GPU processors: (uniformly) log CPU vs GPU use
bertsky opened this issue · comments
Due to the lack of a specification on that aspect, our processors have no or no uniform way to inform the user whether or not a GPU device is used (or even parameterise which one to prefer). Here's the current state of affairs in this regard:
- TF based processors:
- ocrd-eynollah-segment: WARNING
no GPU device available
- ocrd-anybaseocr-block-segmentation: WARNING
Tensorflow cannot detect CUDA installation. Running without GPU will be slow
- ocrd-cor-asv-ann-process: INFO
using ... LSTM implementation to compile ...
- ocrd-keraslm-rate: INFO
using ... LSTM implementation to compile ...
- all others (ocrd-calamari-recognize, ocrd-sbb-binarize, ocrd-anybaseocr-tiseg, ocrd-anybaseocr-layout-analysis): rely on TF logging, which due to Tensorflow gabbiness we have to mute via envvar
TF_CPP_MIN_LOG_LEVEL=3
- ocrd-eynollah-segment: WARNING
- Torch based processors:
- ocrd-kraken-segment: nothing
- ocrd-kraken-recognize: nothing
- ocrd-typegroups-classifier: nothing
- ocrd-detectron2-segment: INFO
Using compute device ...
- ocrd-anybaseocr-dewarp: WARNING
torch cannot detect CUDA installation
The good news is that when there is a GPU, we try to utilise it. But the user has no way to validate this works (other than perhaps make test-cuda
in ocrd_all).
Personally, I like the warnings best. Could we agree on standardising and enforcing these?
@bertsky : Many thanks for opening this issue. The "Warning"-variant would be very nice.
In my opinion, this is needed,
e.g. in my environment I can only "believe" that ocrd-kraken-segment
uses the GPU.
e.g. in my environment I can only "believe" that
ocrd-kraken-segment
uses the GPU.
BTW I just amended OCR-D/ocrd_kraken#38 to include such warnings for Kraken.
Having consistent WARNING
about this across processors would be a good first step indeed.