Faster R-CNN: too many resources requested for launch

Autonomous Machines Jetson & Embedded Systems Jetson TX2
21

Hi,

Has there been any progress on this?

I am trying to use the Tensorflow Mask RCNN inception v2 coco (from the model zoo: https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md). It fails with the ‘to many resources requested for launch’ error on the first image I try to process when I use the Jetson GPU.

If I do as was suggested in #20 (instruct TF to use the CPU), then it works, but it is very slow :(

I am using Tensorflow 1.5, Jetson TX2 with Jetpack L4T 3.2 b157.

This works fine on my laptop and uses less than 6GB of the CPU memory.

I really need a solution to this otherwise I will need to give up on the TX2 since I need somthing that can do Mask RCNN inference in real time!

Thx,
Heidi

22

Hi,

Comment #20 is for debugging, not a fix of this issue.

First, it’s recommended to reboot your device since this can fix some TensorFlow resource issue.
If no luck, please check this page for an alternative:

The page contains the profiling result of various models which are available for Jetson.

Thanks.

23

@AastaLLL So are there any updates? When I use the pytorch version of mask r-cnn, the same error occurs: too many resources requested for launch. The version I used is : GitHub - multimodallearning/pytorch-mask-rcnn

@heidi7y5lm Moreover, how was your project, did you fix the problem by any methods or any alternatives?

Thanks a lot!

24

Hi,

Could you check the memory amount used by your application first?

sudo ./tegrastats

A user has implemented mask r-cnn for Jetson but with the TensorFlow interface:

Thanks.

25

hi @AastaLLL the repo you shared contains my implementation of mask rcnn with mobilenet as backbone (based on the keras model of github/matterport).

It worked and accuracy is at acceptable levels but still it is not deployable on the jetson, as i stated in other issues: (https://devtalk.nvidia.com/default/topic/1033209/general-question-about-jetsons-gpu-cpu-shared-memory-usage/?offset=15#5265335) (https://devtalk.nvidia.com/default/topic/1013464/jetson-tx2/gpu-out-of-memory-when-the-total-ram-usage-is-2-8g/2)

I furthermore implemented a mask rcnn with mobilenet feature extractor based on the tensorflow object detection api.

I did a lightweight model trained only on one class with input size 224x224 and a few other modifiactions to make it smaller. On my host with 4GB it runs super fast, but on the Jetson the same error as others stated for faster r-cnn, although it consumes only roughly 1.8GB in total.

26

Hi, gustavvz

Let’s check this memory issue on the topic 1013464:
[url]https://devtalk.nvidia.com/default/topic/1013464/jetson-tx2/gpu-out-of-memory-when-the-total-ram-usage-is-2-8g/post/5267017/#5267017[/url]

Thanks.

27

hi @AastaLLL and @gustavvz even i am having a hard time running the faster rcnn model o the tx2 board
i tried moving some operations n gpu and cpu and it worked but the detections are not happening since their is loss in the data during the moving of operations.
Did u guys try any other method to deploy faster rcnn on board

28

Hi,

We have faster RCNN sample in the TensorRT package.
It’s recommended to give it a try first:

cp /usr/src/tensorrt/ ~
cd tensorrt
wget --no-check-certificate https://dl.dropboxusercontent.com/s/o6ii098bu51d139/faster_rcnn_models.tgz?dl=0 -O data/faster-rcnn/faster-rcnn.tgz
tar zxvf data/faster-rcnn/faster-rcnn.tgz -C data/faster-rcnn --strip-components=1 --exclude=ZF_*
cd samples/sampleFasterRCNN/
cd ../../bin/
./sample_fasterRCNN

Thanks.