This is the besstttttttt tutorial ever... believe me i have seen more than 20 channels implement this... but if u have landed here this is the bessttt Yolo application on custom dataset video till now... Kudos bruhh
I have gtx 1060 3gb, and this setup causes Error: cuDNN isn't found FWD algo for convolution. I saw in my task manager that it means that i'm out of memory, so I changed width=320, height=320, batches=64, subdivisions=64 and it worked!
@@TheCodingBug I think also that this may due to my big training dataset, it weights 800mb, and I think that this gpu opencv stores them at gpu VRAM, so from 3gb it would be only 2.2gb. My images are quite wide (cyclists dataset) (375,1242) so I guess I could crop them a little + they are png, so converting to jpg should save same space. Thanks for the tutorial!
Awesome work!! I've tried multiple times to get some form of custom object detection up and running. Not until your tutorial have I succeeded! Subscribed and liked!
bro did you run it on the webcam i am trying but couldn't do it (base) C:\darknet\darknet-master\build\darknet\x64>darknet.exe detector data/obj.data cfg/yolov4-obj.cfg backup/yolov4-obj_last.weights -c 0 CUDA-version: 11040 (11040), cuDNN: 8.2.2, GPU count: 1 OpenCV version: 4.5.3 There isn't such command: data/obj.data would you pls help me
@@TheCodingBug Just a side note... this might be a bug with newer versions or updates. Currently rebuilding this on a new system. create_list_of_images.py didn't seem to care whether the files were .jpg or .jpeg on my old system. Now using this tutorial for my new system with new CUDA, opencv, etc... create_list_of_images.py only seems to work if you change the imglist to .jpg if the files are jpg or vice versa for .jpeg. Just figured I'd mention that for people if they end up with a train.txt file that's empty after execution.
Shutdown computer :D JK. You can press cntrl+c and the training would stop. Make sure you have at least one checkpoint in your model output folder so that you can use it.
@@TheCodingBug yess... there's a checkpoint created -yolov4-last_weight...... actually when i started training i got cuda out of memory error so i changed the subdivision to 64.....and now after 12 hours it again showing same error......i think it's bcoz i was using one more anaconda prompt at the same time and running some detection part and it was using cuda memory too
@The Coding bug, Its a great tutorial . I have question that if weights is trained for 2 classes and wanted to add another 2 classes to the existing weights .is it possible ?? .should training start from the scratch ?
How to add an extra class ( for ex : Auto Rickshaw ) to the existing 80 classes from the Coco Dataset. I felt that your explanation was lucid and i would be thankful if you could give a solution to this problem.
You'll have to use original coco dataset and classes and append your custom data and classes with that. So now you'll have 81 classes (80 from coco and 1 custom). Then finetune the existing model.
I want to train the YOLOv4 on custom dataset and then deploy on the Jetson TX2. Can we convert the trained YOLOv4 model (shown using the above method) to Jetson compatible model?
I have not tried that on Jetson Nano. But you can convert these trained models to tflite, which can be used on Jetson Nano. To convert Custom Darknet YOLOV4 model to tflite, I have made another video th-cam.com/video/vzTCJM18uoM/w-d-xo.html
Thanks alot for such amazing and helpful tutorial. I am training for "person" but when i train, i am getting IOU=0.000, i dont know where is the problem or how much IOU threshold I need to set?
There shouldn't be any issue if the training images are around 100. PS: You don't need to train for person class as original YOLOv4 is already trained to detect person.
@@TheCodingBug i have 1500 images for training.. but I want to detect only one class i.e. person thats why i am training for it. its detecting multiclasses which i dont want
@@TheCodingBug what avg loss u think acceptable? because if i aim for the max iteration value it takes more than 24 hour lol, dont want to torture my laptop hahah
@@muhammadrezafairuzi7328 I guess around 0.5 works good for single class. However, there would be a lot of false positives. You can always stop training, evaluate the trained model, amd resume training from last checkpoint created inside backup folder.
I tried your tutorial on installing darknet but im getting error, but when I tried Using vcpkg its working like a charm. But the problem is darknet.exe is in darknet-master not in x64. Can I just transfer it to x64?
Hello, with which command can I detect object on all photos in a particular folder in yolov4? I want it to be done one by one in all the photos one after the other.
Great! and I have a question that when the customized training is finished. I have 3 weight files: (1) _10000.weights, _last.weights, _final_weights. Can you tell me which one is the "final" weight file which I should use later. (I set iteration = 10000). Thank you
darknet.exe detector train data/obj.data cfg yolov4.conv.137 after that i am getting darknet.exe' is not recognized as an internal or external command, operable program or batch file.that type of error
Hi, I know that a weights file is created in every 1000 iterations. But there is only weight file that named "yolov4_last.weights". I am at 3500. iteration right now. How can i provide to be new weight file in every 1000 iterations batch=64 subdivisions=64 (commond prompt wanted the value to be 64) max_batches=50000 steps=40000,45000 classes=3 filters=24
Hello! Thank you very much for this tutorial. it's really helpfull and well explain. I have a question, what can I do if I keep getting the "out of memory" error despite the fact that I increased the subdivision to 64 and even 128 already and deacrease the batch to 32? I only have one classe and 54 images.
If the darknet is built without any errors, you should have darknet.exe inside x64 folder. If you have not built it, you need to do it first by following this tutorial: th-cam.com/video/FE2GBeKuqpc/w-d-xo.html
Thank u for the greatest video!!! I have a question. How do i change iteration save file??? Default is every per 1000 iteration. I wanna every per 100 iteration. Also i want to different name, like yolov4-obj_l100.weights, yolov4-obj_200.weights, ....yolov4-obj_10000.weights this. Thank u so much
Yes you can do that. It would be darknet.exe detector data/obj.data cfg/yolov4-obj.cfg backup/yolov4-obj_last.weights -c 0 Kindly try it and let me know if it works for you!
After giving darknet_no_gpu.exe detector train data/obj.data cfg/yolov4-obj.cfg yolov4.conv.137 I'm getting couldn't open file:data/obj.data.Can you help me?. Thanks in advance.
If you're trying to train a dataset different from the video, verify the image extension (i.e. .jpeg or .jpg) and change it in the script accordingly. Hope it helps!
CUDA-version: 11040 (11040), cuDNN: 8.2.2, GPU count: 1 OpenCV version: 4.5.3 Not an option: demo sir pls help me There isn't such command: data/obj.data (base) C:\darknet\darknet-master\build\darknet\x64>darknet.exe detector data/obj.data cfg/yolov4-obj.cfg backup/yolov4-obj_last.weights -c 1 CUDA-version: 11040 (11040), cuDNN: 8.2.2, GPU count: 1 OpenCV version: 4.5.3 There isn't such command: data/obj.data pls sir help me what should i do pls pls
Try decreasing the learning rate and make sure that the labels for bounding boxes are not absolute values but the ratio (absolute value / width(height))
Is there anyway we can load the last.Weights file inside a python file and run the prediction on images/videos? How can I load it using Python code. Thank you
@@TheCodingBug Thank you for the quick replay! As for my question, In your video we were able to convert .weights file to TensorFlow and .pb file. What I want know is Python code to load the file and start making predictions on video not using CMD. Thank you
I am getting an error at 8:40 . Last 5 lines of whole error is: Wrong annotation: class_id = 29. But class_id should be [from 0 to 0], file: data/obj/john wick_30.txt Wrong annotation: class_id = 12. But class_id should be [from 0 to 0], file: data/obj/john wick_13.txt Wrong annotation: class_id = 9. But class_id should be [from 0 to 0], file: data/obj/john wick_10.txt Wrong annotation: class_id = 23. But class_id should be [from 0 to 0], file: data/obj/john wick_24.txt Error: cuDNN isn't found FWD algo for convolution. I am getting a lot of errors/warning that titled "Wrong annotation:"so ı am stuck at this stage. What should i do? My english is not good. I hope i could explain myself. I am sorry for my english
You've made a mistake where we define number of classes in obj.data file. Classes =1 Also, make sure you have class id 0 against each annotation in data/obj/john wick_30.txt
@@TheCodingBug Thanks for your answer. I did all changes you said. Now contents of obj.data is ; classes= 1 train = data/train.txt names = data/obj.names backup = backup/ I have changed the whole id. For example; I changed the id number at the beginning of the value in all txt files with 0. Now all values start with 0, like that; 0 0.473333 0.355450 0.416000 0.706161 But ı still have an other error like this; [yolo] params: iou loss: ciou (4), iou_norm: 0.07, obj_norm: 1.00, cls_norm: 1.00, delta_norm: 1.00, scale_x_y: 1.05 nms_kind: greedynms (1), beta = 0.600000 Total BFLOPS 59.563 avg_outputs = 489778 Allocate additional workspace_size = 82.58 MB Loading weights from yolov4.conv.137... seen 64, trained: 0 K-images (0 Kilo-batches_64) Done! Loaded 137 layers from weights-file Learning Rate: 0.001, Momentum: 0.949, Decay: 0.0005 Detection layer: 139 - type = 28 Detection layer: 150 - type = 28 Detection layer: 161 - type = 28 If error occurs - run training with flag: -dont_show Resizing, random_coef = 1.40 608 x 608 Create 6 permanent cpu-threads Error: cuDNN isn't found FWD algo for convolution. What should i do?
@@sefadogan4795 you also need yo change kernel sizes in all YOLO layers, as explained in the video. Also, make sure compatible versions of CUDA and cuDNN are installed, as explained in building yolov4 darknet video.
@@TheCodingBug I installed yolov4, CUDA and cuDNN as explained in the yolov4 darknet video, it run as seamless. just have a problems on this stage. Also i checked the video to change kernel size but ı could not find part of kernel. Maybe because of my listening is poor. Could you please write me the video minute so that I can find it?
Can someone help me, I have this error in the process of training: Error: cuDNN isn't found FWD algo for convolution. idk what to do. I already reinstall CUDA 10.1 and cuDNN 8.0.5 for CUDA 10.1
Hi. I have completed the training. However, the output is null. No identification with boundary box. Could you pls help me to sort this out? I had only one class pole. Trained for almost 20 hrs. Thanks in advance. *** output Enter Image Path: d0100000810.jpg Detection layer: 139 - type = 28 Detection layer: 150 - type = 28 Detection layer: 161 - type = 28 d0100000810.jpg: Predicted in 50.263000 milli-seconds. ***output.
@@pmunozuio thank you for the appreciation. RTX models always have some problems because it requires updated CUDA and drivers. I would suggest building opencv 4.5.1 with CUDA 11 and then build darknet. The links to these tutorials are in the description.
Hello, first of all, great tutorial ! I just have a question, how i know if during the training im actually using the GPU instead of the cpu? and second, sorry if the question is dumb, (im new in this) but we shouldnt change in the Makefile GPU = 1, Opencv = 1, and so on? thank you in advanced, and new subscriber here!
Hi. You can use task manager to see CUDA utilization. If you have built darknet using my tutorial, you don't need to change anything in make file as darknet would detect GPU automatically.
@@TheCodingBug okay thank you. I asked because at first it was really fast the training, but then came an error "CUDA OOM" i restarted it and then went very slow. after that I have not been able to make them enter faster.
Why are enabling tensor cores after 3000 iterations ,why not for the start? i have another question my gpu memory is being used but no gpu processing power is being used is it normal
I do not understand what you meant by enabling tensor cores after 3000 iterations. I did not do anything like that! If GPU utilization is at 0%, its alarming. However, normally, it fluctuates between 10-40%, depending on the batch size. If you have high end GPU with more than 6GB Memory, you can increase batch size or reduce subdivisions (from 32 to 16).
@@TheCodingBug when I start training (it is training) I got a message after each iteration that tensor cores are disabled until first 3000 iterations. (My gpu is rtx 2070)
Why not use them from the strat of the processing. I think you might not have the same problem as the gpu you used for the tetorial did not had any tensor cores.
@@hamzashah9543 Mine is 1060. I do not get this error. Let the training run over night. Hopefully it is running on GPU. For 2070, you should consider changing subdivisions=16 as its 8GB card.
@@TheCodingBug I will leave it over night. Just wanted to let you it's not a error To my limited knowledge either by default or somewhere in the configuration it's set to do so. Anyway I will let the training continue. I am also getting a Warning of performance bottle neck on cpu or disk hdd/ssd May be that's the issue
Link to Part 1: th-cam.com/video/sKDysNtnhJ4/w-d-xo.html Build YOLOv4 Darknet on Windows 10: th-cam.com/video/FE2GBeKuqpc/w-d-xo.html Building OpenCV 4.4.0 with GPU from source: th-cam.com/video/tjXkW0-4gME/w-d-xo.html
@TheCodingBug I'm facing this issue can someone help when starting training model loads but at last when graph window come it stops CUDA status Error: file: C:\dark\darknet-master\darknet-master\src\dark_cuda.c : cuda_push_array() : line: 589 : build time: Aug 12 2021 - 20:06:17 CUDA Error: an illegal memory access was encountered
darknet.exe : The term 'darknet.exe' is not recognized as the name of a cmdlet, function, script file, or operable program. Check the spelling of the name, or if a path was included, verify that the path is correct and try again. Dose anyone face the same problem? I'm using Windows 10 to train with the command line＂darknet.exe detector train data/obj.data yolo-obj.cfg yolov4.conv.137＂ and it shown this error.
608 x 608 Create 6 permanent cpu-threads`enter code here` used slow CUDNN algo without Workspace! Need memory: 2217992, available: 0 CUDNN-slow used slow CUDNN algo without Workspace! Need memory: 554504, available: 0 CUDNN-slow Try to set subdivisions=64 in your cfg-file. CUDA status Error: file: E:\darknet\src\dark_cuda.c : cuda_make_array() : line: 493 : build time: Apr 27 2021 - 18:51:36 CUDA Error: out of memory CUDA Error: out of memory: No error** this is the error im facing i tried altering subdivisions and batches in the cfg file still the same GPU : Nvidia GEFORCE 940MX 2GB
Can someone help me, I have this error in the process of training: Error: cuDNN isn't found FWD algo for convolution. my gpu is Nvidia gtx 1050Ti and I have cuda 11.3 with cudnn 8.2.0
thanks god, someone did a tutorial without using google collab
Ty so much
This is the besstttttttt tutorial ever... believe me i have seen more than 20 channels implement this... but if u have landed here this is the bessttt Yolo application on custom dataset video till now... Kudos bruhh
I am glad it was helpful!
I have gtx 1060 3gb, and this setup causes Error: cuDNN isn't found FWD algo for convolution.
I saw in my task manager that it means that i'm out of memory, so I changed width=320, height=320, batches=64, subdivisions=64 and it worked!
Thank you for the update. I hope it'll help someone here.
@@TheCodingBug I think also that this may due to my big training dataset, it weights 800mb, and I think that this gpu opencv stores them at gpu VRAM, so from 3gb it would be only 2.2gb.
My images are quite wide (cyclists dataset) (375,1242) so I guess I could crop them a little + they are png, so converting to jpg should save same space.
Thanks for the tutorial!
thanks.. this helped me in resolving the error
This worked for me also. Thanks
Nice nice it worked as well for bro THANKYOU
Thank you For all your tutorials. You have the best tutorial out of all Yolo tutorials out there!
sorry I'm new I can't find the darkne.exe file... did I skip any steps?
Awesome work!! I've tried multiple times to get some form of custom object detection up and running. Not until your tutorial have I succeeded! Subscribed and liked!
I am glad it was helpful.
bro did you run it on the webcam i am trying but couldn't do it
(base) C:\darknet\darknet-master\build\darknet\x64>darknet.exe detector data/obj.data cfg/yolov4-obj.cfg backup/yolov4-obj_last.weights -c 0
CUDA-version: 11040 (11040), cuDNN: 8.2.2, GPU count: 1
OpenCV version: 4.5.3
There isn't such command: data/obj.data
would you pls help me
@@TheCodingBug Just a side note... this might be a bug with newer versions or updates. Currently rebuilding this on a new system. create_list_of_images.py didn't seem to care whether the files were .jpg or .jpeg on my old system. Now using this tutorial for my new system with new CUDA, opencv, etc... create_list_of_images.py only seems to work if you change the imglist to .jpg if the files are jpg or vice versa for .jpeg. Just figured I'd mention that for people if they end up with a train.txt file that's empty after execution.
@@mdasrafuzzamansakib9812 idk if this is too late or not but the shell probably expected 'demo' after 'detector' in the command
thank you.
This tutorial really helps me a lot.
Great video! only a small question - how to stop the training?
Shutdown computer :D
JK. You can press cntrl+c and the training would stop. Make sure you have at least one checkpoint in your model output folder so that you can use it.
@@TheCodingBug yess... there's a checkpoint created -yolov4-last_weight...... actually when i started training i got cuda out of memory error so i changed the subdivision to 64.....and now after 12 hours it again showing same error......i think it's bcoz i was using one more anaconda prompt at the same time and running some detection part and it was using cuda memory too
great tutorial! thank you so muchh.. but can I set the line_thickness to some spesific size? thankyou..
@The Coding bug, Its a great tutorial . I have question that if weights is trained for 2 classes and wanted to add another 2 classes to the existing weights .is it possible ?? .should training start from the scratch ?
yes
How to add an extra class ( for ex : Auto Rickshaw ) to the existing 80 classes from the Coco Dataset. I felt that your explanation was lucid and i would be thankful if you could give a solution to this problem.
You'll have to use original coco dataset and classes and append your custom data and classes with that. So now you'll have 81 classes (80 from coco and 1 custom). Then finetune the existing model.
When I start Training I got this error, Why it is how to solve this.
Error: cuDNN isn't found FWD algo for convolution.
Can you also make a video on counting objects ?
I love your tutorials! but What will be the max_batch size if I labeled 25000 images and have 1 class? Please help me!
I want to train the YOLOv4 on custom dataset and then deploy on the Jetson TX2. Can we convert the trained YOLOv4 model (shown using the above method) to Jetson compatible model?
I have not tried that on Jetson Nano. But you can convert these trained models to tflite, which can be used on Jetson Nano. To convert Custom Darknet YOLOV4 model to tflite, I have made another video th-cam.com/video/vzTCJM18uoM/w-d-xo.html
i get an error "Error: cuDNN isn't found FWD algo for convolution." when training data. any way to solve this?
Thanks alot for such amazing and helpful tutorial.
I am training for "person" but when i train, i am getting IOU=0.000, i dont know where is the problem or how much IOU threshold I need to set?
There shouldn't be any issue if the training images are around 100.
PS: You don't need to train for person class as original YOLOv4 is already trained to detect person.
@@TheCodingBug i have 1500 images for training.. but I want to detect only one class i.e. person thats why i am training for it. its detecting multiclasses which i dont want
@@khazinahasan7432 you could just filter the coco dataset though through the use of python wrapper
Ok much thanks 😊
Great video and explanation, thanks!
I am glad I could help.
@@TheCodingBug what avg loss u think acceptable? because if i aim for the max iteration value it takes more than 24 hour lol, dont want to torture my laptop hahah
@@muhammadrezafairuzi7328 I guess around 0.5 works good for single class. However, there would be a lot of false positives.
You can always stop training, evaluate the trained model, amd resume training from last checkpoint created inside backup folder.
I tried your tutorial on installing darknet but im getting error, but when I tried Using vcpkg its working like a charm. But the problem is darknet.exe is in darknet-master not in x64. Can I just transfer it to x64?
I haven't tried it. But you can do it I guess.
Keep us posted if it works.
Hello, with which command can I detect object on all photos in a particular folder in yolov4? I want it to be done one by one in all the photos one after the other.
Great! and I have a question that when the customized training is finished. I have 3 weight files: (1) _10000.weights, _last.weights, _final_weights. Can you tell me which one is the "final" weight file which I should use later. (I set iteration = 10000). Thank you
The one with "last" in it's name.
darknet.exe detector train data/obj.data cfg yolov4.conv.137 after that i am getting darknet.exe' is not recognized as an internal or external command,
operable program or batch file.that type of error
Make sure you have darknet.exe and you're in the same folder when issuing the command.
If we have two class of Images. then can we save those images in data/obj/ folder with seperate folder or we can save it into same folder .
Hey mate, is it possible to train if you got amd graphics card?
Can we run this on a youtube video instead of a video from our local repository ?
How to enable gpu in flask? Is there any gpu inference for yolov4 in python ? Any help regarding this will appreciated , thank
Hi, I know that a weights file is created in every 1000 iterations. But there is only weight file that named "yolov4_last.weights". I am at 3500. iteration right now. How can i provide to be new weight file in every 1000 iterations
batch=64
subdivisions=64 (commond prompt wanted the value to be 64)
max_batches=50000
steps=40000,45000
classes=3
filters=24
Hello!
Thank you very much for this tutorial. it's really helpfull and well explain.
I have a question, what can I do if I keep getting the "out of memory" error despite the fact that I increased the subdivision to 64 and even 128 already and deacrease the batch to 32?
I only have one classe and 54 images.
Decrease the batch to 2.
@@TheCodingBug thanks for the hint. But is there a way to increase the allocated GPU?
You are the best Thank you so much 💗
I am glad it was helpful.
Can u tell the command for mAP and AP calculation?
AFTER TRAINING
How can i take accuracy percentage value from command prompt to use in my code?
So, we can say that YOLO v4 is good for Face recognition as well?
Yes indeed. And that too with very small number of training images.
"'darknet.exe' is not recognized as an internal or external command,..." I got the error. I cannot found the file darknet.exe in the x64 folder
If the darknet is built without any errors, you should have darknet.exe inside x64 folder. If you have not built it, you need to do it first by following this tutorial:
th-cam.com/video/FE2GBeKuqpc/w-d-xo.html
@@TheCodingBug ok! thank you for your instant response. I will follow the tutorial. Thanks
@@TheCodingBug Is it the same solution if ./darknet not working in linux
@@lujashakya4399 Yes. You have to build darknet from source.
Thank u for the greatest video!!!
I have a question.
How do i change iteration save file???
Default is every per 1000 iteration. I wanna every per 100 iteration.
Also i want to different name, like yolov4-obj_l100.weights, yolov4-obj_200.weights, ....yolov4-obj_10000.weights this.
Thank u so much
Can we check it on webcam as well ? Instead of video and image , if so what's the command ?
Yes you can do that. It would be
darknet.exe detector data/obj.data cfg/yolov4-obj.cfg backup/yolov4-obj_last.weights -c 0
Kindly try it and let me know if it works for you!
@@TheCodingBug thanks champ
I am glad it worked out for you.
While I train my yolov4 it doesn't use my GPU although it has detected my GPU and rest of the setup done
You can verify through task manager under performance tab whether or not CUDA is being utilized.
After giving darknet_no_gpu.exe detector train data/obj.data cfg/yolov4-obj.cfg yolov4.conv.137 I'm getting couldn't open file:data/obj.data.Can you help me?. Thanks in advance.
It seems obj.data is not present at the provided path.
@@TheCodingBug actually I tried giving absolute and relative path. Still the error exist
Hello, please help me when I run the commend create_list_of_images.py after running there is not any data in the train.txt file
It seems that you need to edit paths to your images directory.
If you're trying to train a dataset different from the video, verify the image extension (i.e. .jpeg or .jpg) and change it in the script accordingly.
Hope it helps!
same issue with me
CUDA-version: 11040 (11040), cuDNN: 8.2.2, GPU count: 1
OpenCV version: 4.5.3
Not an option: demo sir pls help me
There isn't such command: data/obj.data
(base) C:\darknet\darknet-master\build\darknet\x64>darknet.exe detector data/obj.data cfg/yolov4-obj.cfg backup/yolov4-obj_last.weights -c 1
CUDA-version: 11040 (11040), cuDNN: 8.2.2, GPU count: 1
OpenCV version: 4.5.3
There isn't such command: data/obj.data
pls sir help me what should i do pls pls
Hi, while custom training I'm getting nan avg loss at 30 iteration. How to resolve it?
Try decreasing the learning rate and make sure that the labels for bounding boxes are not absolute values but the ratio (absolute value / width(height))
@Coding bug, after decreasing the learning rate still I'm getting nan avg loss
Hi friend. How can I cut the detected object?
Is there anyway we can load the last.Weights file inside a python file and run the prediction on images/videos? How can I load it using Python code. Thank you
Yes you can use OpenCV to directly read Darknet Weights files. th-cam.com/video/FjyF03uawsA/w-d-xo.html
@@TheCodingBug Thank you for the quick replay!
As for my question, In your video we were able to convert .weights file to TensorFlow and .pb file. What I want know is Python code to load the file and start making predictions on video not using CMD. Thank you
I am getting an error at 8:40 . Last 5 lines of whole error is:
Wrong annotation: class_id = 29. But class_id should be [from 0 to 0], file: data/obj/john wick_30.txt
Wrong annotation: class_id = 12. But class_id should be [from 0 to 0], file: data/obj/john wick_13.txt
Wrong annotation: class_id = 9. But class_id should be [from 0 to 0], file: data/obj/john wick_10.txt
Wrong annotation: class_id = 23. But class_id should be [from 0 to 0], file: data/obj/john wick_24.txt
Error: cuDNN isn't found FWD algo for convolution.
I am getting a lot of errors/warning that titled "Wrong annotation:"so ı am stuck at this stage. What should i do?
My english is not good. I hope i could explain myself. I am sorry for my english
You've made a mistake where we define number of classes in obj.data file.
Classes =1
Also, make sure you have class id 0 against each annotation in data/obj/john wick_30.txt
@@TheCodingBug Thanks for your answer. I did all changes you said.
Now contents of obj.data is ;
classes= 1
train = data/train.txt
names = data/obj.names
backup = backup/
I have changed the whole id.
For example;
I changed the id number at the beginning of the value in all txt files with 0. Now all values start with 0, like that;
0 0.473333 0.355450 0.416000 0.706161
But ı still have an other error like this;
[yolo] params: iou loss: ciou (4), iou_norm: 0.07, obj_norm: 1.00, cls_norm: 1.00, delta_norm: 1.00, scale_x_y: 1.05
nms_kind: greedynms (1), beta = 0.600000
Total BFLOPS 59.563
avg_outputs = 489778
Allocate additional workspace_size = 82.58 MB
Loading weights from yolov4.conv.137...
seen 64, trained: 0 K-images (0 Kilo-batches_64)
Done! Loaded 137 layers from weights-file
Learning Rate: 0.001, Momentum: 0.949, Decay: 0.0005
Detection layer: 139 - type = 28
Detection layer: 150 - type = 28
Detection layer: 161 - type = 28
If error occurs - run training with flag: -dont_show
Resizing, random_coef = 1.40
608 x 608
Create 6 permanent cpu-threads
Error: cuDNN isn't found FWD algo for convolution.
What should i do?
@@sefadogan4795 you also need yo change kernel sizes in all YOLO layers, as explained in the video.
Also, make sure compatible versions of CUDA and cuDNN are installed, as explained in building yolov4 darknet video.
@@TheCodingBug I installed yolov4, CUDA and cuDNN as explained in the yolov4 darknet video, it run as seamless. just have a problems on this stage.
Also i checked the video to change kernel size but ı could not find part of kernel. Maybe because of my listening is poor.
Could you please write me the video minute so that I can find it?
Can someone help me, I have this error in the process of training:
Error: cuDNN isn't found FWD algo for convolution.
idk what to do. I already reinstall CUDA 10.1 and cuDNN 8.0.5 for CUDA 10.1
set batch value to 64
Hi. I have completed the training. However, the output is null. No identification with boundary box. Could you pls help me to sort this out? I had only one class pole. Trained for almost 20 hrs. Thanks in advance.
*** output
Enter Image Path: d0100000810.jpg
Detection layer: 139 - type = 28
Detection layer: 150 - type = 28
Detection layer: 161 - type = 28
d0100000810.jpg: Predicted in 50.263000 milli-seconds.
***output.
I had the same problem with you, can I know did you solve it?
hello every one me , i have problem beacause i am using for the annotation file xml extension
How many image data used for this tutorial
95 images I guess.
yOOO THANK YOU, U MADE YOLO EASER
Thank you for the appreciation.
@@TheCodingBug btw where to download the darknet files??
@@boringchannel6313 You have to build it. th-cam.com/video/FE2GBeKuqpc/w-d-xo.html
@@TheCodingBug ok ty
'darknet.exe' is not recognized as an internal or external command , What can i do?
You need to build darknet and make sure you have darknet.exe in x64 folder.
What is your graphics card with which you run the program? I see you are running 21 FPS, please help me
I have GTX 1060 and I have built opencv with CUDA support.
@@TheCodingBug your post its very good, I have gpu rtx 3070, but I have problems running, because there are no drivers yet, maybe some suggestion
@@pmunozuio thank you for the appreciation.
RTX models always have some problems because it requires updated CUDA and drivers. I would suggest building opencv 4.5.1 with CUDA 11 and then build darknet.
The links to these tutorials are in the description.
can i do without gpu support for andriod app
Hello, first of all, great tutorial !
I just have a question, how i know if during the training im actually using the GPU instead of the cpu?
and second, sorry if the question is dumb, (im new in this) but we shouldnt change in the Makefile GPU = 1, Opencv = 1, and so on?
thank you in advanced, and new subscriber here!
Hi. You can use task manager to see CUDA utilization.
If you have built darknet using my tutorial, you don't need to change anything in make file as darknet would detect GPU automatically.
@@TheCodingBug okay thank you. I asked because at first it was really fast the training, but then came an error "CUDA OOM" i restarted it and then went very slow. after that I have not been able to make them enter faster.
@@AndresPerez-qd8pn Restart the computer. Then change subdivisions in config file as you have less memory for GPU.
@@TheCodingBug Thank you, great helpful channel!!
if I have 5 classes and i put the batches only on 4000. What it does effect?
4000 for 2 classes.
if you use 5 classes you can use 10000 batch, but thats need more time for works.
please link to download "darknet.exe" not the video
Why are enabling tensor cores after 3000 iterations ,why not for the start?
i have another question my gpu memory is being used but no gpu processing power is being used
is it normal
I do not understand what you meant by enabling tensor cores after 3000 iterations. I did not do anything like that!
If GPU utilization is at 0%, its alarming. However, normally, it fluctuates between 10-40%, depending on the batch size. If you have high end GPU with more than 6GB Memory, you can increase batch size or reduce subdivisions (from 32 to 16).
@@TheCodingBug when I start training (it is training) I got a message after each iteration that tensor cores are disabled until first 3000 iterations.
(My gpu is rtx 2070)
Why not use them from the strat of the processing.
I think you might not have the same problem as the gpu you used for the tetorial did not had any tensor cores.
@@hamzashah9543 Mine is 1060. I do not get this error. Let the training run over night. Hopefully it is running on GPU. For 2070, you should consider changing subdivisions=16 as its 8GB card.
@@TheCodingBug I will leave it over night.
Just wanted to let you it's not a error
To my limited knowledge either by default or somewhere in the configuration it's set to do so.
Anyway I will let the training continue.
I am also getting a
Warning of performance bottle neck on cpu or disk hdd/ssd
May be that's the issue
How to evaluate the model ?
How can we test it ?
Make an android app for Yolov4
I am in the process of making the video. Hopefully, it will be uploaded this week. Stat tuned.
@@TheCodingBug ❤️
First 😀
Link to Part 1:
th-cam.com/video/sKDysNtnhJ4/w-d-xo.html
Build YOLOv4 Darknet on Windows 10:
th-cam.com/video/FE2GBeKuqpc/w-d-xo.html
Building OpenCV 4.4.0 with GPU from source:
th-cam.com/video/tjXkW0-4gME/w-d-xo.html
@TheCodingBug
I'm facing this issue can someone help
when starting training model loads but at last when graph window come it stops
CUDA status Error: file: C:\dark\darknet-master\darknet-master\src\dark_cuda.c : cuda_push_array() : line: 589 : build time: Aug 12 2021 - 20:06:17
CUDA Error: an illegal memory access was encountered
darknet.exe : The term 'darknet.exe' is not recognized as the name of a cmdlet, function, script file, or operable program. Check
the spelling of the name, or if a path was included, verify that the path is correct and try again.
Dose anyone face the same problem? I'm using Windows 10 to train with the command line＂darknet.exe detector train data/obj.data yolo-obj.cfg yolov4.conv.137＂ and it shown this error.
SIR CAN I CONTACT YOU
opencv_world451.dill not found. this is the error is am getting, please help
You need to build OpenCV with option "opencv_world" enabled.
608 x 608
Create 6 permanent cpu-threads`enter code here`
used slow CUDNN algo without Workspace! Need memory: 2217992, available: 0
CUDNN-slow used slow CUDNN algo without Workspace! Need memory: 554504, available: 0
CUDNN-slow Try to set subdivisions=64 in your cfg-file.
CUDA status Error: file: E:\darknet\src\dark_cuda.c : cuda_make_array() : line: 493 : build time: Apr 27 2021 - 18:51:36
CUDA Error: out of memory
CUDA Error: out of memory: No error**
this is the error im facing
i tried altering subdivisions and batches in the cfg file still the same
GPU : Nvidia GEFORCE 940MX 2GB
With 608*608, it'll require 4GB GPU memory. Try 416*416
I edited to 416 in yolov4-obj.cfg file but still its showing 608*608
Can someone help me, I have this error in the process of training:
Error: cuDNN isn't found FWD algo for convolution.
my gpu is Nvidia gtx 1050Ti and I have cuda 11.3 with cudnn 8.2.0
set batch value to 64