WHY! v4 detection ETA is 10 times than v3 ??!!

[12343954]

Yolo V4 ， training avg-loss=3.0 , CUDA 10.2 + cuDNN 7.6.5

D:\Darknet-YoloV4\darknet-yolov4\build\darknet\x64>darknet_images.cmd
Flag value false not forcing CPU mode
 Try to load cfg: ./training/11/voc_custom/yolov4_custom.cfg, weights: ./training/11/voc_custom/backup/yolov4-custom_best.weights, clear = 0
 0 : compute_capability = 750, cudnn_half = 0, GPU: GeForce RTX 2060
net.optimized_memory = 0
mini_batch = 1, batch = 64, time_steps = 1, train = 0
   layer   filters  size/strd(dil)      input                output
   0 conv     32       3 x 3/ 1    608 x 608 x   3 ->  608 x 608 x  32 0.639 BF
   1 conv     64       3 x 3/ 2    608 x 608 x  32 ->  304 x 304 x  64 3.407 BF
   2 conv     64       1 x 1/ 1    304 x 304 x  64 ->  304 x 304 x  64 0.757 BF
   3 route  1                                      ->  304 x 304 x  64

Yolo V3 ， training avg-loss=0.2 , CUDA 10.1 + cuDNN 7.6.4

darknet detector test cfg/coco.data cfg/yolov3.cfg yolov3.weights data/dog.jpg
 CUDA-version: 10010 (10010), cuDNN: 7.6.4, CUDNN_HALF=1, GPU count: 1
 OpenCV version: 4.2.0
 compute_capability = 750, cudnn_half = 1
net.optimized_memory = 0
batch = 1, time_steps = 1, train = 0
   layer   filters  size/strd(dil)      input                output
   0 conv     32       3 x 3/ 1    416 x 416 x   3 ->  416 x 416 x  32 0.299 BF
   1 conv     64       3 x 3/ 2    416 x 416 x  32 ->  208 x 208 x  64 1.595 BF

WHERE am i WRONG ??? Why my v4 cudnn_half = 0 ??

[12343954]

put v3 and v4 under the same CUDA 10.1 + cuDNN 7.6.4 same code detect same image on the same computer

ETA v3:v4 = 39ms : 622ms

v3

D:\Tensorflow2\darknet\build\darknet\x64>detect.cmd
 Try to load cfg: training\11\voc_custom\yolov3_custom.cfg, weights: training\11\voc_custom\backup\yolov3_custom_63000.weights, clear = 0
 compute_capability = 750, cudnn_half = 1
net.optimized_memory = 0
batch = 1, time_steps = 1, train = 0
   layer   filters  size/strd(dil)      input                output
   0 conv     32       3 x 3/ 1    416 x 416 x   3 ->  416 x 416 x  32 0.299 BF
   1 conv     64       3 x 3/ 2    416 x 416 x  32 ->  208 x 208 x  64 1.595 BF
   2 conv     32       1 x 1/ 1    208 x 208 x  64 ->  208 x 208 x  32 0.177 BF
   3 conv     64       3 x 3/ 1    208 x 208 x  32 ->  208 x 208 x  64 1.595 BF

v4

D:\Darknet-YoloV4\darknet-yolov4\build\darknet\x64>detect.cmd
name 'DARKNET_FORCE_CPU' is not defined
 Try to load cfg: training\11\voc_custom\yolov4_custom.cfg, weights: training\11\voc_custom\backup\yolov4-custom_best.weights, clear = 0
 0 : compute_capability = 750, cudnn_half = 0, GPU: GeForce RTX 2060
net.optimized_memory = 0
mini_batch = 1, batch = 64, time_steps = 1, train = 0
   layer   filters  size/strd(dil)      input                output
   0 conv     32       3 x 3/ 1    608 x 608 x   3 ->  608 x 608 x  32 0.639 BF
   1 conv     64       3 x 3/ 2    608 x 608 x  32 ->  304 x 304 x  64 3.407 BF
   2 conv     64       1 x 1/ 1    304 x 304 x  64 ->  304 x 304 x  64 0.757 BF
   3 route  1

[AlexeyAB]

1 Run the same command for both v3 and v4 ./darknet detector test coco.data yolov4.cfg yolov4.weights data/dog.jpg on the same PC

2. Do you use implementation for rotated-bbox? What implementation do you use?

[stephanecharette]

put v3 and v4 under the same CUDA 10.1 + cuDNN 7.6.4 same code detect same image

Is this on the same computer, same darknet directory, etc?

Cause I can confirm that when I run YOLO v3 or v4 detection on the same computer, I get approximately the same results. Around 3-4 milliseconds for most of my neural networks. And then if I rebuild darknet without GPU support, then I get numbers like 300-700 milliseconds since it runs on the CPU.

I suspect that whatever you are doing, your v4 is running on the CPU.

[12343954]

@AlexeyAB @stephanecharette Thank you very much for replying.

1, I tested v4 & v3 on the same PC via python. 2, My rotated-bbox code is running on both v4 and v3 darknet framework under GPU. but I don't test with one platform two yolo.cfgs. You can see my two different folders.

3. I am sure that the above results are all tested under GPU, but the ETAs are very different.
   because when I tested on CPU，the v4 ETA is more than 1000ms！！！

I suspect that cudnn is not turned on under v4. what's means of cudnn_half = 0 ?

I used v4 & v3 with their own cfg, wight. I trained their own data set separately, but the same name in different folders. This will ensure that my code changes minimally.

[AlexeyAB]

1 Run the same command for both v3 and v4 ./darknet detector test coco.data yolov4.cfg yolov4.weights data/dog.jpg on the same PC, and show screenshot

2. Do you use implementation for rotated-bbox? What implementation do you use? Show a link.

[12343954]

all test under CUDA 10.1 + cuDNN 7.6.4，same PC

v4，darknet.exe detector test data/coco.data cfg/yolov4.cfg yolov4.weights data/dog.jpg

 CUDA-version: 10010 (11000), cuDNN: 7.6.4, CUDNN_HALF=1, GPU count: 1
 CUDNN_HALF=1
 OpenCV version: 4.2.0
 0 : compute_capability = 750, cudnn_half = 1, GPU: GeForce RTX 2060
net.optimized_memory = 0
mini_batch = 1, batch = 8, time_steps = 1, train = 0
   layer   filters  size/strd(dil)      input                output
   0 conv     32       3 x 3/ 1    608 x 608 x   3 ->  608 x 608 x  32 0.639 BF
   1 conv     64       3 x 3/ 2    608 x 608 x  32 ->  304 x 304 x  64 3.407 BF
   2 conv     64       1 x 1/ 1    304 x 304 x  64 ->  304 x 304 x  64 0.757 BF

v4 python darknet_images.py --input=data/dog.jpg --config_file=cfg/yolov4.cfg --weights=yolov4.weights --data_file=data/coco.data

name 'DARKNET_FORCE_CPU' is not defined
 Try to load cfg: cfg/yolov4.cfg, weights: yolov4.weights, clear = 0
 0 : compute_capability = 750, cudnn_half = 0, GPU: GeForce RTX 2060
net.optimized_memory = 0
mini_batch = 1, batch = 8, time_steps = 1, train = 0
   layer   filters  size/strd(dil)      input                output
   0 conv     32       3 x 3/ 1    608 x 608 x   3 ->  608 x 608 x  32 0.639 BF
   1 conv     64       3 x 3/ 2    608 x 608 x  32 ->  304 x 304 x  64 3.407 BF
   2 conv     64       1 x 1/ 1    304 x 304 x  64 ->  304 x 304 x  64 0.757 BF

v4 (No GPU) python darknet_images.py --input=data/dog.jpg --config_file=cfg/yolov4.cfg --weights=yolov4.weights --data_file=data/coco.data

Environment variables indicated a CPU run, but we didn't find D:\Darknet-YoloV4\darknet-yolov4\build\darknet\x64\yolo_cpp_dll_nogpu.dll. Trying a GPU run anyway.
 Try to load cfg: cfg/yolov4.cfg, weights: yolov4.weights, clear = 0
mini_batch = 1, batch = 8, time_steps = 1, train = 0
   layer   filters  size/strd(dil)      input                output
   0 conv     32       3 x 3/ 1    608 x 608 x   3 ->  608 x 608 x  32 0.639 BF
   1 conv     64       3 x 3/ 2    608 x 608 x  32 ->  304 x 304 x  64 3.407 BF

v3，darknet.exe detector test data/coco.data cfg/yolov3.cfg yolov3.weights data/dog.jpg

CUDA-version: 10010 (11000), cuDNN: 7.6.5, CUDNN_HALF=1, GPU count: 1
 OpenCV version: 4.2.0
 compute_capability = 750, cudnn_half = 1
net.optimized_memory = 0
batch = 1, time_steps = 1, train = 0
   layer   filters  size/strd(dil)      input                output
   0 conv     32       3 x 3/ 1    416 x 416 x   3 ->  416 x 416 x  32 0.299 BF
   1 conv     64       3 x 3/ 2    416 x 416 x  32 ->  208 x 208 x  64 1.595 BF

[AlexeyAB]

@12343954

So from your screenshots on GeForce RTX 2060:

• YOLOv4-608 - 39ms
• YOLOv3-416 - 20ms (2x faster)

---

This fully matches the chart from the article: https://arxiv.org/abs/2004.10934 GPU Tesla V100

• YOLOv4-608 - 16ms (62 FPS)
• YOLOv3-416 - 8ms (120 FPS) (2x faster)

[12343954]

@AlexeyAB Thank you for reply!

Though my test, I see darknet.exe different performance between v4 and v3. but i don't understand why the same code running under python are very different. 10 times!! is this because v4's cudnn_half is unavailable ? I don't change your code darknet_images.py.

[AlexeyAB]

Does https://github.com/AlexeyAB/darknet/blob/master/darknet_images.py support rotated bboxes?

[12343954]

😄 sorry, forgot answer your 2nd question. I implement the rotated-bbox via openCV myself, I combined some technical information on the Internet.

[robotaiguy]

Were you, by chance, using the default 416x416 network size for Yolo v3, and now using the default 608x608 network size for Yolo v4? That should be enough to make a notable difference in side-by-side comparisons.