YOLOv5n Nano Study

[glenn-jocher]

Starting this discussion to track smaller model experiments as several users have expressed interest. Experiments in https://github.com/ultralytics/yolov5/tree/tests/v6.0. Note: only inference speeds shown, NMS not included in charts below.

Train

python train.py --cfg yolov5n1.yaml
python train.py --cfg yolov5n2.yaml
python train.py --cfg yolov5n3.yaml
python train.py --cfg yolov5n4.yaml
python train.py --cfg yolov5s_v6.yaml

Val: V100 --batch 32

python val.py --batch 32 --data coco.yaml --task speed --weights runs/train/exp/weights/best.pt runs/train/exp2/weights/best.pt runs/train/exp3/weights/best.pt runs/train/exp4/weights/best.pt runs/train/exp5/weights/best.pt yolov5s.pt

Model	size (pixels)	mAPval 0.5:0.95	mAPtest 0.5:0.95	mAPval 0.5	Speed V100 (ms)		params (M)	FLOPs 640 (B)
YOLOv5n1	640	26.1	-	41.9	0.9		1.9	4.5
YOLOv5n2	640	24.1	-	39.8	0.7		1.3	3.3
YOLOv5n3	640	22.9	-	38.4	0.7		1.4	3.2
YOLOv5n4	640	27.6	-	50.1	0.9		2.9	6.7
YOLOv5s release 5.0 (current)	640	36.7	-	55.4	1.0		7.3	17.0
YOLOv5s release 6.0 (pending)	640	37.2	-	56.0	1.0		7.2	16.5

Val: V100 --batch 1

python val.py --batch 1 --data coco.yaml --task speed --weights runs/train/exp/weights/best.pt runs/train/exp2/weights/best.pt runs/train/exp3/weights/best.pt runs/train/exp4/weights/best.pt runs/train/exp5/weights/best.pt yolov5s.pt

Model	size (pixels)	mAPval 0.5:0.95	mAPtest 0.5:0.95	mAPval 0.5	Speed V100 (ms)		params (M)	FLOPs 640 (B)
YOLOv5n1	640	-	-	-	7.8		1.9	4.5
YOLOv5n2	640	-	-	-	7.4		1.3	3.3
YOLOv5n3	640	-	-	-	6.9		1.4	3.2
YOLOv5n4	640	-	-	-	6.8		2.9	6.7
YOLOv5s release 5.0 (current)	640	36.7	-	55.4	8.2		7.3	17.0
YOLOv5s release 6.0 (pending)	640	37.2	-	56.0	8.2		7.2	16.5

Val: V100 --batch 1 --device cpu

python val.py --batch 1 --device cpu --data coco128.yaml --task speed --weights runs/train/exp/weights/best.pt runs/train/exp2/weights/best.pt runs/train/exp3/weights/best.pt runs/train/exp4/weights/best.pt runs/train/exp5/weights/best.pt yolov5s.pt

Model	size (pixels)	mAPval 0.5:0.95	mAPtest 0.5:0.95	mAPval 0.5	Speed CPU (ms)		params (M)	FLOPs 640 (B)
YOLOv5n1	640	-	-	-	35.2		1.9	4.5
YOLOv5n2	640	-	-	-	34.0		1.3	3.3
YOLOv5n3	640	-	-	-	32.3		1.4	3.2
YOLOv5n4	640	-	-	-	54.8		2.9	6.7
YOLOv5s release 5.0 (current)	640	36.7	-	55.4	83.5		7.3	17.0
YOLOv5s release 6.0 (pending)	640	37.2	-	56.0	76.0		7.2	16.5

[ppogg]

Sir, if you have tested the speed of yolov5n on mobile devices? It seems that the Apple m1 chip will speed up the model inference, but how about the performance of yolov5n for ordinary mobile devices, like Android?

By the way, good job！Sir！I like it！

[glenn-jocher]

@ppogg we have not deployed v6.0 to iOS yet, we will do this in the next few days. The nano models should see the most speedup on CPU and low power mobile devices compared to the small models.

[ppogg]

Great, sir. I look forward to your work progress！

[glenn-jocher]

UDPATE: The result of the above study is that the YOLOv5n2 model was selected as the best speed-mAP compromise candidate of the four experimental nano models. So the v6.0 release YOLOv5n model is the YOLOv5n2 model from this study.

This model was also the simplest to understand as we only apply a reduction in width scaling from YOLOv5s (0.5) to YOLOv5n (0.25), which is in line with the scaling of the entire family (0.25 changes between each size).

[henbucuoshanghai]

the nano model of yolov5-6.0 can be 30fps in 4G Nano? img is 6406403

[glenn-jocher]

@henbucuoshanghai I don't understand your question. The above profiling results are at --img 640

[njustczr]

是的 yolov5 n 在nano 4GB上 640x640x3的输入能达到30fps+

[henbucuoshanghai]

tks for reply.   I mean the fps of the yolov5-6.0 on NVIDIA JETSON NANO?

[njustczr]

yolov5n-5.0 on NVIDIA JETSON NANO, the input size is 640x640x3, i can get 30 fps+.

[danurrahmajati]

halo, since the focus on mobile use, is there any example of your implementation to the android device?

[kurtgenc]

Take a look at pytorch web there are many examples:
https://pytorch.org/mobile/android/

[danurrahmajati]

great thanks

[kurtgenc]

I have a custom model trained with yolov5l that I'd like to test inference on Android, is there a way to convert the trained model from yolov5l.pt to yolov5s.pt or other nano models? Or do I need to train for the smaller weight again?

[senstar-hsoleimani]

Train it again

[grobomacOakD]

Is it possible to convert a trained yolov5s to yolov5n?

[glenn-jocher]

@grobomacOakD haha, no. This would be like converting a dog brain into a cat brain.

What you can do is start training a YOLOv5n model from a pretrained YOLOv5s model, though very few layers will transfer:

python train.py --weights yolov5s.pt --cfg yolov5n.yaml
...
Transferred 60/349 items from yolov5s.pt

You're probably just better of training a new YOLOv5n model starting from our pretrained model:

python train.py --weights yolov5n.pt

[henningscale]

Why are these models not run as .engine or .onnx files?
Will the .pt files get published for us to try? I cannot find them manually or with model = torch.hub.load('ultralytics/yolov5', 'yolov5n1') or is it the case that Yolov5n is the same as Yolov5n2?

[glenn-jocher]

@henningscale yes, result of this study was selection of n2 as the YOLOv5n model for the v6.0 release.

[henningscale]

@glenn-jocher How do you get inference speeds that quick? With my 4gb nano and using the v5n6 I get 61ms inference speeds average from 210 images with: python3 detect.py --weights yolov5n6 --nosave --img640 640 --source .../data.

[glenn-jocher]

@henningscale commands and hardware to reproduce are clearly indicated in my original post #5027 (comment), which have zero in common with your hardware and commands. If you compare apples to oranges you shouldn't be surprised to find differences.

[Sohna-Ctrl]

can you share the config file of the yolov5n1、yolov5n2、yolov5n3、yolov5n4，the website(https://github.com/ultralytics/yolov5/tree/tests/v6.0.) is disabled

[glenn-jocher]

@Sohna-Ctrl oh sorry, it seems the branch was auto-deleted upon PR merge. I think you can rest assured that the current yolov5n.yaml is the best balance of speed and accuracy though among the ones tested.

[Sohna-Ctrl]

thanks！but now i have some project needs，maybe your config setting can inspire me to change the model to better fit my datasets.

[Sohna-Ctrl]

Can you re-upload the files for me to study，if it is convenient.

[R97416032]

Hello,sir.Can you tell me how to compute the size of model? The params number of v5n is 1.9M but why the size is about 4MB？

[glenn-jocher]

Size is a function of quantization. FP16 models are 4MB.

[R97416032]

hi，sir！I get into trouble again. After training，it generates a results.txt. There are lr0、lr1 and lr2 in results.txt. I guess lr is learning rate, but I don't know why it has three lr. Thank you sincerely!

[glenn-jocher]

@R97416032 YOLOv5 has 3 parameter groups defined here:

yolov5/train.py

Lines 153 to 175 in b24a576

	g = [], [], [] # optimizer parameter groups
	bn = tuple(v for k, v in nn.__dict__.items() if 'Norm' in k) # normalization layers, i.e. BatchNorm2d()
	for v in model.modules():
	if hasattr(v, 'bias') and isinstance(v.bias, nn.Parameter): # bias
	g[2].append(v.bias)
	if isinstance(v, bn): # weight (no decay)
	g[1].append(v.weight)
	elif hasattr(v, 'weight') and isinstance(v.weight, nn.Parameter): # weight (with decay)
	g[0].append(v.weight)
	if opt.optimizer == 'Adam':
	optimizer = Adam(g[2], lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum
	elif opt.optimizer == 'AdamW':
	optimizer = AdamW(g[2], lr=hyp['lr0'], betas=(hyp['momentum'], 0.999)) # adjust beta1 to momentum
	else:
	optimizer = SGD(g[2], lr=hyp['lr0'], momentum=hyp['momentum'], nesterov=True)
	optimizer.add_param_group({'params': g[0], 'weight_decay': hyp['weight_decay']}) # add g0 with weight_decay
	optimizer.add_param_group({'params': g[1]}) # add g1 (BatchNorm2d weights)
	LOGGER.info(f"{colorstr('optimizer:')} {type(optimizer).__name__} with parameter groups "
	f"{len(g[1])} weight (no decay), {len(g[0])} weight, {len(g[2])} bias")
	del g

[R97416032]

genius!

[jjjonathan14]

Is it possible to get the architecture of yolov5n and pytorch implementation architecture from scratch

[glenn-jocher]

@jjjonathan14 👋 Hello! Thanks for asking about YOLOv5 🚀 architecture visualization. We've made visualizing YOLOv5 🚀 architectures super easy. There are 3 main ways:

model.yaml

Each model has a corresponding yaml file that displays the model architecture. Here is YOLOv5s, defined by yolov5s.yaml:

yolov5/models/yolov5s.yaml

Lines 12 to 48 in 1a3ecb8

	# YOLOv5 v6.0 backbone
	backbone:
	# [from, number, module, args]
	[[-1, 1, Conv, [64, 6, 2, 2]], # 0-P1/2
	[-1, 1, Conv, [128, 3, 2]], # 1-P2/4
	[-1, 3, C3, [128]],
	[-1, 1, Conv, [256, 3, 2]], # 3-P3/8
	[-1, 6, C3, [256]],
	[-1, 1, Conv, [512, 3, 2]], # 5-P4/16
	[-1, 9, C3, [512]],
	[-1, 1, Conv, [1024, 3, 2]], # 7-P5/32
	[-1, 3, C3, [1024]],
	[-1, 1, SPPF, [1024, 5]], # 9
	]
	# YOLOv5 v6.0 head
	head:
	[[-1, 1, Conv, [512, 1, 1]],
	[-1, 1, nn.Upsample, [None, 2, 'nearest']],
	[[-1, 6], 1, Concat, [1]], # cat backbone P4
	[-1, 3, C3, [512, False]], # 13
	[-1, 1, Conv, [256, 1, 1]],
	[-1, 1, nn.Upsample, [None, 2, 'nearest']],
	[[-1, 4], 1, Concat, [1]], # cat backbone P3
	[-1, 3, C3, [256, False]], # 17 (P3/8-small)
	[-1, 1, Conv, [256, 3, 2]],
	[[-1, 14], 1, Concat, [1]], # cat head P4
	[-1, 3, C3, [512, False]], # 20 (P4/16-medium)
	[-1, 1, Conv, [512, 3, 2]],
	[[-1, 10], 1, Concat, [1]], # cat head P5
	[-1, 3, C3, [1024, False]], # 23 (P5/32-large)
	[[17, 20, 23], 1, Detect, [nc, anchors]], # Detect(P3, P4, P5)
	]

TensorBoard Graph

Simply start training a model, and then view the TensorBoard Graph for an interactive view of the model architecture. This example shows YOLOv5s viewed in our Notebook –

# Tensorboard
%load_ext tensorboard
%tensorboard --logdir runs/train

# Train YOLOv5s on COCO128 for 3 epochs
python train.py --weights yolov5s.pt --epochs 3

Netron viewer

Use https://netron.app to view exported ONNX models:

python export.py --weights yolov5s.pt --include onnx --simplify

Good luck 🍀 and let us know if you have any other questions!

[jjjonathan14]

Thank you so much for the quick reply Glenn Jocher and the support. Thanks again

…

On Mon, Jul 11, 2022 at 3:10 AM Glenn Jocher ***@***.***> wrote: @jjjonathan14 <https://github.com/jjjonathan14> 👋 Hello! Thanks for asking about YOLOv5 🚀 *architecture visualization*. We've made visualizing YOLOv5 🚀 architectures super easy. There are 3 main ways: model.yaml Each model has a corresponding yaml file that displays the model architecture. Here is YOLOv5s, defined by yolov5s.yaml: https://github.com/ultralytics/yolov5/blob/1a3ecb8b386115fd22129eaf0760157b161efac7/models/yolov5s.yaml#L12-L48 TensorBoard Graph Simply start training a model, and then view the TensorBoard Graph for an interactive view of the model architecture. This example shows YOLOv5s viewed in our Notebook <https://github.com/ultralytics/yolov5/blob/master/tutorial.ipynb> – [image: Open In Colab] <https://colab.research.google.com/github/ultralytics/yolov5/blob/master/tutorial.ipynb> [image: Open In Kaggle] <https://www.kaggle.com/ultralytics/yolov5> # Tensorboard %load_ext tensorboard %tensorboard --logdir runs/train # Train YOLOv5s on COCO128 for 3 epochs python train.py --weights yolov5s.pt --epochs 3 [image: Screenshot 2021-04-11 at 01 10 09] <https://user-images.githubusercontent.com/26833433/114286928-349bd600-9a63-11eb-941f-7139ee6cd602.png> Netron viewer Use https://netron.app to view exported ONNX models: python export.py --weights yolov5s.pt --include onnx --simplify [image: Screen Shot 2022-04-29 at 11 09 23 AM] <https://user-images.githubusercontent.com/26833433/165999628-0d6095ce-58dd-49b7-a04e-25d0995e5493.png> Good luck 🍀 and let us know if you have any other questions! — Reply to this email directly, view it on GitHub <#5027 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AZAGX3VYTVSIV3V4LDNTFSLVTM7NLANCNFSM5FFWHEJA> . You are receiving this because you were mentioned.Message ID: ***@***.***>