masyagin1998 / robin Goto Github PK

.

Hello, thank you for the awesome project. I tried the base model on my own images and found that some image does not perform very well. I want to finetune the weights on my own dataset. May I know how to made the ground truth of the document images? Thanks!

I have
ICDAR 2017 Competition on Document Image Binarization
you can contact me if you still need this
my email: [email protected]

@masyagin1998 hi,

Robin seem to be very interesting and fast, I am currently working on a semantic segmentation code, and I want to achieve the highest prediction fps frames per second and accuracy along with memory efficiency, so what do you recommend?

• Should I use patches in training & prediction, or do you recommend something else?
• Does training fully rely on the cropped patches of images, or does it further have some tricks like (resizing, downscaling, upscaling, etc...)?

The pretrained weight works quit well for various document samples. However, when I train my own network from scratch, it seems that the performance is always worse than the provided weight. Therefore, can you please tell me the optimization method and training strategy you used?

@masyagin1998 Thank you for your amazing and well-written code, you are appreciated,

Have you considered the option of finetuning an existing model, or perhaps the ability to resume the training using an existing model?

Hi,masyagin,It is my glad to learn your method,when I run the train.py ,the error occured as follow:
File "train.py", line 391, in main
for i in range(len(os.listdir(os.path.join(args.input, 'in'))))]
FileNotFoundError: [Errno 2] No such file or directory: './input/in'
I thought I should create my dataset,but I don not know how to do and modify the path ,could you help me?Thanks very much.

Hey @masyagin1998,

Thank you so much for sharing the code. I appreciate it. I'm trying to create a python package for robin. I tried to add LFS support to upload the weight file(for running tests), but it seems we can't have LFS support to public forks if the original repo doesn't have it git-lfs/git-lfs#1906 . It would be great if you could add it.

Could you also add some kinda format for referring the code in the README.md?

Thanks and best regards

tensorflow is missing from requirements.txt. Can someone add it ?

Hi .. Thanks for your sharing
Would you please add a sample command to run train.py with its parameters?

Hi,

thanks a lot for your work.
would you please share the tensorlow and keras version used in this work?
it seems like that it cannot work with the latest version.

thank you