At the begining there are some WARNs
[WARN] Cannot find rule for <class 'torch.nn.modules.container.Sequential'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'backbone.darknet.Conv_BN_LeakyReLU'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'backbone.darknet.DarkNet_19'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'utils.modules.Conv'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'utils.modules.reorg_layer'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'models.yolov2_d19.YOLOv2D19'>. Treat it as zero Macs and zero Params.

And it will stop at 3rd epoch and raise error when I train:
Traceback (most recent call last):
File "train.py", line 474, in
train()
File "train.py", line 432, in train
dist.barrier()
File "/home/ggy/anaconda3/lib/python3.8/site-packages/torch/distributed/distributed_c10d.py", line 1959, in barrier
_check_default_pg()
File "/home/ggy/anaconda3/lib/python3.8/site-packages/torch/distributed/distributed_c10d.py", line 210, in _check_default_pg
assert _default_pg is not None,
AssertionError: Default process group is not initialized

That's my param:
python train.py -d coco --cuda -v yolov2_d19 -hr -ms --ema

How can I solve this problem?
Thanks!

As far as I know, a convolutional layer should not have bias when a Batch Normalization applied after it. I think it is a serious issue according to the theory of Batch Normalization. I would appreciate if you learn more about Batch Normalization and fix the issue.
Thus you can add a bias=False as a kwarg of torch.nn.Conv2d in utils.modules.Conv2d module, then retrain the models.

您好!请问提供预训练模型下载吗?

Can you please share the drive link of weights mentioned in the repo?

This is my mistake, I found that I set trainable to False, Sorry!

"txty_loss_function = nn.BCEWithLogitsLoss(reduction='none')" doesn't match the original paper?
can you explain it.

thanks!!

How can I download google drive or other link?

Hey friend. I just want to use the COCO dataset to train a model that is used to detect people. That's means I only have one class to predict. But I have a question about the value of the "num_classes". Does the "num_classes" include "background"? 1 or 2 should I write?

ins = np.where(iou <= self.nms_threshold)[0]
orders = orders[ins+1]

这里是否要修改为：
ins = np.where(iou <= self.nms_threshold)
orders = orders[ins]

Hi, thanks for your nice job!
I want to apply your code on other dataset (such as fire detection), however, in this case, num of case is only 1 (20 for voc and 80 for coco). So, should I remove the loss part for classification(because there exists loss for object confidence) , or just change number of class to 1?

Is the default value of args.* your hyper parameters for training?

env

• GPU 2080ti
• ubuntu 16.04
• python 3.6
• pytorch 1.5.0

errors

when i use yolo-v2 to train the voc dataset, and set the parameters as follows:

parser.add_argument('-v', '--version', default='yolo_v2',
                        help='yolo_v2, yolo_v3, slim_yolo_v2, tiny_yolo_v3')

    parser.add_argument('-d', '--dataset', default='VOC',
                        help='VOC or COCO dataset')

    parser.add_argument('-hr', '--high_resolution', action='store_true', default=True,
                        help='use high resolution to pretrain.')

    parser.add_argument('-ms', '--multi_scale', action='store_true', default=True,
                        help='use multi-scale trick')

    parser.add_argument('--batch_size', default=32, type=int,
                        help='Batch size for training')

    parser.add_argument('--lr', default=1e-3, type=float,
                        help='initial learning rate')

    parser.add_argument('-cos', '--cos', action='store_true', default=False,
                        help='use cos lr')

    parser.add_argument('-no_wp', '--no_warm_up', action='store_true', default=False,
                        help='yes or no to choose using warmup strategy to train')

    parser.add_argument('--wp_epoch', type=int, default=2,
                        help='The upper bound of warm-up')

    parser.add_argument('--dataset_root', default="/home/xxx/tmp/tmp/yolo_v1_v2/data/VOCdevkit",
                        help='Location of VOC root directory')

    parser.add_argument('--num_classes', default=20, type=int,
                        help='The number of dataset classes')

    parser.add_argument('--momentum', default=0.9, type=float,
                        help='Momentum value for optim')

    parser.add_argument('--weight_decay', default=5e-4, type=float,
                        help='Weight decay for SGD')

    parser.add_argument('--gamma', default=0.1, type=float,
                        help='Gamma update for SGD')

    parser.add_argument('--num_workers', default=8, type=int,
                        help='Number of workers used in dataloading')

    parser.add_argument('--cuda', action='store_true', default=True,
                        help='use cuda.')

    parser.add_argument('--save_folder', default='weights/voc/', type=str,
                        help='Gamma update for SGD')

    parser.add_argument('--tfboard', action='store_true', default=False,
                        help='use tensorboard')

    parser.add_argument('--resume', type=str, default=None,
                        help='fine tune the model trained on MSCOCO.')

An error occured while running the code, it looks like mismatch the size in somewhere, but when i set the multi-scale False, the code is working well. So i think there must be something wrong with multi-scale.

[Epoch 1/250][Iter 0/517][lr 0.000000][Loss: obj 433.45 || cls 8.17 || bbox 19.32 || total 460.94 || size 608 || time: 13.28]
[Epoch 1/250][Iter 10/517][lr 0.000000][Loss: obj 431.94 || cls 7.31 || bbox 13.79 || total 453.04 || size 608 || time: 9.52]
Traceback (most recent call last):
  File "/home/xxx/tmp/tmp/yolo_v1_v2/train_voc.py", line 294, in <module>
    train()
  File "/home/xxx/tmp/tmp/yolo_v1_v2/train_voc.py", line 245, in train
    conf_loss, cls_loss, txtytwth_loss, total_loss = model(images, target=targets)
  File "/home/xxx/anaconda3/envs/pytorch1.5.0/lib/python3.7/site-packages/torch/nn/modules/module.py", line 550, in __call__
    result = self.forward(*input, **kwargs)
  File "/home/xxx/tmp/tmp/yolo_v1_v2/models/yolo_v2.py", line 217, in forward
    x1y1x2y2_pred = (self.decode_boxes(txtytwth_pred) / self.scale_torch).view(-1, 4)
  File "/home/xxx/tmp/tmp/yolo_v1_v2/models/yolo_v2.py", line 90, in decode_boxes
    xywh_pred = self.decode_xywh(txtytwth_pred)
  File "/home/xxx/tmp/tmp/yolo_v1_v2/models/yolo_v2.py", line 74, in decode_xywh
    xy_pred = torch.sigmoid(txtytwth_pred[:, :, :, :2]) + self.grid_cell
RuntimeError: The size of tensor a (361) must match the size of tensor b (100) at non-singleton dimension 1

https://github.com/yjh0410/pytorch-yolo-v2-v3/blob/e18c5671dcfac59038717b262e2729f9f0a05fcb/tools.py#L8

Have you ever trained the model with different IGNORE_THRESH like 0.7?

I know original YOLOv3 used 0.7.

https://github.com/developer0hye/YOLOv3-PyTorch-Colab

https://colab.research.google.com/drive/1ikcj4kJ2N1ny3jIHuEFIZTnwlV9-nCw4

I am experimenting with the variant of your code.

I got the higher mAP with normalized wh prediction instead of wh divided by the stride.

The results of 170 epoch trained yolv3 tiny

AP for aeroplane = 0.7843
AP for bicycle = 0.8139
AP for bird = 0.6702
AP for boat = 0.6116
AP for bottle = 0.5703
AP for bus = 0.8028
AP for car = 0.8394
AP for cat = 0.8012
AP for chair = 0.5465
AP for cow = 0.7497
AP for diningtable = 0.6677
AP for dog = 0.7366
AP for horse = 0.8391
AP for motorbike = 0.8023
AP for person = 0.7880
AP for pottedplant = 0.4613
AP for sheep = 0.7257
AP for sofa = 0.6991
AP for train = 0.8025
AP for tvmonitor = 0.7194
Mean AP = 0.7216

Hello, I would like to ask you a question:
In the multi_gt_creator function of the tools.py file, why does box_w only divide by strides [-1] before calculating iou, does this not always cause iou to be calculated on the minimum or maximum downsampling multiple? Instead of calculating at all downsampling multiples.

box_ws = box_w / strides[-1]
box_hs = box_h / strides[-1]

请问下载了预训练集和代码，yolo2 和yolo3的AP都达不到，大概只有30~40那个样子？是什么情况呢

Have you trained and tested your model using Sigmoid function instead of Softmax function?

In the YOLOv3 paper, the author mentioned that sigmoid function is used to predict multi-class.
Actually, I compared the mAP of 50 epoch trained 2 models with sigmoid and softmax and I observed that the model with softmax has higher mAP. But, I think 50 epochs trained model is not sufficient to compare the performance.

Hello @yjh0410,

Thanks for the share of your work! I am training the yolov2 darknet19 model on my machine, and have realized 2 problems:

1. After some time, training stops although the training process is running. I realized this only after forwarding the stdouts to be logged into a txt file, and had to kill and resume training 5 times within 120 epochs to be able to continue.
2. When the training pid is killed, the memory of the gpus didn't get flushed out. When I do sudo fuser -v /dev/nvidia*, I saw that 8 more processes were created along with the parent process shown on nvtop/nvidia-smi. Do you know why 9 instances of the same process are created when the code isn't multithreaded?
   

I'm running on ubuntu 18.04 lts, python 3.7.6, torch 1.7.0 and using GeForce GTX 1080 Ti.

According to generate_txtytwth(), there's no corresponding operation of sigmoid.

Hello @yjh0410,
Thanks for the share of your work! I am training the yolov2slim model on my machine, and meet a problem:when I use 3 GPUs to train the yolov2slim model on COCO2017 dataset. When I log in to the server again after running the code in the background, the connection fails.Here are my circumstances and instructions :
python 3.6.0 torch 1.6.0 torchvision 0.7.0
nohup python -m torch.distributed.launch --nproc_per_node=3 train.py
-dist
--sybn
--num_gpu 3
--batch_size 96 &

Hi, yjh!
It's been a while.

Have you ever measured the mAP of yolov3-tiny on VOC?

My environment

GPU: gtx1080ti
Cuda: 9.2
Cudnn: not download

The program crashes during training sometimes, have you ever experienced the crash?

Hi, yjh!

https://github.com/yjh0410/pytorch-yolo-v2/blob/baecdae27e4b75883088729dbd60a8f2a1270c80/tools.py#L317-L318

Why do you divide width and height of object by largest stride size?

I also saw this policy in other yolov3 implementations, but I couldn't understand that reason.
Can you expect that we can predict the normalized width and height with normalized anchor box?

When I trained YOLOV3 at Batch Size = 4, ‘nan’ appeared. After Debug, I found the infinite number appears at models/yolov3.py 118th code,
This is because the output of the network txtytwth_pred [:::,:, 2:] some values are too large, resulting in an infinite exp value.

Therefore, this is not the author's implementation problem, it is the reason for the unstable training, so I reduced the learning rate to 1e-4, and the training was normal.
Hope can help to other friend.

请问为什么tools.py里面
txty_loss_function采用的是BCEWithLogitsLoss，而conf_loss_function采用的是MSEWithLogitsLoss，两者是否弄反了？

First of all, thank you for your work and open source，The problem I met before was caused by a mistake of mine. But I still want to make suggestions to you, you can post the performance of the infer speed

Thanks for sharing your great works.

I tried to download your ImageNet pretrained weight file with baidu. But I failed to it.

Could you share pre-trained weight with Google drive?

Thanks for sharing great works!

I have one question.

What is your input image normalization policy used in pretrained backbone network?

Did you put the image into model without normalization?

Is the range [0, 255]? or you followed the normalization method proposed in VGG(Z-normalization using mean and std)?

It would be great if you could add a license so we know how we may use the code! Thank you!

Hi, I met the same problem in yjh0410/PyTorch_YOLOv3#1 when I train yolov3 in this project.

Hi!

Could you upload the pretrained darknet-light model?

https://github.com/yjh0410/pytorch-yolo-v2-v3/blob/550661313e36d86bdfa2e4d5dc2a3b5ece411e79/utils/augmentations.py#L350-L373

I think, in the real application, there is no case that process swapped axis image. So, I think this augmentation may harm the model's performance.

What do you think of this augmentation?

self.backbone = darknet19(pretrained=trainable, hr=hr)

I met this warn when I type the commad python train.py -v yolov2_r50 in the terminate

[WARN] Cannot find rule for <class 'torch.nn.modules.container.Sequential'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'backbone.resnet.Bottleneck'>. Treat it as zero Macs and zero Params.
[WARN] Cannot find rule for <class 'backbone.resnet.ResNet'>. Treat it as zero Macs and zero Params.

I just can't figure out the reason why this warn occured and hope you can help me
thank you!

4900 images and batchsize：64 gpu：2080TI

https://github.com/yjh0410/pytorch-yolo-v2-v3/blob/e18c5671dcfac59038717b262e2729f9f0a05fcb/backbone/darknet.py#L161-L227

Why do you replace the MaxPooling with strided convolution??

I think that convolution has more complexity than MaxPooling.

4 anchor box's size is hard coded in your code. I wonder how you get them.

Hi! It's been a long time.

Did you compare the performance with or without normalization?