Hi,
I try to reproduce your deeplabv3+ baseline on CIHP data, But I don't know your paper result about CIHP dataset is based on Graphonomy code or Google Deeplabv3 code.
I use the google official code to train model on CIHP data, but only get MIOU about 50%, much lower than your paper result 57.13.

My implement detail is following:

1. Use the coco pretrain model provide by deeplabv3
2. Then I train the model similar to your paper setting

python deeplab/train.py \
    --training_number_of_steps=180000 \
    --train_split="train" \
    --model_variant="xception_65" \
    --atrous_rates=6 \
    --atrous_rates=12 \
    --atrous_rates=18 \
    --output_stride=16 \
    --decoder_output_stride=4 \
    --train_crop_size="513,513" \
    --train_batch_size=16 \
    --base_learning_rate=0.007 \
    --fine_tune_batch_norm=True \
    --initialize_last_layer=False \
    --dataset="cihp" \
    --tf_initial_checkpoint=deeplab/pretrain/xception_65_coco_pretrained/x65-b2u1s2p-d48-2-3x256-sc-cr300k_init.ckpt \
    --train_logdir=datasets/cihp/train/xception_65_coco_pretrained_cihp_batch16 \
    --dataset_dir=deeplab/datasets/cihp/tfrecord/

3. The google deeplab already include the random crop, scale and left right flip in their code
4. But finally I only get MIOU 50%

So is there any big issue of my training precess.
I find my model can not distinguish left and right hand and leg, is there any problem about flip?

Result from google code

Result use your code (using data/pretrained_model/inference.pth model)

My model result is not very smooth and can not class left and right.
I'm new to semantic segment, Could you please give me some advice on this issue.
Thanks
Gordon.

Dear author,
Thanks for your neat implementation.
I found the report deeplabv3 performance in this paper is relatively close to Graphonomy. Therefore I want to reproduce the deeplabv3 related experirments. Could you kindly give some tips on how to train and evaluate on deeplabv3 baselines using your repo.? Thank you.

I need some information on training on my custom dataset with 2 labels - for human and background. Thanks

Hi, is there a training script to train pascal from chip dataset?

Hi, thanks for your work. Would you please update the model weight of the universal model? It seems that the google link is invalid and the file was removed. Thanks!

In your paper, I don't know how to get the value of v for a node.
Can you explain it in detail for me? Really grateful!

Hello
Thanks for the awesome work

In the README, there are link to datasets that can be downloaded via google drive, what are the licenses for using these datasets ?
b.r

I tried to download the universal model and got the following message. Could you please double-check the link?

i'm trying to prepare the data to run this but there are no links to the pascal-part dataset. i found this http://roozbehm.info/pascal-parts/pascal-parts.html but after downloading and unzipping the folder only contains 15 photos. where do i get the full dataset?

Thanks for sharing such excellent work! I find the code here looks wired because the source_graph_con3 is never used in the function. I guess that it is a mistake maybe?

How can we get the instances?
If there are two humans in the image, can we get a list of the areas that belong to each of them?

Can this model be mapped to a single node graph for human parsing ( Semantic Segmentation - 2 classes - human / Background), if yes, how should I calculate the adj matrix for the same. Any help or guidance on the same would be appreciated.

Hi,
Thanks for your contribution!
I know that images in "Category_rev_ids" directory is obtained by flipping. But I also have to exchange the label of the body parts such as 'left-arm' and 'right-arm'? Am I right?

I have found three sets of weights from these issues:
#29 (comment)
#16 (comment)
And these are the outputs of using them to segment messi.jpg



But none of them is the same as the segmented image in the repo.
Do you still have a copy of that instance that generated the repo segmentation?

How to train this model to segment different part(48 classes) of the cars. I have data set and their mask. How to train this model on my custom dataset.

Hi, Gaoyiminggithub

I'm look for a good human semantic parser recently I found serval projects.

• one is your one,
• another is SSL

Because the LIP_SSL didn't annotated the neck part of human, but your code can handle this.
But why the speed of Graphonomy is much more slower than LIP_SSL, is there anyway I can speed up.

Normalize all class?I think this should be 1 instead of -1

hi, great work！
currently I try to use your inference model to predict my data. But when I use universal_trained.pth as the pretrained model with gpu memory 12GB, it always report error meaning not enough cuda memory. so I wonder how much memory does the model need? any ideas to optimize this? should I use small resized pictures or any other things I could do?
waiting for your reply, thanks!

Thanks for sharing the repo.

I notice that another repo mentions that there exist no ImageNet pre-trained Aligned Xception model, and the performance is worse than the ResNet-101.

In other words, I mean that the model below is not the Aligned Xception trained on ImageNet~

pretrain_dict = model_zoo.load_url('http://data.lip6.fr/cadene/pretrainedmodels/xception-b5690688.pth')

I am wondering whether you have trained the Aligned Xception model on the ImageNet. It would be great if you could share the models~

Besides, I am also wondering if the Imagnet pre-trained Aligned Xception model performs better than modified ResNet-101.

Hi ,
Please help me to resolve the issue moment i run the program i get this error
new_node1 = torch.matmul(res_feature_after_view1, self.node_fea_for_res)
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 1.41 GiB (GPU 0; 2.00 GiB total capacity; 330.44 MiB already allocated; 1.23 GiB free; 410.00 MiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

approach tried

1. import os
   os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:2000" -> did'nt work
2. os.environ["PYTORCH_CUDA_ALLOC_CONF"] = 'max_split_size_mb:2000' -> did'nt work
3. torch.cuda.empty_cache() -> in the beginning of the file and several places within the loop
   for iii, sample_batched in enumerate(zip(testloader_list, testloader_flip_list)):

I am using the model for segmentation of human images. I am following the instructions
python exp/inference/inference.py --loadmodel data\pretrained_model\inference.pth --img_path ./img/00891_00_resize.jpg --output_path ./img/ --output_name 00891_00_parse.

The output is correct for images with resolution 192X256. However the output is not correct for images with resolution 768X1024. pls help me out.

Dear author:
thanks for share the implementation code!
When i dive into your code, a trivial detail confused me. as stated below:

        net.train()
        for ii, sample_batched in enumerate(trainloader):

            inputs, labels = sample_batched['image'], sample_batched['label']
            # Forward-Backward of the mini-batch
            inputs, labels = Variable(inputs, requires_grad=True), Variable(labels)
            global_step += inputs.data.shape[0]

I dnk why the inputs requires gradient? Please correct me! thank you.

In the ReadMe file, for dataset preparation, it is said that You need to download the human parsing dataset, prepare the images and store in /data/datasets/dataset_name/. We recommend to symlink the path to the dataets to /data/dataset/ as follows.

What are the steps to prepare this Pascal Dataset?

Hi, I have no access to google drive for the trained model, could you please give me a hand?

Hi ,
Thank you for sharing your great work!

I saw that due to some bug (reset parameters) you retrained the CHIP and Pascal models and uploaded them to Google Drive.

1. Unfortunately the link you updated is private so I get "Access Denied"
2. Is the Universal model also influenced by this bug? if yes, will it be updated?
3. is it possible to extract both Pascal and CHIP in one inference? is it recommended accuracy wise?

Many Thanks,
Yarin

Thanks for sharing this repo.
I have downloaded the checkpoint 'inference.pth' in the Model section in README.
And when i load the model weight i find that the 'target_featuremap_2_graph.pre_fea' and 'target_featuremap_2_graph.weight', which are contained in the feature to graph projection module, are all 0, which i think is not so reasonable.
Am i having any misunderstanding or maybe I have downloaded a wrong model?

Can't approch to DataSet and PreTrained model..

Please approve the access

Hi,

Thanks for such a great work and neat code!

Currently I am trying to reproduce the result on the augmented ATR dataset, but unfortunately I didn't find the train/validation/test split here, here and here.

Could you please kindly provide the list for the ATR dataset?

Many thanks,
Charles

I find some cases show its worse than deeplab v3+.

BadCase

raw image

graphonomy

the model is download from your release:CIHP pretrained model download link

deeplab v3+

Potential reasons

I guess maybe i have downloaded a wrong model or something wrong, but i cannot figure it out. Thank you for your share and read this issue.

great work!!
does it can inference the different dataset labels, like cihp and pascal, at one model?

@Gaoyiminggithub @Engineering-Course Hello sir, I would like to ask if this model of inference.pth can use only the CPU？

I can not find ATR and CIHP test datasets. I want to know how you tested on these two datasets.
Besides，How to Train Deeplabv3+ on human parsing datasets. Thank you!!!

Can you share the Pascal pretrained model-----ascal_base_trained.pth？thanks！