation is producing the noise in the final output.
Please help me on this.

the order of img above is real_image, clothes, pre_clothes_mask, mask_clothes, all_clothes_label, target_label_in_CELoss, inference_result_of_G1, I wonder why the inference could learn the shape of try-on clothes and show it on the label while the target label keep the old clothes label?

Hi~ I'm trying to reproduce your work on our dataset, so I'd appreciate it if you could tell me the quantity and format of your training dataset, Thank you~

Below are the testing results on my data.

1. I have used pose_keypoints from coco_18 model and verified the keypoints by drawing on the image and are good for the below results.
2. Generated labels with pretrained model o LIP dataset and modified the numbering accordingly.
3. I am using test_color and test_edges same as test data provided.

The results are not completely satisfying. Any suggestions are welcome.

Thank you

Little v shape in neck is same as query image

Sleeves are same as query image

Overlaying the back layer of t-shirt on front neck

Left arms cloth is not overlaying

Cloth is same as query image

Thanks for providing open source code, can you provide test.py ?

Can you please help with this error?

chethan@ex5820:~/DeepFashion_Try_On/ACGPN_inference$ python3 test.py
?
------------ Options -------------
batchSize: 1
beta1: 0.5
checkpoints_dir: ./checkpoints
continue_train: False
data_type: 32
dataroot: ../Data_preprocessing/
debug: False
display_freq: 100
display_winsize: 512
fineSize: 512
gpu_ids: [0]
input_nc: 3
isTrain: True
label_nc: 20
lambda_feat: 10.0
loadSize: 512
load_pretrain: ./checkpoints/label2city
lr: 0.0002
max_dataset_size: inf
model: pix2pixHD
nThreads: 2
n_blocks_global: 4
n_blocks_local: 3
n_downsample_global: 4
n_layers_D: 3
n_local_enhancers: 1
name: label2city
ndf: 64
netG: global
ngf: 64
niter: 100
niter_decay: 100
niter_fix_global: 0
no_flip: False
no_ganFeat_loss: False
no_html: False
no_lsgan: False
no_vgg_loss: False
norm: instance
num_D: 2
output_nc: 3
phase: test
pool_size: 0
print_freq: 100
resize_or_crop: scale_width
save_epoch_freq: 10
save_latest_freq: 1000
serial_batches: False
tf_log: False
use_dropout: False
verbose: False
which_epoch: latest
-------------- End ----------------
CustomDatasetDataLoader
dataset [AlignedDataset] was created
../Data_preprocessing/test_label label
../Data_preprocessing/test_label label
../Data_preprocessing/test_img img
../Data_preprocessing/test_img img
../Data_preprocessing/test_edge edge
../Data_preprocessing/test_edge edge
../Data_preprocessing/test_mask mask
../Data_preprocessing/test_mask mask
../Data_preprocessing/test_colormask colormask
../Data_preprocessing/test_colormask colormask
../Data_preprocessing/test_color color
../Data_preprocessing/test_color color

Inference images = 10
latest_net_U.pth
latest_net_G1.pth
latest_net_G2.pth
latest_net_G.pth
/home/chethan/.local/lib/python3.6/site-packages/torchvision/transforms/transforms.py:188: UserWarning: The use of the transforms.Scale transform is deprecated, please use transforms.Resize instead.
"please use transforms.Resize instead.")
/home/chethan/.local/lib/python3.6/site-packages/torchvision/transforms/transforms.py:188: UserWarning: The use of the transforms.Scale transform is deprecated, please use transforms.Resize instead.
"please use transforms.Resize instead.")
Traceback (most recent call last):
File "test.py", line 104, in
for i, data in enumerate(dataset, start=epoch_iter):
File "/home/chethan/.local/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 336, in next
return self._process_next_batch(batch)
File "/home/chethan/.local/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 357, in _process_next_batch
raise batch.exc_type(batch.exc_msg)
IndexError: Traceback (most recent call last):
File "/home/chethan/.local/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 106, in _worker_loop
samples = collate_fn([dataset[i] for i in batch_indices])
File "/home/chethan/.local/lib/python3.6/site-packages/torch/utils/data/dataloader.py", line 106, in
samples = collate_fn([dataset[i] for i in batch_indices])
File "/home/chethan/DeepFashion_Try_On/ACGPN_inference/data/aligned_dataset.py", line 160, in getitem
C_path = self.C_paths[test]
IndexError: list index out of range

thanks for your work. I still have a quesiton that your person-keypoints has length of 54，not equal to coco format whose length is 51，can you tell me your difference between this gap？

I have searched the entire repo but still couldnt find the solution

Traceback (most recent call last):
File "train.py", line 192, in
model.module.optimizer_G.zero_grad()
File "/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py", line 539, in getattr
type(self).name, name))
AttributeError: 'Pix2PixHDModel' object has no attribute 'optimizer_G'

Thanks for this great work, however, I still have some questions about this work.

1. How should I evaluate SSIM? I directly calculate SSIM score on image reconstruction task (reference and target are from the same image). However, the pretrained model and my own trained model (20 epoch) get 0.7980 and 0.7594 on test set.

2. Does this model only need 20 epochs training? In default options, this model would be trained for 200 epochs. I found that SSIM score is still increasing after 20 epochs and reached 0.7653 in epoch 40.

3. In released training and inference code, average skin color (skin_color) of each class area is used in the input of content fusion module instead of synthesized clothing mask (M_c^S) mentioned in paper. 🤔
   

   DeepFashion_Try_On/ACGPN_train/models/pix2pixHD_model.py

   Line 323 in a628ca5

   G_in=torch.cat([img_hole_hand,masked_label,real_image*clothes_mask,skin_color,self.gen_noise(shape)],1)

When running the train only see the pytorch warning log about the input image size? I also use VITON Dataset.
Is this the problem that causes the output quality error as shown below:

/usr/local/lib/python3.6/dist-packages/torchvision/transforms/transforms.py:211: UserWarning: The use of the transforms.Scale transform is deprecated, please use transforms.Resize instead.
"please use transforms.Resize instead.")
/usr/local/lib/python3.6/dist-packages/torchvision/transforms/transforms.py:211: UserWarning: The use of the transforms.Scale transform is deprecated, please use transforms.Resize instead.
"please use transforms.Resize instead.")
/usr/local/lib/python3.6/dist-packages/torch/nn/_reduction.py:43: UserWarning: size_average and reduce args will be deprecated, please use reduction='mean' instead.
warnings.warn(warning.format(ret))
/usr/local/lib/python3.6/dist-packages/torch/nn/functional.py:1558: UserWarning: nn.functional.tanh is deprecated. Use torch.tanh instead.
warnings.warn("nn.functional.tanh is deprecated. Use torch.tanh instead.")
/usr/local/lib/python3.6/dist-packages/torch/nn/functional.py:3226: UserWarning: Default grid_sample and affine_grid behavior has changed to align_corners=False since 1.3.0. Please specify align_corners=True if the old behavior is desired. See the documentation of grid_sample for details.
warnings.warn("Default grid_sample and affine_grid behavior has changed "
/pytorch/torch/csrc/utils/python_arg_parser.cpp:756: UserWarning: This overload of nonzero is deprecated:
nonzero(Tensor input, *, Tensor out)
Consider using one of the following signatures instead:
nonzero(Tensor input, *, bool as_tuple)
/usr/local/lib/python3.6/dist-packages/torch/nn/modules/loss.py:88: UserWarning: Using a target size (torch.Size([2, 1, 256, 192])) that is different to the input size (torch.Size([2, 256, 192])). This will likely lead to incorrect results due to broadcasting. Please ensure they have the same size.
return F.l1_loss(input, target, reduction=self.reduction)

Runtime environment is:

Cuda compilation tools, release 10.1, V10.1.243
torch.version: 1.5.1+cu101

I realize that data needs to be preprocessed and put in Data_preprocessing folder, but where is the data from? It would be greatly appreciated if someone write a brief instruction on how to train the model on the given data or custom data. Thanks.

There is a inference model , but I don't the input:label, label_ref, image_ref；Can you tell me what they mean? Because inference should only need to input model pictures and clothes to successfully predict, but there are 11 parameters in train model .

Thank you for your work, but there is no information about how to test with custom data.

When I trained the network with LIP dataset (20 labels' segmentation input), this error raised

Based on NVlabs/SPADE#57 it should be some channel problem, but I believe I had changed relevant channels and total channel numbers except noise channel. Since Lip dataset doesn't have a counterpart, I used the dress channel and it seems not work properly. Any idea to fix this issue?

Hello, thanks for the repo and the datasets, I tried to show the key_points on the img with same preifx but found that the points are disorderly. Could you tell me the correct order for the annotation key_point in train_pose repos？

Hello, thank you very much for your great work! The pipeline is deeply thought and designed smartly. It would be really great to know the full label mapping between original VITON label maps and your generated new maps. Also, segmentation map you provided in the README has some number gaps. Can you please tell me what's the reason? Thank you.

Hello, in the dataset, what does the train colormask and train_mask mean? How to make this? I want to generate it use my own dataset, Could you please guide me?

How could I get the vgg19-dcbb9e9d.pth?
Thanks

Hi, I intend to add a new image to the dataset and train it again, but after 20 epochs, the test image has a strange color in the neck:


I use a Google colab for training with Pytorch 1.15.
Do you know this error?
And can you give more information about the python environment which runs out the same as the result in the article?

In the file DeepFashion_Try_On/ACGPN_train/models/pix2pixHD_model.py, in the 323 line, why put real_image in the G_in, is it an error? real_image is label, why put label in the network input?

Hey can you help me with this? I have used parsing model and openpose what editing should i do ? I am using the normal test.py file given by you. I have added 14 images in test_img.... Same for test_pose and test_label. Added test_colormask and test_mask as it is. And 1 image in test_color and test_edge.

I am a student and now I am doing some research in computer vision. Your paper and code are of great help! However, I do not understand where the data with poses in json came from. What software did you utilize? I want to use it in my project

According to the paper, one of the inputs to the second generator is supposed to be the output from the first generator, which is dis_label in the code. But in the training coding feeding the dis_label information to the generator, masked_label is fed to the generator which is from the train data segmentation, not the generated one. But in for the inference code, the generated variable is fed to the second generator.
The same thing is found for the inputs to the third generator.
Could you please explain these inconsistencies in training and inference code?

I use LIP model to generate my own dataset, but the label is different with yours , can you tell me which segmentation model you use to generate your dataset?

I have a segmentation model and a keypoints model trained with 17 keypoints.
Your model seems to need 18 keypoints with_center.
Can I use your model with 17 keypoints?

thanks for your contribution ！ but i still have a question about how to generate your dataset . for example ,your train_lable and test_lable?

Hi! Thanks for your great work.
I am confused about the effect of g1.
Why need to train a g1 to output m_w^S?
Can m_w^S be obtained directly by a human parser?

thanks for your kindly reply，but I still have a question that test_edge only consists 0 and 255 in pixel？how to get it？

Thank you very much for contributing to this amazing project.

I am trying to test the model with custom images and I have:

• test_color
• test_edge
• test_img
• test_label
• test_pose

How can you generate the files from the folders ?:

• test_mask
• tes_colormask

Thanks

In vtion and cp_vton, they both use LIP_JPPNet to get person parsing label, which has 20 class labels. But in your code, your person pasing label only has 14 class, what does 14 class labels mean?
If you map the 20 class to 14 class, could you please provide map dict? thank you.

Hi，Where is the generated warped clothes？