microsoft / Recursive-Cascaded-Networks

[ICCV 2019] Recursive Cascaded Networks for Unsupervised Medical Image Registration
https://arxiv.org/abs/1907.12353
MIT License
361 stars 87 forks source link

my training result is not as good as paper #28

Closed w19787 closed 4 years ago

w19787 commented 4 years ago

thanks for excellent paper and work. i try to re-train the project after migrating the project to tensorflow 2.0. Only api is migrated, no network, loss, structure, data preprocessing ,etc changed.

The following is the screen shot of training tensorboard on liver dataset: image image

and the evalation result: image

any idea what might be go wrong? thanks.

zsyzzsoft commented 4 years ago

What setting are you running? I guess you are running the 1-cascade VTN version that should correspond to these numbers. Specify -n to train more cascades.

w19787 commented 4 years ago

What setting are you running? I guess you are running the 1-cascade VTN version that should correspond to these numbers. Specify -n to train more cascades.

thanks for reply. it is my bad not to notice the n should be changed for better performance according paper.

w19787 commented 4 years ago

What setting are you running? I guess you are running the 1-cascade VTN version that should correspond to these numbers. Specify -n to train more cascades.

in order to process the 10-cascade VTN training, what kinds of GPU is required? it is OOM on 16G v100 when train on 5 or 10 cascades.

zsyzzsoft commented 4 years ago

It would be fine if using 4 GPUs.

w19787 commented 4 years ago

It would be fine if using 4 GPUs.

after migrating to tf2.0 (since my server is installed new version's cuda which cannot run on tf1.4), the multi-gpu cannot work properly. the compute_gradients seems only work on gpu0. Currently, no idea how to fix it. image

image

w19787 commented 4 years ago

It would be fine if using 4 GPUs.

@zsyzzsoft i have upload https://github.com/w19787/Recursive-Cascaded-Networks-TF2.0, if you can advice how to fix the issue of multi-gpu support on this version will be appreciated!!!

zsyzzsoft commented 4 years ago

I haven't met this issue before and I'm not familiar with TF 2.0. Maybe the GPU specification doesn't work properly for TF 2, but I'm not sure.

w19787 commented 4 years ago

I haven't met this issue before and I'm not familiar with TF 2.0. Maybe the GPU specification doesn't work properly for TF 2, but I'm not sure.

got it. thanks!