Documentation bug: model weights in object detection documentation do not exist
dgtlmoon opened this issue · comments
https://www.deepdetect.com/platform/docs/object-detection/
dd@a0a6e47be381:/opt/deepdetect/build/main$ ls -al /opt/platform/models/pretrained
ls: cannot access '/opt/platform/models/pretrained': No such file or directory
In the latest GPU docker image, I have only
dd@a0a6e47be381:/opt/deepdetect/build/main$ find /opt|grep caffemodel
/opt/models/resnet_50/ResNet-50-model.caffemodel
/opt/models/ggnet/bvlc_googlenet.caffemodel
https://www.deepdetect.com/platform/docs/object-detection/
dd@d8bfd627e7fd:/opt/deepdetect/build/main$ find /opt|grep caffemod
/opt/models/resnet_50/ResNet-50-model.caffemodel
/opt/models/ggnet/bvlc_googlenet.caffemodel
I checked the docker images for cpu platform docker
, gpu
and cpu
and not in those either
Hi, yes it's a known issue on some installs. The tarball is available from https://www.deepdetect.com/downloads/platform/pretrained_latest.tar.gz
Could be nice to add the following comments (as an example)
template - used in the "mllib":{ "template":
field, this is required, if we were not finetuning (finetuning
not set or is false
) then we can just use this template to train new weights on, otherwise you need to provide the template
and the current pretrained weights
, which should be for the same model.
@alx easy solution would be, under the table of models, just include a link "Download the pretrained weights file for the model here https://www.deepdetect.com/downloads/platform/pretrained/"
Hi @dgtlmoon
Thanks again for your feedbacks, I've updated the documentation page using them: https://www.deepdetect.com/platform/docs/object-detection/
I'm not sure where to insert your comment about the template parameter, you provide a curl parameter while this page is showing an example of a jupyter notebook snippet.
Documentation has been updated and published, feel free to re-open this issue to append new comments about this section.
We're planning to move the documentation on github, you should soon be able to make PR on it.
Have a nice day,
Alex