Daniil Pakhomov
Daniil Pakhomov
@barryridge Great! Thank you so much! I am glad that you like it. I will merge once we switch to a new TF version soon.
Hi, I might have messed up the model -- I will have a look at it. Could you try the resnet 34 8s instead ? -- that should work Thank...
Seems like you are trying to init 34 model and load weights of 18 model. Try loading the weights of 34 model -- I took it from the table: https://www.dropbox.com/s/91wcu6bpqezu4br/resnet_34_8s_68.pth?dl=0
Make sure you changed your model into eval mode Also this is not the best model -- try psp
I will upload them a bit later For now you can try reducing the size of the input image. This will help since network was trained on images of 380...
Hi, it's MIT -- I will do so, thank you for pointing out :)
Hi @j-min I can add it later but as it was mentioned in the paper by Chen et al. they didn't notice an improvement when using networks with output stride...
Hi @ron-weiner You can try this software to create training data: https://github.com/wkentaro/labelme and then train using our library Let me know if you have any further questions :)
@ron-weiner I will have a look if such tool exists :) Or if you find something better by that time, I would also be interested in the way you approached...
@sharanry Hi, It should work with the latest available version, Could you, please, give me the line where exception occurred ?