Nvidia DIGITS 6.0 unable to parse half precision param

We have a server running DIGITS v6.0 with the following command: docker run --runtime=nvidia --restart unless-stopped --name digits -d -p 5000:5000 -p 6006:6006 -v /media/STORAGE/digits-jobs:/jobs -v /media/STORAGE2/:/media/STORAGE2 -v /media/STORAGE/MEDIA_DB:/media/STORAGE/MEDIA_DB nvidia/digits

Trying to train a network using the fp16 capabilities:

name: "net"

default_forward_type: FLOAT16
default_backward_type: FLOAT16
default_forward_math: FLOAT16
default_backward_math: FLOAT16

...

We encounter the following error: Bad network: Not a valid NetParameter: 2:1 : Message type "caffe.NetParameter" has no field named "default_forward_type"..

How do we fix this?

I don’t think its fixable. I don’t think digits was ever updated to include FP16. See here

I don’t think its fixable. I don’t think digits was ever updated to include FP16. See here

Can I update the version of caffe that DIGITS is using by modifying the official digits docker? is there a workaround?