OSError: SavedModel file does not exist about models HOT 5 CLOSED

There should be a saved_model.pbtxt in fsl_fast.tar.gz\fsl_fast\0\1513180310, maybe try pointing to this? As far as I remember, it was sufficient with tf v.1.10 to point to the directory as you did above. I assume test.csv has been replaced by the CSV you posted above.

Another note, there might be an issue running these models on other data. Typically, MR weighted images exhibit quite a large variation between acquisition protocols.

from models.

Thanks, I'll try with fsl_fast/0/1513180310/saved_model.pb and let you know if it worked.

Maybe I can try histogram matching to the training dataset. Also, should my image be MNI-registered? Would a rigid registration be enough?

from models.

Histogram matching might improve things, yes. Registration is a standard rigid to MNI 1mm isotropic resampling as done in the UK Biobank (c.f. the paper).

from models.

Great, thanks!

from models.

I modified the model path in config.json to:

"model_path": "/tmp/neuronet/models/fsl_fast/0/",

and it worked.

I ran it on three MNI-registered images from three different scanners and the segmentations are reasonable. @mrajchl, thanks for your help.

from models.