gaoyiminggithub / graphonomy Goto Github PK
View Code? Open in Web Editor NEWGraphonomy: Universal Human Parsing via Graph Transfer Learning
License: MIT License
Graphonomy: Universal Human Parsing via Graph Transfer Learning
License: MIT License
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?
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?
Can you share the Pascal pretrained model-----ascal_base_trained.pth?thanks!
@Gaoyiminggithub @Engineering-Course Hello sir, I would like to ask if this model of inference.pth can use only the CPU?
I find some cases show its worse than deeplab v3+.
the model is download from your release:CIHP pretrained model download link
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.
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?
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.
Hi, Gaoyiminggithub
I'm look for a good human semantic parser recently I found serval projects.
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.
Awesome job !does the inference model use transfer? I want to change the body parts into the same labe to extract person contours,will the training effect be better? The inferrence model didn't work well for my some small and low definition images,(I just change scale list into 2).
There is only one person in my picture ,should i train ATR or LIP use deeplabv3 directly ,can you give some advice or train step ? Thankyou!
Can't approch to DataSet and PreTrained model..
Please approve the access
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?
I need some information on training on my custom dataset with 2 labels - for human and background. Thanks
my system is using simple cpu please help needed
Normalize all class?I think this should be 1 instead of -1
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.
great work!!
does it can inference the different dataset labels, like cihp and pascal, at one model?
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?
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!!!
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:
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/
random crop
, scale
and left right flip
in their codeSo 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 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.
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
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!
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!
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?
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.
Hi, is there a training script to train pascal from chip dataset?
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?
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
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.
Many Thanks,
Yarin
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.
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.
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.
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.