:zap: A newly designed ultra lightweight anchor free target detection algorithm， weight only 250K parameters， reduces the time consumption by 10% compared with yolo-fastest, and the post-processing is simpler
License: BSD 3-Clause "New" or "Revised" License
使用OpenCV部署FastestDet，包含C++和Python两种版本的程序。模型文件不超过1M
https://github.com/hpc203/FastestDet-opencv-dnn
import json
import cv2
import os
import matplotlib.pyplot as plt
import shutil
from tqdm import tqdm
def load_images_from_folder(input_path):
file_names = []
for filename in tqdm(os.listdir(input_path)):
file_names.append(filename)
return file_names
def get_img_ann(image_id):
img_ann = []
isFound = False
for ann in data['annotations']:
if ann['image_id'] == image_id:
img_ann.append(ann)
isFound = True
if isFound:
return img_ann
else:
return None
def get_img(filename):
for img in data['images']:
if img['file_name'] == filename:
return img
if __name__ == "__main__":
input_path = "/home/dzhang/data/coco/val2017/"
output_path = "/home/dzhang/data/coco_fastest_net/val2017"
annotation_path = "/home/dzhang/data/coco/annotations/instances_val2017.json"
output_txt_path = "/home/dzhang/data/coco_fastest_net/val2017.txt"
os.system(f"rm -rf {output_path}")
os.makedirs(output_path, exist_ok=True)
with open(annotation_path, 'r') as f:
data = json.load(f)
file_names = load_images_from_folder(input_path)
labeled_file_names = []
count = 0
print("Processing labels...")
for filename in tqdm(file_names):
# Extracting image
img = get_img(filename)
img_id = img['id']
img_w = img['width']
img_h = img['height']
# Get Annotations for this image
img_ann = get_img_ann(img_id)
if img_ann:
# Opening file for current image
file_object = open(f"{output_path}/img{count}.txt", "a")
for ann in img_ann:
current_category = ann['category_id'] - 1 # As yolo format labels start from 0
if current_category > 79:
print("Current category larger than 79")
continue
current_bbox = ann['bbox']
x = current_bbox[0]
y = current_bbox[1]
w = current_bbox[2]
h = current_bbox[3]
# Finding midpoints
x_centre = (x + (x+w))/2
y_centre = (y + (y+h))/2
# Normalization
x_centre = x_centre / img_w
y_centre = y_centre / img_h
w = w / img_w
h = h / img_h
# Limiting upto fix number of decimal places
x_centre = format(x_centre, '.6f')
y_centre = format(y_centre, '.6f')
w = format(w, '.6f')
h = format(h, '.6f')
# Writing current object
file_object.write(f"{current_category} {x_centre} {y_centre} {w} {h}\n")
file_object.close()
# copy image here if successful
source = os.path.join(input_path, filename)
destination = f"{output_path}/img{count}.jpg"
labeled_file_names.append(destination)
shutil.copy(source, destination)
count += 1 # This should be outside the if img_ann block.
else:
print(f"Image {filename}'s label not found.")
with open(output_txt_path, 'w') as f:
for filename in labeled_file_names:
f.write(f"{filename}\n")
print("Labels processed successfully.")I use the above code to transfer coco dataset to standard format, but the evaluation result seems not to be right:
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.028
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.054
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.027
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.006
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.025
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.051
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.031
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.043
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.045
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.013
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.047
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.079
Do you know why?
本地gpu太low了。没办法跑测试。
按照：
dog-qiuqiu/Yolo-FastestV2@3922174
修改失败了。但能跑test.py，但无法跑train.py，不晓得怎么改里面具体的。
hello, dear author.
能否提供一下网络的结构示意图?
Hi :)
Thanks for this repo
I have a problem with setting a custom input resolution in YAML config
DATASET:
TRAIN: path\to\train\txt
VAL: path\to\test\txt
NAMES: path\to\names
MODEL:
NC: 1
INPUT_WIDTH: 240
INPUT_HEIGHT: 240
TRAIN:
LR: 0.001
THRESH: 0.25
WARMUP: true
BATCH_SIZE: 64
END_EPOCH: 200
MILESTIONES:
- 50
- 100
- 150When I try to train FastestDet with different resolution, not default 352x352 (for example, 240x240), I got RuntimeError:
Load yaml sucess...
<utils.tool.LoadYaml object at 0x00000299198CD8B0>
Initialize params from:./module/shufflenetv2.pth
Traceback (most recent call last):
File "E:\Repositories\FastestDet\train.py", line 134, in <module>
model = FastestDet()
File "E:\Repositories\FastestDet\train.py", line 42, in __init__
summary(self.model, input_size=(3, self.cfg.input_height, self.cfg.input_width))
File "C:\Users\Reutov\.anaconda3\envs\experimental_env\lib\site-packages\torchsummary\torchsummary.py", line 72, in summary
model(*x)
File "C:\Users\Reutov\.anaconda3\envs\experimental_env\lib\site-packages\torch\nn\modules\module.py", line 1501, in _call_impl
return forward_call(*args, **kwargs)
File "E:\Repositories\FastestDet\module\detector.py", line 25, in forward
P = torch.cat((P1, P2, P3), dim=1)
RuntimeError: Sizes of tensors must match except in dimension 1. Expected size 15 but got size 16 for tensor number 2 in the list.
Shapes of P1, P2 and P3 (240х240): torch.Size([2, 48, 30, 30]) torch.Size([2, 96, 15, 15]) torch.Size([2, 192, 8, 8])
Shapes of P1, P2 and P3 (358x358): torch.Size([2, 48, 44, 44]) torch.Size([2, 96, 22, 22]) torch.Size([2, 192, 11, 11])
I can guess that I need to change the architecture of the backbone network a little bit, but pretrained weights will not to load correctly in this case
How can I to change input resolution for training FastestDet?
Can I to change input resolution of FastestDet in convert to ONNX process?
Thanks in advance :)
Thanks for your contribution!
The require.txt file is missing.
Please check and upload.
模型预测得到两个预测框,这两个框的iou接近于1.0.
这种情况不合乎常理吧?
是否可以修改后处理算法,使这两个框只保留一个?
Can you give me solution to calculate mAP, Precision and Recall?
@dog-qiuqiu Are the shufflenetv2.pth weights which are loaded by default in shufflenetv2.py from ImageNet training ?
Do you plan to train an architecture like yolov5 on multiple GPUs, which is more friendly to the use of large-scale data?
ncnn model load sucess...
output: 85, 22, 22
Time: 14.54 ms
x1:4 y1:167 x2:304 y2:268 person:95.12%
x1:262 y1:59 x2:658 y2:1046 bicycle:94.27%
x1:-29 y1:175 x2:134 y2:267 person:87.71%
x1:210 y1:144 x2:355 y2:239 person:78.87%
ncnn model load sucess...
input: 3, 352, 352
output: 0, 0, 0
Time: 6.87 ms
output 维度不见了，请问是为什么呢？
In line 89 of the file /utils/datasets.py, there is a statement: label_path = img_path.split(".")[0] + ".txt" which may cause some errors, such as using the format of './xxx/xxx. txt '.
To avoid such errors, label_path = os.path.splitext(img_path)[0] + ".txt" might be better?
感谢
有在相同平台测试过，onnxruntime和ncnn两个framework的速度差异吗
您好，请问这个实际检测的准确率很差，是我自己的问题嘛。大家有在端侧或虚拟机上尝试嘛？
我在试验训练环境的时候发现，无法保存模型权重
Sorry if I asked a stupid question. I thought it was a Spatial Pyramid Pooling (SPP) code but it looks different from what I know. Can someone enlighten me?
def __init__(self, input_channels, output_channels):
super(SPP, self).__init__()
self.Conv1x1 = Conv1x1(input_channels, output_channels)
self.S1 = nn.Sequential(nn.Conv2d(output_channels, output_channels, 5, 1, 2, groups = output_channels, bias = False),
nn.BatchNorm2d(output_channels),
nn.ReLU(inplace=True)
)
self.S2 = nn.Sequential(nn.Conv2d(output_channels, output_channels, 5, 1, 2, groups = output_channels, bias = False),
nn.BatchNorm2d(output_channels),
nn.ReLU(inplace=True),
nn.Conv2d(output_channels, output_channels, 5, 1, 2, groups = output_channels, bias = False),
nn.BatchNorm2d(output_channels),
nn.ReLU(inplace=True)
)
self.S3 = nn.Sequential(nn.Conv2d(output_channels, output_channels, 5, 1, 2, groups = output_channels, bias = False),
nn.BatchNorm2d(output_channels),
nn.ReLU(inplace=True),
nn.Conv2d(output_channels, output_channels, 5, 1, 2, groups = output_channels, bias = False),
nn.BatchNorm2d(output_channels),
nn.ReLU(inplace=True),
nn.Conv2d(output_channels, output_channels, 5, 1, 2, groups = output_channels, bias = False),
nn.BatchNorm2d(output_channels),
nn.ReLU(inplace=True)
)
self.output = nn.Sequential(nn.Conv2d(output_channels * 3, output_channels, 1, 1, 0, bias = False),
nn.BatchNorm2d(output_channels),
)
self.relu = nn.ReLU(inplace=True)
def forward(self, x):
x = self.Conv1x1(x)
y1 = self.S1(x)
y2 = self.S2(x)
y3 = self.S3(x)
y = torch.cat((y1, y2, y3), dim=1)
y = self.relu(x + self.output(y))
return y```
Hi I get this error with batchsize = 32 after 10 epoch
creating index...
index created!
Traceback (most recent call last):
File "train.py", line 134, in
model.train()
File "train.py", line 126, in train
mAP05 = self.evaluation.compute_map(self.val_dataloader, self.model)
File "C:\final\dataset\FastestDet-main\utils\evaluation.py", line 98, in compute_map
mAP05 = self.coco_evaluate(gts, pts)
File "C:\final\dataset\FastestDet-main\utils\evaluation.py", line 45, in coco_evaluate
coco_pred.dataset["images"].append({"id": i})
MemoryError
Layer (type) Output Shape Param #
use SGD optimizer
Starting training for 10 epochs...
Epoch:0 LR:0.000000 IOU:0.683203 Obj:0.118574 Cls:1.722579 Total:9.085380: 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:03<00:00, 3.00s/it]
Epoch:1 LR:0.000002 IOU:0.682660 Obj:0.117858 Cls:1.543631 Total:8.890644: 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:01<00:00, 1.23s/it]
Epoch:2 LR:0.000026 IOU:0.642826 Obj:0.115852 Cls:1.679555 Total:8.675802: 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:01<00:00, 1.13s/it]
Epoch:3 LR:0.000130 IOU:0.680697 Obj:0.117061 Cls:1.506619 Total:8.825181: 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:01<00:00, 1.28s/it]
Epoch:4 LR:0.000410 IOU:0.693652 Obj:0.116068 Cls:1.497365 Total:8.903667: 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:01<00:00, 1.25s/it]
Epoch:5 LR:0.001000 IOU:0.667514 Obj:0.118184 Cls:1.399224 Total:8.630288: 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:01<00:00, 1.17s/it]
Epoch:6 LR:0.001000 IOU:0.641047 Obj:0.113830 Cls:1.065889 Total:8.015538: 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:01<00:00, 1.14s/it]
Epoch:7 LR:0.001000 IOU:0.598921 Obj:0.112864 Cls:0.718651 Total:7.315843: 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:01<00:00, 1.23s/it]
Epoch:8 LR:0.001000 IOU:0.596256 Obj:0.111394 Cls:0.472929 Total:7.025274: 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:01<00:00, 1.25s/it]
Epoch:9 LR:0.001000 IOU:0.594581 Obj:0.108997 Cls:0.344135 Total:6.844735: 100%|████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:01<00:00, 1.38s/it]
Epoch:10 LR:0.001000 IOU:0.586639 Obj:0.108114 Cls:0.251320 Total:6.674245: 100%|███████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:01<00:00, 1.29s/it]
computer mAP...
0%| | 0/1 [00:00<?, ?it/s]D
:\python\anaconda\envs\fast38\lib\site-packages\torch\functional.py:568: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at
C:\actions-runner_work\pytorch\pytorch\builder\windows\pytorch\aten\src\ATen\native\TensorShape.cpp:2228.)
return _VF.meshgrid(tensors, **kwargs) # type: ignore[attr-defined]
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00, 2.46s/it]
creating index...
index created!
creating index...
index created!
Running per image evaluation...
Evaluate annotation type bbox
DONE (t=0.03s).
Accumulating evaluation results...
DONE (t=0.01s).
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000
ModuleNotFoundError: No module named 'pycocotools._mask'
能否像yoloV5那样,提供预训练模型(加快训练速度)?
train.py导出路径有个冒号，windows训练了半天，一个都没保存下来.
pth文件命名带了冒号，windows不太好使，建议改一下
你好，请问代码中提到的那些配置文件需要在哪里下载呢
按照说明，转出 onnx的模型，在 https://netron.app/导入查看到，输出是 float32[1,85,22,22]
但看到命令执行得到的结果形如
tensor([0.0814, 0.1866, 0.3560, 0.8782, 0.9012, 0.0000])
tensor([0.5778, 0.4236, 0.9649, 0.8865, 0.8948, 0.0000])
tensor([0.3183, 0.4065, 0.5119, 0.7926, 0.8064, 0.0000])
tensor([0.4625, 0.4142, 0.6451, 0.8280, 0.7805, 0.0000])
这里是如何转化的呢？
由于项目需要单线程运行，对于YOLO-fastestv2，在amd平台的虚拟机和rk3588上单线程设置都是正常的，查看cpu占用一般只有一到两个核心高。但是配置在Fastestdet的用同样设置上对于虚拟机平台，1-12cpu都在30-40%，而对于rk3588，1-8cpu都在70%左右。
上述都是设置的单线程。使用
ncnn::Net net;
net.opt.num_threads = 1; //此处配置的单线程和下面的ex.set_num_threads(1);保持一致，对于yolofastestv2也是同样的设置
net.load_param("FastestDet.param");
net.load_model("FastestDet.bin");
但是实际上则不像fastestv2的占用，特别是rk3588上的测试，设置单线程，但8核都在70%左右。
运行ncnn例子，无任何修改，出现这种情况，哪位大佬知道是什么问题
大神您好，
我这边训练了一个自己的3类的检测模型，直接评测的情况尚可，但是转换ncnn模型并且用大神给的example去进行推理，发现检测结果差别很大，不知道这个是什么原因呢
Hi,
developer,
I have noticed that the weight_AP05:0.253207_280-epoch.pth in weights folder has been changed to weight_AP05_0.253207_280-epoch.pth
Would you like to update the READEME.md colon's issue?
python3 test.py --yaml configs/coco.yaml --weight weights/weight_AP05:0.253207_280-epoch.pth --img data/3.jpg
发现当label中大小目标差异比较大的时候，用f = iou > iou.mean()过滤后，很多小目标的label完全过滤掉，无法参加训练。比如我的label分别为0,1. 0是大的label，1是小的label。过滤后，只有0的label参加训练。1对应的被忽略
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
Django
The Web framework for perfectionists with deadlines.
Laravel
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
Microsoft
Open source projects and samples from Microsoft.
Google
Google ❤️ Open Source for everyone.
Alibaba
Alibaba Open Source for everyone
D3
Data-Driven Documents codes.
Tencent
China tencent open source team.