hi, I don't see the Mutil-scale training.If I go from [416, 416] scale to [512, 512] scale model, will I reset model after saving?
By loading the trained model of [416, 416], the img_size is set to [512, 512] for retraining. Is it possible to do multi-scale training?
Because I understand that multi-scale training has been set up several scales [320, 416, 512, 640], in the training process each batch randomly selected a size.
Thank you in advance for your assistance。@wizyoung

I want to get the average time cost of my test set. How do I do that in code?

I noticed that you mention in README.md that

NOTE: The yolo anchors should be scaled to the rescaled new image size. Suppose your image size is [W, H], and the image will be rescale to 416*416 as input, for each generated anchor [anchor_w, anchor_h], you should apply the transformation anchor_w = anchor_w / W * 416, anchor_h = anchor_g / H * 416.

however, the size of the images in my dataset are varied, do I need to calculate the mean H and W of my images and apply the above transformation?

博主你好，我尝试使用ｐｏｓｃａｌ　ｖｏｃ２００７数据进行训练，出现一下错误，请帮忙看下。谢谢

ｔｒain.txt, val.txt,　已经按照数据格式替换成如下的格式：
/.../JPEGImages/002611.jpg 14 1 397 1 351 11 152 337 131 298
/.../JPEGImages/002332.jpg 14 121 254 39 231 14 337 494 61 182 14 101 500 81 362
/.../JPEGImages/005445.jpg 7 32 353 233 497
之前bbox中坐标使用相对于ｗｉｄth/ｈｅｉght的ｓｃａｌｅ值，可以进行训练，到了３个ｉｔｒｅ后就全部是ＮＡＮ。
估计是值太小，出现以下报错，仅仅是更换了ｂｂｏｘ中的xmin,ymin,xmax,ymax的值的大小。有时间的话帮忙看下哈

.names文件替换成了ｖｏｃ的２０个类，

报错信息：
File "/home/jiapy/workspace/yolo-v3/YOLOv3_TensorFlow/utils/data_utils.py", line 115, in process_box
y_true[feature_map_group][y, x, k, 0:2] = box_centers[i]
IndexError: index 59 is out of bounds for axis 0 with size 52

When I train my own data by your model , I meet this problem as flow, and I wolud give your my data , thank you！！！

train.txt

File "D:\Project_Data\YOLOv3_TensorFlow\utils\data_utils.py", line 122, in process_box
y_true[feature_map_group][y, x, k, 5+c] = 1.

IndexError: index 10 is out of bounds for axis 3 with size 10

I am training in my own dataset.Follow the documentation exactly.
The last time the output was normal, the next time it was nan.
Help,please.

data_utils.py文件中生成y_true的时候

for i, idx in enumerate(best_match_idx):

idx: 0,1,2 ==> 2; 3,4,5 ==> 1; 6,7,8 ==> 2

    feature_map_group = 2 - idx // 3

这里是不是有问题，按照anchors_mask = [[6,7,8], [3,4,5], [0,1,2]]，
idx对应关系应该是 0,1,2 ==> 2; 3,4,5 ==> 1; 6,7,8 ==> 0，
与feature_map_group的计算公式对应不上。

Thanks for your meticulous works. Have you trained from scratch using your codes here and what is the performance？ I find there are many differences with original keras codes

Traceback (most recent call last):
Epoch: 1, global_step: 200 | loss: total: nan, xy: 11.83, wh: 1146.74, conf: nan, class: nan | Last batch: rec: 0.000, prec: 0.000 | lr: 0.0004329
File "E:/net/new/YOLOv3_TensorFlow-master/train.py", line 160, in
'Gradient exploded! Please train again and you may need modify some parameters.')

ArithmeticError: Gradient exploded! Please train again and you may need modify some parameters.

如图，训练到中间还会出现nan,什么情况。。

could u tell me how to use the gpu

when I run the func evaluate_on_gpu, why it is soooooooooo slow???

In the train.txt, it is the path of jpg and bbox, so where is the training image?

I am training with custom data which looks like this.
img size is 720*1280
my training data looks like this

Training is not starting and
i am getting an error like this

OutOfRangeError: End of sequence
[[node IteratorGetNext (defined at C:/Users/madhu/Desktop/YOLOv3_TensorFlow-master/train.py:160) = IteratorGetNextoutput_shapes=[, , , ], output_types=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT], _device="/job:localhost/replica:0/task:0/device:CPU:0"]]
[[{{node IteratorGetNext/_535}} = _Recvclient_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device_incarnation=1, tensor_name="edge_671_IteratorGetNext", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:GPU:0"]]

please help me resolve this issue

笔者你好，能系统介绍一些复现神经网络的技巧吗，看你复现的这份代码写得很漂亮，我自学这方面貌似很困难，常常遇到要写一个功能时却不知道怎样去寻找相关需要调用的函数有没有、在哪儿、叫什么名字，你是如何解决这些问题的呢，不能先把各个库的API先全部记住吧，感谢回复

Traceback (most recent call last):
File "test_single_image.py", line 52, in
saver.restore(sess, args.restore_path)
File "C:\Program Files\Python35\lib\site-packages\tensorflow\python\training\saver.py", line 1538, in restore
+ compat.as_text(save_path))
ValueError: The passed save_path is not a valid checkpoint: ./data/darknet_weights/yolov3.ckpt

笔者你好,我用您的模型训练4类,人 ,车,交通灯,自行车.
我重新实现了您的代码.网络也能跑起来! 在此十分感谢

前期跑很正常.loss都下降.但是不知何时我的loss就变NaN了.请问这是什么问题呢?

我没有做自己的数据增强.
谢谢啦.

Does anyone try to train on VOC 2007+2012 trainval?

I got only 0.742 mAP on VOC 2007 test

recall: 0.703, precision: 0.793, total_loss: 0.440, loss_xy: 0.017, loss_wh: 0.007, loss_conf: 0.376, loss_class: 0.040

I apply the 2-stage training:

1. I load the darknet weight first and only train the yolov3_head.
2. When the loss < 2, I lower the learning rate and train the whole model

hello wizyoung， when i run the python convert_weights.py ， it happend following errors：
Traceback (most recent call last): File "convert_weight.py", line 30, File"D:\ocr\YOLOv3_TensorFlow\utils\misc_utils.py", line 101, in load_weights (shape[3], shape[2], shape[0], shape[1]))
ValueError: cannot reshape array of size 285787 into shape (256,128,3,3)
what should i do ？ thx

1. If i want to export output graph and use the graph in C++ api, do i need to change the name of the input placeholder? Currenly 'image' has no name. Only phase_train gets to decide whether to get images from data generator

2. How to train with negative samples ( image with no annotations to reduce false positives)

3. Is it possible to get original image size before image is reshaped with 'img_size" ??

Thank you in advance :)

Thanks for repository. would you implementation work with Yolo-SPP configuration from original darknet repository ?

Train. Py file running error, I hope you can help, thank you!

File "D:\Tensorflow yolov3\YOLOv3_TensorFlow-master\utils\data_utils.py", line 142, in parse_data
img, boxes = resize_image_and_correct_boxes(img, boxes, img_size)

File "D:\Tensorflow yolov3\YOLOv3_TensorFlow-master\utils\data_utils.py", line 53, in resize_image_and_correct_boxes
boxes[:, 0] = boxes[:, 0] / ori_width * new_width

IndexError: too many indices for array

2019-03-06 22:22:51.632340: W T:\src\github\tensorflow\tensorflow\core\framework\op_kernel.cc:1306] Invalid argument: TypeError: a bytes-like object is required, not 'str'
Traceback (most recent call last):

File "C:\Users\Xiaos\Anaconda3\lib\site-packages\tensorflow\python\ops\script_ops.py", line 158, in call
ret = func(*args)

File "F:\DL\YOLOv3_TensorFlow\utils\data_utils.py", line 130, in parse_data
pic_path, boxes, labels = parse_line(line)

File "F:\DL\YOLOv3_TensorFlow\utils\data_utils.py", line 19, in parse_line
s = line.strip().split(' ')

TypeError: a bytes-like object is required, not 'str'

what does the error mean? could someone tell me how to construt the train.txt? I use the coco dataset, the txt file is the following
F:/PascalVOCdataset/VOCdevkit/VOC2007/JPEGImages/003949.jpg 11 210 151 245 193

any people know where is the error?

目前正在尝试v3-tiny版本，不过效果还不理想

I've already change the restore_path and set it to the right path, but it didn't work. Would anyone help me out? (I want to change the lr.)

Hello, your project is doing very well. Thank you very much. However, there was a problem with my speed during the test. I trained my own training set with only one category.When testing a picture, I need 2.4s, which is much worse than your test time. I also tried to put model and nms in front of the session in test_single_image.py, but it still has no effect. what should I do?（By the way, my tensorflow version is 1.11.0 ，with the cuda 9.0 and cudnn 7.）

您好，博主，我细看了源码，有个地方很不懂！其中get_kmeans.py的kmean聚类prior anchors时，所使用的数据[W,H]是没有scale到[0,1]，但在训练时scale了anchor_w = anchor_w / W * 416, anchor_h = anchor_g / H * 416，我想问的是将样本数据[W,H]宽高scale到[0,1]，再kmean聚类会不会更好？望回复~，谢谢~

When I followed your step to train a model I met the mistake. Can you give me some advice,thanks a lot.

OutOfRangeError (see above for traceback): End of sequence
[[Node: IteratorGetNext = IteratorGetNextoutput_shapes=[, , , ], output_types=[DT_FLOAT, DT_FLOAT, DT_FLOAT, DT_FLOAT], _device="/job:localhost/replica:0/task:0/device:CPU:0"]]
[[Node: IteratorGetNext/_535 = _Recvclient_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device_incarnation=1, tensor_name="edge_671_IteratorGetNext", tensor_type=DT_FLOAT, _device="/job:localhost/replica:0/task:0/device:GPU:0"]]

thanks for your great job,I need convert .ckpt.meta or .pb to .dlc file
but i just have .weight,
use your convert_weight.py ,it can success convert to .ckpt
but i can't convert to .dlc, it just need some node,this is the node,

why it different to link
can you give me some advise,thank you so much.

Hi wizyoung:

Thank you for providing this awesome repo!

Based on my test, it takes around 100s to run boxes_, scores_, labels_ = sess.run([boxes, scores, labels], feed_dict={input_data: img}) 1000 times with a 416x416 image. This speed is nearly 5 times slower than 23ms you claimed. I didn't change anything on your test_single_image.py file.

Could you help me a little bit with this?

Thank you!

Hi, thank you so much for sharing your code! I have some questions and hope you could help me.I try to freeze graph and convert *.pb to TFlite model, so that can transport to Android mobile phones.This is my source code as follow:

But,I get a unsupported operation error as shown below:

b'2019-01-29 16:12:22.550452: I tensorflow/contrib/lite/toco/import_tensorflow.cc:1080] Converting unsupported operation: ResizeNearestNeighbor\n

Yes, about TFlite unsupported operation error,Could you help how to modify the yolov3 source code to support operation?

楼主好，请问一下为什么yolov3要用reshape将四维tensor转换成5维，是因为这样更适配于GPU还是其他的原因？

How to solve this problem, great god, help me!

I got the above error when I start training,

however, when I remove num_parallel_calls=args.num_threads, in

Should i transform the origin images' bx,by,bw,bh ?
e.g. if my image is 213213 bx =1 by = 1, i should transform the image to 416416 and replace bx = 2 . and put the transformed image's path and bx,by,bh,bw in data.txt ?

作者你好，看原版yolov3的cfg文件，激活层用的是leaky，而这里用的是默认的relu，这个没有影响吗？

I get a training error as shown below

I have no negative samples in my train.txt and i left a blank line at the end of both train.txt and val.txt
test.txt
train.txt

What can be the problem for this error?

I have a question about the precision and recall.I have trained model on voc2007, and the precision and recall can get 90%+ during training, but when I run on the eval dataset, the precision is 80%, more Important is that the recall is only 30%!!!!!! , what can i do to improve the precision especially recall!!!

Hello, How to understand the following sentence :
First stage: Restore darknet53_body part weights from COCO checkpoints, train the yolov3_head with big learning rate like 1e-3 until the loss reaches to a low level, like less than 1.
Second stage: Restore the weights from the first stage, then train the whole model with small learning rate like 1e-4 or smaller. At this stage remember to restore the optimizer parameters if you use optimizers like adam.

i got an error when i try to train my own dataset.

is that because i work it on windows?
i found someone saying it working correctly on Linux when i google this error.
could you please help me?

tensorflow.python.framework.errors_impl.InvalidArgumentError: 0-th value returned by pyfunc_0 is int32, but expects int64
出现这个错误怎么办啊 大佬

I have trained a model successfully, but when i use test_single_image.py to test it showed error as follow, I can't find the reason, can anyone help me? thanks.

InvalidArgumentError (see above for traceback): Restoring from checkpoint failed. This is most likely due to a mismatch between the current graph and the graph from the checkpoint. Please ensure that you have not altered the graph expected based on the checkpoint. Original error:

Assign requires shapes of both tensors to match. lhs shape= [1,1,1024,21] rhs shape= [1,1,1024,255]
[[Node: save/Assign_350 = Assign[T=DT_FLOAT, _class=["loc:@yolov3/yolov3_head/Conv_6/weights"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/device:GPU:0"](yolov3/yolov3_head/Conv_6/weights, save/RestoreV2/_701)]]
[[Node: save/RestoreV2/_372 = _SendT=DT_FLOAT, client_terminated=false, recv_device="/job:localhost/replica:0/task:0/device:GPU:0", send_device="/job:localhost/replica:0/task:0/device:CPU:0", send_device_incarnation=1, tensor_name="edge_378_save/RestoreV2", _device="/job:localhost/replica:0/task:0/device:CPU:0"]]

Hello

I tried to train my data, but I received the below error:
[[{{node PyFunc}}]]
[[{{node ITeratorGetNext}}]]

'During handling of the above exception, another exception occurred:'

[[{{node PyFunc}}]]
[[node ITeratorGetNext (defined at train.py: 60) ]]

and line 60 is:
image_ids, image, y_true_13, y_true_26, y_true_52 = iterator.get_next()

My system has no GPU and I am running the program on CPU. How can I solve it?

Regards

hey, I found the loss of code in other version(ie. https://github.com/BobLiu20/YOLOv3_PyTorch/blob/master/nets/yolo_loss.py) is a little different from here. Anyone can explain this? Thanks!

Whenever i try to restore from lastly saved checkpoint, code seems to train from scratch again.

I don't restore from yolo.ckpt (from COCO dataset) after training.

笔者，您好！希望您在百忙之中可以看到我的请教，希望您可以赐教。万分感谢！
请问您写的这个yolov3版本的新数据集的制作是如何制作的呢？是不是和原版yolov3的不太一样呢？
请问在您示例当中写的xxx/xxx/1.jpg 0 453 369 473 391 1 588 245 608 268中数字的顺序是：类别，方框左上角x坐标，方框左上角y坐标，方框右下角x坐标，方框右下角y坐标的书序吗？
关于自己数据集的制作请问您有相关文章链接吗？大致步骤是怎样的呢？
谢谢

where is video_result.mp4 ? when I python video_test.py ./data/demo_data/vedio.mp4