search

NVIDIA / DIGITS

Deep Learning GPU Training System
https://developer.nvidia.com/digits
BSD 3-Clause "New" or "Revised" License
4.12k stars 1.38k forks source link

GPU selection is not offered in DetectNet Model Build. #2001

Open panzerfrank opened 6 years ago

panzerfrank commented 6 years ago

I am working through the Two Days to a Demo Tutorial from NVIDIA following download of the new release of Jetpack 3.2

Caffe and DIGITS run to make the datasets mnist, coco-dog and ImageNet-ILSVRC12-subset

Models GoogleNet-ILSVRC12-subset and MNIST DIGITS run happily

However when I try to build the DetectNet-COCO-Dog I am not offered the select GPU option as shown on the tutorial and the build crashes out on lack of memory.

I am using Server: UBUNTU-DT DIGITS version: 6.1.1 Caffe version: 0.15.14 Caffe flavor: NVIDIA

Has anyone else seen this fault?

adithya-p commented 6 years ago

Can you post a screenshot of the log after running make command for building Caffee?

panzerfrank commented 6 years ago

This is the last screenshot of the Caffe Log after running the program

screenshot from 2018-03-22 17-01-40

is this the information you are looking for. Protobuf seems to be thewrong version

panzerfrank commented 6 years ago

make says it is up to date in caffe

panzerfrank commented 6 years ago

Make for caffe says it is up to date

sindhurau2009 commented 6 years ago

If you have 2 or more GPU's available on your system, the DIGITS server would ask you to select a GPU of your choice. If there is just one available and the server has been set up using GPU and CUDA correctly, then the model runs on the GPU by default.

What is the RAM for your GPU? Also, can you specify the values of batch size and batch allocation specified while building the model? It might be that the batch size is too high for your GPU's memory, in which case you might want to bring it down.