Hi there,
I am using tf_trt_models on jetson NANO with JP 4.2.3, tensorflow 1.14.0.
In detection.py, there is an entry in "MODEL" dict, ssd_resnet_50_fpn_coco. Following the example in detection.ipynb seems to convert successfully to tensorRT model. But the Benchmark gives an Average runtime: 0.57 seconds.
It is werid since ssd_inception_v2_coco gives 0.087 seconds on jetson NANO, that is almost 7 times faster than ssd_resnet_50_fpn_coco. From the model zoo page (https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf1_detection_zoo.md), these two models should not give such a huge difference in inference time (ssd_inception_v2_coco@42ms vs ssd_resnet_50_fpn_coco@76ms).
Another issue is on Faster-Rcnn models. I figured out that inside function build_detection_graph, the line
config.model.faster_rcnn.second_stage_post_processing.score_threshold = score_threshold
should be changed to
config.model.faster_rcnn.second_stage_post_processing.batch_non_max_suppression.score_threshold = score_threshold
in order to convert a faster-rcnn model. However, the Benchmark gives an Average runtime > 1 second! for faster_rcnn_inception_v2_coco. That's really slow.
Hi there, I am using tf_trt_models on jetson NANO with JP 4.2.3, tensorflow 1.14.0.
In detection.py, there is an entry in "MODEL" dict, ssd_resnet_50_fpn_coco. Following the example in detection.ipynb seems to convert successfully to tensorRT model. But the Benchmark gives an Average runtime: 0.57 seconds. It is werid since ssd_inception_v2_coco gives 0.087 seconds on jetson NANO, that is almost 7 times faster than ssd_resnet_50_fpn_coco. From the model zoo page (https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf1_detection_zoo.md), these two models should not give such a huge difference in inference time (ssd_inception_v2_coco@42ms vs ssd_resnet_50_fpn_coco@76ms).
Another issue is on Faster-Rcnn models. I figured out that inside function build_detection_graph, the line
config.model.faster_rcnn.second_stage_post_processing.score_threshold = score_threshold
should be changed toconfig.model.faster_rcnn.second_stage_post_processing.batch_non_max_suppression.score_threshold = score_threshold
in order to convert a faster-rcnn model. However, the Benchmark gives an Average runtime > 1 second! for faster_rcnn_inception_v2_coco. That's really slow.Has anybody encoutered similar problems ?
Many thanks !