How to use GPU on embedded devices

[ChenZhenGui]

I tried to run the demo on the Jetson Nano using the GPU. I set CUDA_VISIBLE_DEVICES=0, but found that the GPU did not work after running

[ly015]

Most of our demos use 'cuda:0' as the default device for model inference. Could you please specify the demo script and arguments you used when encountering this issue?

[ChenZhenGui]

webcam_demo.py:

Copyright (c) OpenMMLab. All rights reserved.

import logging from argparse import ArgumentParser

from mmengine import Config, DictAction

from mmpose.apis.webcam import WebcamExecutor from mmpose.apis.webcam.nodes import model_nodes

def parse_args(): parser = ArgumentParser('Webcam executor configs') parser.add_argument( '--config', type=str, default= './demo/webcam_cfg/pose_estimation.py')

parser.add_argument(
    '--cfg-options',
    nargs='+',
    action=DictAction,
    default={},
    help='Override settings in the config. The key-value pair '
    'in xxx=yyy format will be merged into config file. For example, '
    "'--cfg-options executor_cfg.camera_id=1'")
parser.add_argument(
    '--debug', action='store_true', help='Show debug information.')
parser.add_argument(
    '--cpu', action='store_true', help='Use CPU for model inference.')
parser.add_argument(
    '--cuda',
    action='store_true',
    default=True,
    help='Use GPU for model inference.')

return parser.parse_args()

def set_device(cfg: Config, device: str): """Set model device in config.

Args:
    cfg (Config): Webcam config
    device (str): device indicator like "cpu" or "cuda:0"
"""

device = device.lower()
assert device == 'cpu' or device.startswith('cuda:')

for node_cfg in cfg.executor_cfg.nodes:
    if node_cfg.type in model_nodes.__all__:
        node_cfg.update(device=device)

return cfg

def run(): args = parse_args() cfg = Config.fromfile(args.config) cfg.merge_from_dict(args.cfg_options)

if args.debug:
    logging.basicConfig(level=logging.DEBUG)

if args.cpu:
    cfg = set_device(cfg, 'cpu')

if args.cuda:
    cfg = set_device(cfg, 'cuda:0')

webcam_exe = WebcamExecutor(**cfg.executor_cfg)
webcam_exe.run()

if name == 'main': run()

[ChenZhenGui]

pose_estimation.py:

Copyright (c) OpenMMLab. All rights reserved.

executor_cfg = dict( # Basic configurations of the executor name='Pose Estimation', camera_id='http://admin:[email protected]:8081/', # Define nodes. # The configuration of a node usually includes: # 1. 'type': Node class name # 2. 'name': Node name # 3. I/O buffers (e.g. 'input_buffer', 'output_buffer'): specify the # input and output buffer names. This may depend on the node class. # 4. 'enable_key': assign a hot-key to toggle enable/disable this node. # This may depend on the node class. # 5. Other class-specific arguments nodes=[ # 'DetectorNode': # This node performs object detection from the frame image using an # MMDetection model. dict( type='DetectorNode', name='detector', model_config='./demo/mmdetection_cfg/' 'ssdlite_mobilenetv2-scratch_8xb24-600e_coco.py', model_checkpoint='./demo/webcam_cfg/'
'ssdlite_mobilenetv2_scratch_600e_coco_20210629_110627-974d9307.pth', input_buffer='input', # _input_ is an executor-reserved buffer output_buffer='det_result'), # 'TopDownPoseEstimatorNode': # This node performs keypoint detection from the frame image using an # MMPose top-down model. Detection results is needed. dict( type='TopDownPoseEstimatorNode', name='human pose estimator', model_config='./configs/body_2d_keypoint/topdown_heatmap/coco/mobilevit_coco-256x192.py', model_checkpoint="./work_dirs/AP_epoch_300.pth", labels=['person'], input_buffer='det_result', output_buffer='human_pose'), # 'ObjectAssignerNode': # This node binds the latest model inference result with the current # frame. (This means the frame image and inference result may be # asynchronous). dict( type='ObjectAssignerNode', name='object assigner', frame_buffer='frame', # _frame_ is an executor-reserved buffer object_buffer='human_pose', output_buffer='frame'), # 'ObjectVisualizerNode': # This node draw the pose visualization result in the frame image. # Pose results is needed. dict( type='ObjectVisualizerNode', name='object visualizer', enable_key='v', enable=True, show_bbox=True, must_have_keypoint=False, show_keypoint=True, input_buffer='frame', output_buffer='vis'), # 'SunglassesNode': # This node draw the sunglasses effect in the frame image. # Pose results is needed. dict( type='SunglassesEffectNode', name='sunglasses', enable_key='s', enable=False, input_buffer='vis', output_buffer='vis_sunglasses'), # # 'BigeyeEffectNode': # # This node draw the big-eye effetc in the frame image. # # Pose results is needed. dict( type='BigeyeEffectNode', name='big-eye', enable_key='b', enable=False, input_buffer='vis_sunglasses', output_buffer='vis_bigeye'), # 'NoticeBoardNode': # This node show a notice board with given content, e.g. help # information. dict( type='NoticeBoardNode', name='instruction', enable_key='h', enable=False, input_buffer='vis_bigeye', output_buffer='vis_notice', content_lines=[ 'This is a demo for pose visualization and simple image ' 'effects. Have fun!', '', 'Hot-keys:', '"v": Pose estimation result visualization', '"s": Sunglasses effect B-)', '"b": Big-eye effect 0_0', '"h": Show help information', '"m": Show diagnostic information', '"q": Exit' ], ), # 'MonitorNode': # This node show diagnostic information in the frame image. It can # be used for debugging or monitoring system resource status. dict( type='MonitorNode', name='monitor', enable_key='m', enable=False, input_buffer='vis_notice', output_buffer='display'), # 'RecorderNode': # This node save the output video into a file. dict( type='RecorderNode', name='recorder', out_video_file='webcam_demo.mp4', input_buffer='display', output_buffer='display' # _display_ is an executor-reserved buffer ) ])

[ChenZhenGui]

the command is : CUDA_VISIBLE_DEVICES=0 python demo/webcam_demo.py

[ChenZhenGui]

But I find the GPU:nvidia,tegra210-gm20b did not work

[ly015]

Could you please check the GPU availability by the following commands?

python -c "import torch; print(torch.cuda.is_available())"