Does ocrd_calamari use GPU using ocrd/all:maximum-cuda image?

[mikegerber]

Reported by @jbarth-ubhd in Gitter (Oct 06 10:50)

calamari-recognize does not run on GPU, even with maximum-cuda

This seems to be concluded from this memory usage graph:

gpu memory usage, red=sbb-binarize, blue=eynollah-segment, green=calamari-recognize

[mikegerber]

I suspect that it's actually running on GPU but does not use much memory compared to eynollah or sbb-binarize, but I'll investigate anyway to confirm.

[maxnth]

I can't comment on whether ocrd_calamari actually leverages the GPU or not but AFAIK calamari should seldomly put heavy load on the GPU during recognition anyways.

[mikegerber]

Confirmed: It's not using GPU using the ocrd/all:maximum-cuda image. I'm investigating further. @jbarth-ubhd

[mikegerber]

TensorFlow and CUDA, the endless joy:

(headless-tf2) root@29b1f018638e:/usr/local/sub-venv/headless-tf2# pip list | grep tensorflow
tensorflow               2.6.0

(headless-tf2) root@29b1f018638e:/usr/local/sub-venv/headless-tf2# find /usr/ -name "libcuda.so*"
/usr/local/cuda-10.0/compat/libcuda.so.1
/usr/local/cuda-10.0/compat/libcuda.so
/usr/local/cuda-10.0/compat/libcuda.so.410.129

So this is CUDA Toolkit 10.0 but PyPi's TensorFlow 2.6.0 requires CUDA Toolkit 11.1 or 11.2.1
(see https://github.com/mikegerber/test-nvidia README for the full list of compatible versions.)

I believe @bertsky was already working on putting more than one CUDA Toolkit version into one image?

[mikegerber]

→ OCR-D/ocrd_all#263

[mikegerber]

AFAICS It's a problem in ocrd_all, not here. I'm copying my comment above to OCR-D/ocrd_all#279 and closing this.