GPU-Suport: Mask-RCNN + Minor GPU fixes #2714
Conversation
|
@jahaniam Thanks for the your contribution! |
Thank you and your team. |
|
@jahaniam , could you please help us to fix codacy issues in the PR? |
| build: | ||
| image: cvat/tf.matterport.mask_rcnn | ||
| baseImage: tensorflow/tensorflow:2.1.0-py3 | ||
| baseImage: tensorflow/tensorflow:1.13.1-py3 |
There was a problem hiding this comment.
The latest version of tensorflow 1.x is 1.15.5 and I see the comment Note that this is the last patch release for the TensorFlow 1.x series.
What is a reason to move from 2.1 to 1.13.1?
There was a problem hiding this comment.
Before you were installing RUN pip install tensorflow==1.13.1 . RUN can't be overwritten using nuctl for gpu. If I use tensorflow 2.1-GPU I was getting some errors. Besides the mask-rcnn code is in tensorflow 1.x. It is better to have the docker for 1.x .
There was no reason not to go with v 1.15.5, we might be able to switch to that version probably. I just needed a version that works fine. I used 1.13.1 because it was being installed on line 119.
|
@jahaniam , the patch looks great! Let's clarify a couple of moments and merge. |
There was a problem hiding this comment.
@jahaniam , could you please help us to fix codacy issues in the PR?
vat/apps/documentation/installation_automatic_annotation.md [maximum-line-length] Line must be at most 120 characters Also you will need to add `--resource-limit nvidia.com/gpu=1 --triggers '{"myHttpTrigger": {"maxWorkers": 1}}'` to the nuclio deployment command. You can increase the maxWorker if you have enough GPU memory. [maximum-line-length] Line must be at most 120 characters - Since the model is loaded during deployment, the number of GPU deployed functions will be limited to your GPU memory. [list-item-content-indent] Don’t use mixed indentation for children, remove 2 spaces - Since the model is loaded during deployment, the number of GPU deployed functions will be limited to your GPU memory. [list-item-indent] Incorrect list-item indent: add 2 spaces - Since the model is loaded during deployment, the number of GPU deployed functions will be limited to your GPU memory. serverless/tensorflow/faster_rcnn_inception_v2_coco/nuclio/model_loader.py Trailing whitespace
Do you know how I can recheck it? I did some changes, I wanna make sure it is ok then commit it.
|
All comments are addressed. The only thing remaining is the codacy complaining about two spaces for the list item (my indentings are correct). I am not able to fix that. If you can fix it please go ahead. |
|
@jahaniam , thanks for the great contribution! Really appreciate your time and efforts. |
* fixed cpu mask rcnn+preparation for gpu * fix-limit gpu memory to 30% of total memory per worker Co-authored-by: Nikita Manovich <nikita.manovich@intel.com>
|
Hello, @jahaniam you reported these times Mask-RCNN GPU: ~1.2 Sec / Image this time is to use the model to predict one image annotations ?? |
Yes. |
Summary:
Here is a short result on my laptop with Nvidia 1060m :
Related Issues:
#2635,
#2489,
Boosted to nuclio to 1.5.16, based on nuclio/nuclio#2058 ,
processorMountModeis deprecated and is replaced withmountMode, therefore, #2578 needs to be revisited.Checklist
developbranchcvat-core, cvat-data and cvat-ui)
License
Feel free to contact the maintainers if that's a concern.