search

taoxugit / AttnGAN

MIT License
1.33k stars 415 forks source link

Would you like to make the implementation of R-precision available? #56

Open ShihuaHuang95 opened 5 years ago

ShihuaHuang95 commented 5 years ago

Hi, I am going to compare AttnGAN with our method, I am wondering if you can release the code for the R-precision proposed in your AttnGAN paper and make the comparison more fairly.

gryBox commented 5 years ago

What method are you using? Is it published?

ShihuaHuang95 commented 5 years ago

No, still in progress.

gryBox commented 5 years ago

Quick question: From your understanding of this model. Does the quality of the images generated depend on the hyperparameter CAPTIONS_PER_IMAGE. For example does training a dataset with 5 captions per image vs 1 (long) caption per image make a difference?

FWIW: I've been working with @gar1t on using guild with this model. Makes it super easy to do hyperparamater optimization and keep track of tests.

ShihuaHuang95 commented 5 years ago

Emmm...., I have not taken such issue into account. But I think they would make a difference, as more captions will provide more meaningful samples for the training of the text encoder.

gryBox commented 5 years ago

Thank you for your answer. It was what I was looking for. For context -- This is what I am currently working on. A data pipeline to automatically prepare training data for a text to image model.

Pycon 2019 Poster presentation (It's a 4ft x 8ft poster) Not to be viewed on a small screen.

PyCon2019-Image-DataPieline.pdf

ShihuaHuang95 commented 5 years ago

Cool! Excellent work!

gryBox commented 5 years ago

Thank you. Again from your understanding of this model. Do the captions have to be fixed? For example: Can one image have 5 captions and another 20?

The whole point of this is minimize the amount of data the machine needs as inputs and do some automatic labeling.

Thanks again for taking the time to answer these questions.

venuszhou commented 5 years ago

@Hshihua so have you implemented the R-precision?

ShihuaHuang95 commented 5 years ago

@venuszhou Not yet, but I am trying my best to re-implement.

ShihuaHuang95 commented 5 years ago

@venuszhou @gryBox Hi guys, I have opened a new repo for such evaluation metrics, and I am going to finish this work in the next month.

gryBox commented 5 years ago

@Hshihua - Nice. I am currently working on automating the process of running all the steps once data is downloaded. Early draft: https://github.com/gryBox/pytig/tree/photo. At pycon there was a strong interest by a number of senior people to contribute to a python library for text to image generation. The idea is take a few state of the art text to image models, maintain and optimize them for users looking to experiment with models in a couple commands

huqiyii commented 4 years ago

I want to know how to evaluate whether a picture is related to a sentence

Iceland-Leo commented 4 years ago

@Hshihua Excuse me, do you implement the R-precision now? It's important. I think I need your help

FlowerRoad1996 commented 4 years ago

@Hshihua Excuse me, do you implement the R-precision now? It's important. I think I need your help

Excuse me，It’s also very important for me, have you solve this problem and willing to share it? Thank you very much!

maincarry commented 4 years ago

I wrote an R-Precision module for AttnGAN. It measures the R-Precision of evaluated images from a certain checkpoint. The result may be off a bit from the paper because how the image is generated but should be consistent. Hope this can help. @Iceland-Leo @FlowerRoad1996

FlowerRoad1996 commented 4 years ago

I wrote an R-Precision module for AttnGAN. It measures the R-Precision of evaluated images from a certain checkpoint. The result may be off a bit from the paper because how the image is generated but should be consistent. Hope this can help. @Iceland-Leo @FlowerRoad1996

Oh，thank you very much! I'll try to use it, and I also want to evaluate it by myself and compare the result.

Iceland-Leo commented 4 years ago

I wrote an R-Precision module for AttnGAN. It measures the R-Precision of evaluated images from a certain checkpoint. The result may be off a bit from the paper because how the image is generated but should be consistent. Hope this can help. @Iceland-Leo @FlowerRoad1996

Thank you very much! I'll use it.

cripac-sjx commented 3 years ago

I wrote an R-Precision module for AttnGAN. It measures the R-Precision of evaluated images from a certain checkpoint. The result may be off a bit from the paper because how the image is generated but should be consistent. Hope this can help. @Iceland-Leo @FlowerRoad1996

Would you like to share the r-precision module, thanks very much

cloverzyy commented 3 years ago

I wrote an R-Precision module for AttnGAN. It measures the R-Precision of evaluated images from a certain checkpoint. The result may be off a bit from the paper because how the image is generated but should be consistent. Hope this can help. @Iceland-Leo @FlowerRoad1996

Could you please share the r-precision module with me? Thanks! My email address is 1094680771@qq.com.