Hello,how to use Yolo-Fastest( weight and cfg of COCO and VOC)?Thanks for your reply.

Is these speed testing on Kirin 990 CPU or NPU，and could you please give the benchmark testing code demo？

Amazing work!
I test the model of yolo-fastest-xl on a jetson nano, the FPS is more than 10 which is quite good.
I wonder if the TensorRT is introduced according to jkjung's work(https://github.com/jkjung-avt/tensorrt_demos), which could we achieve a real-time detection on edge devices.
Many thanks!

您好，我是初学者。我根据Alexnet做了一个分类网络，然后如何能想您的yolofast一样进行ncnn加速？大概需要做的流程是什么？期待您的回答，谢谢

Hi,

Really interesting work. I was wondering where can I find a simple architecture definition for Yolo-Fastest or Yolo-Fastest-XL? I am asking because I wanted to port it to TensorFlow to compare it with my own work.

Also, if you can provide more details on the training method, like which data augmentations, loss functions etc. you use, it would also be an interesting study.

Even if you are not releasing a paper, a simple doc with such details would be really helpful for someone following up on your work. For example, what exactly is your backbone? Where do you cut it off and what feature aggregation method did you use? Also how many object detection heads does your model have? These details would be really helpful.

Thank You

Can it do INT8 quantization with MNN or NCNN?
I tried to quantify it, the conversion did not report any errors, but I didn't get any predicted results.

I am trying to convert your darknet model to pytorch, but I find that the size of your weight is mismatch with my model, here is my convert code:

def load_darknet_weights(self, weights_path):
     """Parses and loads the weights stored in 'weights_path'"""

     # Open the weights file
     with open(weights_path, "rb") as f:
         header = np.fromfile(f, dtype=np.int32, count=5)  # First five are header values
         print(header)
         self.header_info = header  # Needed to write header when saving weights
         self.seen = header[3]  # number of images seen during training
         weights = np.fromfile(f, dtype=np.float32)  # The rest are weights

     # Establish cutoff for loading backbone weights
     cutoff = None
     if "darknet53.conv.74" in weights_path:
         cutoff = 75

     ptr = 0
     for i, (module_def, module) in enumerate(zip(self.module_defs, self.module_list)):
         if i == cutoff:
             break
         if module_def["type"] in ["convolutional","normal_convolutional"]:
             conv_layer = module[0]
             if int(module_def["batch_normalize"])==1:
                 # Load BN bias, weights, running mean and running variance
                 print(module[0].state_dict()["weight"].shape)
                 bn_layer = module[1]
                 # print(bn_layer.state_dict())
                 num_b = bn_layer.bias.numel()  # Number of biases
                 print(num_b)
                 # Bias
                 bn_b = torch.from_numpy(weights[ptr : ptr + num_b]).view_as(bn_layer.bias)
                 bn_layer.bias.data.copy_(bn_b)
                 ptr += num_b
                 # Weight
                 bn_w = torch.from_numpy(weights[ptr : ptr + num_b]).view_as(bn_layer.weight)
                 bn_layer.weight.data.copy_(bn_w)
                 ptr += num_b
                 # Running Mean
                 bn_rm = torch.from_numpy(weights[ptr : ptr + num_b]).view_as(bn_layer.running_mean)
                 bn_layer.running_mean.data.copy_(bn_rm)
                 ptr += num_b
                 # Running Var
                 bn_rv = torch.from_numpy(weights[ptr : ptr + num_b]).view_as(bn_layer.running_var)
                 bn_layer.running_var.data.copy_(bn_rv)
                 ptr += num_b
                 # Load conv. weights
                 num_w = conv_layer.weight.numel()
                 print(num_w)
                 conv_w = torch.from_numpy(weights[ptr: ptr + num_w]).view_as(conv_layer.weight)
                 conv_layer.weight.data.copy_(conv_w)
                 ptr += num_w
             else:
                 # 对于yolov3.weights,不带bn的卷积层就是YOLO前的卷积层
                 if ".weights" in weights_path:
                     num_b = 255
                     ptr += num_b
                     num_w = int(self.module_defs[i-1]["filters"]) * 255
                     ptr += num_w
                 else:
                     # Load conv. bias
                     num_b = conv_layer.bias.numel()
                     conv_b = torch.from_numpy(weights[ptr : ptr + num_b]).view_as(conv_layer.bias)
                     conv_layer.bias.data.copy_(conv_b)
                     ptr += num_b
                     # Load conv. weights
                     num_w = conv_layer.weight.numel()
                     conv_w = torch.from_numpy(weights[ptr : ptr + num_w]).view_as(conv_layer.weight)
                     conv_layer.weight.data.copy_(conv_w)
                     ptr += num_w

I guess My convolution layer is different with yours, Here is my convolution layer:

elif module_def["type"] == "normal_convolutional":
            # print("==============input filtrs:",output_filters[-1],"=================")
            # print('convolution')
            bn = int(module_def["batch_normalize"])
            filters = int(module_def["filters"])
            kernel_size = int(module_def["size"])
            stride = int(module_def['stride'])
            pad = int(module_def['pad'])
            groups=int(module_def['groups'])
            modules.add_module(
                f"conv_3{module_i}",
                nn.Conv2d(
                    in_channels=output_filters[-1],
                    out_channels=filters,
                    kernel_size=kernel_size,
                    stride=stride,
                    padding=pad,
                    bias=not bn,
                    groups=groups
                    ),
                )
            # modules.add_module(f"batch_norm_{module_i}", nn.BatchNorm2d(filters, momentum=0.9, eps=1e-5))
            if bn:
                modules.add_module(f"batch_norm_1{module_i}", nn.BatchNorm2d(filters))
            if module_def["activation"] == "relu6":
                modules.add_module(f"leaky_{module_i}", nn.ReLU6(inplace=True))

I don't know where is different, if your know, could you finger out? thanks a lot.

Good job！it‘s really fast. and i wanna know why you use leaky relu but not relu?

Hi,how to solve?

Sincerely suggest that you could add “\” before the “`” in ncnn_sample's README.md, so that it may be easier for new comers to get started.

I already know that, Thank you very much.

hello, when i prepare training with a generated pretrained model, and follow darknet yolov3 configure. the total loss is large big because of class_loss is very high ,even to 340+. how can i sovle this problem? ths

hello I want to know where I can find the yolo-fastest.conv.109

To tranform caffe failed！ How can we solute the question？

Hi, i'm converting your models to TensorRT model to use in NVIDIA DeepStream SDK. I already ported MobileNetV2-YOLOv3-Nano and MobileNetV2-YOLOv3-Lite.

The Yolo-Fastest and MobileNetV2-YOLO-Fastest are the same model?

I need to know, to create parse functions.

Thanks.

How can one use this repos with a custom widthxheight? what should I change in config, width, height anchors, train from scratch??

您好，请教一下，我是在alexeyAB的darknet下训练的，会报上面的错误，是要更新cudnn吗？
因为修改环境不太方便，所以我试着将groups都改为=1和将所有的groups行删掉，然后用一个数据集进行训练，跑一小会就nan了。。。请问有什么解决办法吗？

老哥 你两个先验框size是多少哇

我是用预训练模型在训练了10000批次后，发现模型对区域的检测基本都能检测到，但检测归类的class还是预训练模型里的“car,bus.plan”之类的东西，而不是我自定义的分类。我在各种**.names都定义了我自己的分类，voc.data里也指定了name文件，但还是无效，这和我的训练次数或样本数量有关吗？还是使用预训练模型的问题呢？

Hello, I am on 1660TiGPU in ubuntu16.04, yolofastest speed is 30 frames, and about 1 frame on Raspberry Pi. Is this normal?Looking forward to your reply, thank you.

In the Ros system, the release topic of the camera is received, then the NCNNSAMPLE is used for testing and the release topic is the test result.

Hi,
The readme file is not very clear for the demo.
1-what is the difference between darknet_images.py vs image_yolov3.sh and the same for video? Just two ways to run the demo or these are different for comparison?

2- How to compare the speed of yolov3 or v4 with yolo-fastest?

hey,man,thanks for the work,it's cool. When i convert the model to 8bit by ncnn tool,the accuracy has dropped a lot,can you give me some suggetions?

When I convert the weights file to other model files, it always report error: buffer is too small for requested array. I check the .cfg file,then I found there some 'groups' vars in [convolutional] layer, maybe it reads more weights than GroupConv2D needs and make the weights not enough to use, but I don't know how to split groups in yolo-fastest(filters=groups?).
Is the error caused by this problem, or others, and how to solve this error?

What about
Darknet weights (.weights) to TensorFlow weights (.ckpt)
Darknet model (.cfg) to TensorFlow graph (.pb, .meta)

Hello qiuqiu,

Is my understanding below right ?
The commands for getting the pre-trained weights for yolo-fastest-xl and training yolo-fastest-xl are :

./darknet partial yolo-fastest-xl.cfg yolo-fastest-xl.weights yolo-fastest-xl.conv.109 109
./darknet detector train voc.data yolo-fastest-xl.cfg yolo-fastest-xl.conv.109

Thanks.

I trained model using NVIDIA Tesla P100, My cpu is intel i7 1.8G, Why infer 320*320 needing about 30ms, where is wrong?
Kirin 990 CPU’s time only is 6.74ms!! where is wrong?
Thank you very much!

我用NVIDIA Tesla P100训练的模型，在intel i7 1.8G cpu上预测320*320的时间是30ms左右，但是根据官网，Kirin 990 CPU预测同样大小是6.74ms。
这个差距太大了，是哪里出了问题？非常感谢！

I follow the README file to train my dataset. but the total loss is still very large(>=350)at 1000 epochs and without ant decrease tread,HOW CAN I SOLVE THIS PROBLEM?

Hi,good job you have done.I want to read about the artical about the Yolo-Fastest,but I can not find where it is.

How could we calculate anchors for this model?

the code of https://github.com/CaoWGG/TensorRT-YOLOv4 can`t converts the yolo-fastest to onnx model.
typeerror: buffer is too small for requested array

编译时报错，nvcc fatal : Unsupported gpu architecture 'compute_61'，显卡是笔记本rtx 2060 ，请问怎么解决？
when i compile it ，an error occurred，“nvcc fatal : Unsupported gpu architecture 'compute_61'”，my GPU is laptop edition RTX 2060 ，who can help me？

这个分支和yolov4那个分支有啥区别，也可以用来训练v4,v4-tiny吧

I followed your README.md, modified Makefile to open CUDA CUDNN and GPU, and then run "make"
I used ./Yolo-Fastest/VOC/yolo-fastest.weights to run image_yolov3.sh, it costs 29ms on 1080TI GPU.
I do not know where is wrong
looking forward to your reply

How to run a demo, please? It seems not so clear in readme.md.

how can I change *.cfg with EfficientNet-lite1~4

I tested yolo-fastest on my PC, which run quite slowly with almost 25ms per frame on average. The XL version also takes even much more consuming time with 45ms. BTW, the version of my GPU is 2070super. Did I test it not correctly? Need help.

@dog-qiuqiu Hi Sir, Can you please update the COCO mAP for the Yolo-Fastest?
https://github.com/dog-qiuqiu/Yolo-Fastest#evaluating-indicatorbenchmark

Seems like it's missing, or maybe you are still working on it.
Thanks

I used CUDA 10.2，cudnn 7.6.5，GPU NVIDIA GeForce RTX 2060

I have set up in Makefile
GPU=1
CUDNN=1
ifeq ($(GPU), 1)
COMMON+= -DGPU -I/usr/local/cuda-10.2/include/
CFLAGS+= -DGPU
LDFLAGS+= -L/usr/local/cuda-10.2/lib64 -lcuda -lcudart -lcublas -lcurand

C:\Yolo-Fastest\build\darknet\x64> ./darknet.exe detector test ./cfg/voc.data ./cfg/yolo-fastest.cfg ./cfg/yolo-fastest.weights ./data/person.jpg -i 1 -thresh 0.25 -out_filename ./data/person_output.jpg

CUDA status Error: file: c:\yolo-fastest\src\dark_cuda.c : cuda_set_device() : line: 39 : build time: Nov 24 2020 - 18:36:35
CUDA Error: invalid device ordinal

Can anyone help to fix this?

自己训练的Yolo-Fast转换成.param， 一旦执行load_param的时候就会崩溃，怎么办?

Can I use this project（Android_NCNN_yolov4-tiny） to deploy？

I have tried : https://github.com/Linzmin1927/DW2TF

the error:

  File "/home/lebhoryi/RT-Thread/Detection_API/Yolo-Fastest/DW2TF/util/reader.py", line 37, in walk
    assert end_point <= self.size, 'Over-read {}'.format(self.path)
AssertionError: Over-read ../Yolo-Fastest/COCO/yolo-fastest.weights

how could I do it

感谢您的技术分享！！
请问预训练weights yolo-fastest.conv.109要到哪里下载呢？
多谢！

Hi, qiuqiu,

Awesome!
Very you for the great project sharing!

I am facing a little trouble.

I want to replace yolov3-spp model in Alphapose.
however that convert by pytorch like this

elif  module_type == "shortcut":
    from_ = int(modules[i]["from"])
    x = outputs[i-1] + outputs[i+from_]
    outputs[i] = x             

elif module_type == 'yolo':        
                
anchors = self.module_list[i][0].anchors
#Get the input dimensions
inp_dim = int (self.net_info["height"])

#Get the number of classes
num_classes = int (modules[i]["classes"])

#Output the result
x = x.data.to(args.device)
x = predict_transform(x, inp_dim, anchors, num_classes, args)

that support layers like shortcut , yolo ...etc

but this project added dropout layer.
anyone have any idea to convert that?

please help, thanks!

您好！

如果我想训练自己的数据集，需要如何更改.cfg文件？我看这个项目中的.cfg文件格式和darknet yolov4的还是不太一样，比如我的classes数量和预训练的模型不一样，我需要按什么规则更改呢？

您好，我想在RGB-D数据上训练yolo,请问要怎么去配置呢？

Thank you for releasing Yolo-Fastest! It is indeed the fastest model I have used on Raspberry pi.
However, my running time is about 100ms at 2 threads and more than 100ms at 4 threads. I checked the CPU usage(60% percent at most) and memory usage(below 30%), both of which are low.
I wonder what else I can do with the Raspberry pi 4 to reach 33ms with yolo-fastest.
Thanks a lot!