Skip to content

Instantly share code, notes, and snippets.

@ivder
Created July 15, 2019 07:58
Show Gist options
  • Save ivder/f7e3d87fdf204f0fb868a3a5f5cdbfda to your computer and use it in GitHub Desktop.
Save ivder/f7e3d87fdf204f0fb868a3a5f5cdbfda to your computer and use it in GitHub Desktop.
Train
./darknet detector train data/obj.data cfg/obj.cfg darknet53.conv.74 -map
Test
./darknet detector test data/obj.data cfg/obj.cfg backup/obj_last.weights
Test Multiple from test.txt
./darknet detector test data/obj.data cfg/obj.cfg backup/obj_6000.weights -i 0 -thresh 0.25 -dont_show -ext_output < data/obj/test.txt > detected.txt
Test C++
LD_LIBRARY_PATH=./:$LD_LIBRARY_PATH ./uselib data/obj.names cfg/obj.cfg backup/obj_last.weights data/out.mp4
Test mAP
./darknet detector map data/obj.data cfg/obj.cfg backup/obj_6000.weights
Calc Anchors
./darknet detector calc_anchors data/obj.data -num_of_clusters 9 -width 608 -height 608
Important notes:
cfg parameters : https://github.com/AlexeyAB/darknet/issues/279
every 10 iterations network size is resized
In repo https://github.com/AlexeyAB/darknet input size(width and height in .cfg file) matters for Training, since network will be resized from (1/1.4)x to (1.4)x from width x height in cfg file.
* if width=416 height=416 then network will be resized from 288 x 288 to 576 x 576
Batch=64 -> loading 64 images for this "iteration".
Subdivision=8 -> Split batch into 8 "mini-batches" so 64/8 = 8 images per "minibatch" and this get sent to the gpu for process.
That will be repeated 8 times until the batch is completed and a new itereation will start with 64 new images.
Note and Todo:
C : originally resized to network size (416x416 or 608x608)
C++ : original size (960x540)
Accuracy Comparison
-Spp 608x808 : map: 74.8%, avg loss: 0.07, miss: 12/93
-416x416 : map: 70.7%, avg loss: 0.1 , miss: 13/93
-416x416 scratch : map: 60.6%, avg loss: 0.17, miss: 19/93
-Tiny 416x416 : map: 57.5%, avg loss: 0.34, miss: 19/93
Inference Speed Comparison
GTX 1080
-Spp 608x608 : 23 ms
-416x416 : 13 ms
-C++ : 17 ms
-C++ ROI : 16 ms
-C++ resize : 15 ms
-Tiny 416x416 : 2 ms
-C++ ROI : 5 ms
-C++ resize : 4 ms
Jetson TX2
-416x416 : 223 ms
-C++ ROI : 230 ms
-Tiny 416x416 : 26~60 ms (not stable)
-C++ ROI : 30 ms~50 ms (not stable)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment