PRG's Setup of Intel Neural Compute Stick
- Step 1
- Step 2 (Needs PC Restart after this!)
- Step 3:
sudo bash /opt/intel/openvino/deployment_tools/demo/setupvars.sh
- Step 4:
cd /opt/intel/openvino/deployment_tools/demo && ./demo_benchmark_app.sh -d MYRIAD
- NOTE: All tests were run at USB 2.0
- Step 1:
cd NCS-2 && python tf_model.py -s [1,256,256,1] -l [4,8,16]
- Step 2:
python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo_tf.py --input_meta_graph mnist_model.meta --input_shape [1,256,256,1] --data_type FP16
- Step 3:
source /opt/intel/openvino/bin/setupvars.sh
- Step 4:
python3 movi2.py
Image size is [BWHC] [ batch size x width x height x color channels] and ResNet architecture. Num Output layers = 4.
Image size | Filters in subsequent layers | FPS | NumParams | NumFLOPs |
---|---|---|---|---|
1 x 256 x 256 x 1 | [16,32,64,128] | 7.5 | 745364 | 3559699941 |
1 x 512 x 512 x 1 | [16,32,64,128] | 2.0 | 745364 | 14234334693 |
1 x 256 x 256 x 2 | [16,32,64,128] | 7.25 | 745524 | 3580671781 |
1 x 256 x 256 x 6 | [16,32,64,128] | 6.01 | 746164 | 3664559141 |
1 x 256 x 256 x 1 | [8,16,32,64] | 11.2 | 186828 | 917679877 |
1 x 512 x 512 x 1 | [8,16,32,64] | 3.0 | 186828 | 3669602053 |
1 x 256 x 256 x 2 | [8,16,32,64] | 10.7 | 186908 | 928165797 |
1 x 256 x 256 x 6 | [8,16,32,64] | 8.2 | 187228 | 970109477 |
1 x 256 x 256 x 1 | [10,15,23,35,53,80,120] | 10.3 | 917033 | 824121066 |
1 x 512 x 512 x 1 | [10,15,23,35,53,80,120] | 2.8 | 917033 | 3290991978 |
1 x 256 x 256 x 2 | [10,15,23,35,53,80,120] | 9.9 | 917133 | 837228466 |
1 x 256 x 256 x 6 | [10,15,23,35,53,80,120] | 7.7 | 917533 | 889658066 |
1 x 256 x 256 x 6 | [8,12,18,27,41,61,92] | 8.4 | 540871 | 580564264 |
- Increasing Depth has small effect on speed
- Increasing Width has a huge effect on speed
- Input dimension size has a huge effect on speed