Training your own YoloX Object Detection Model on Colab - YoloX Object Detection Model Deployment

[@kubaMAX27]

From the depth of my heart I want to thank you for this amazing tutorial. I have struggled with launching training with custom data for god knows how long but thanks to your clear and simple instructions I have managed to do just that.

[@danielleivy8180]

Very underrated channel ❤. I had to switch YOLO models at the last minute due to possible licensing conflicts and I very much needed to learn YOLOx fast!

[@坪田醒悟]

Very easy to understand lecture. I was very lucky to know this one.

[@neuralearn]

We are glad this was helpful. Much more on the way :)

[@ThamChunHwaAMB]

Can you teach how to do the dataset using your way. I am quite blurring because my datasets is different with yours

[@neuralearn]

Hello my dear Tham, you can check out these videos where we explain how to build your own dataset. - kzbin.info/www/bejne/jIaToXqHm6mNfq8 - kzbin.info/www/bejne/i5CblYd6i7eeopI

[@ahmetklc7589]

Hey can i ask you about yolov7 evaluating adjustment. I learn yolox thanks for you but i cant make evaluating on yolov7 can you help me about that? I use yolo format for train model but when i tried to evaluate moel i always get same error " pycocotools unable to run: Results do not correspond to current coco set "

[@ghassenhajjem7233]

My VOC2007 File is empty, does anyone know why

[@abdullahmohammed6533]

Amazing tutorial , keep going :)

[@neuralearn]

Thanks my dear Abdullah, we shall keep going :)

[@khoaoang1907]

My weights has always been able to predict and visualize it in yolox output but it is now showing this error!! Anyone knows any idea why? Really need help with this Traceback (most recent call last): File "/content/YOLOX/tools/demo.py", line 320, in main(exp, args) File "/content/YOLOX/tools/demo.py", line 285, in main model.load_state_dict(ckpt["model"]) File "/usr/local/lib/python3.7/dist-packages/torch/nn/modules/module.py", line 1605, in load_state_dict self.__class__.__name__, " \t".join(error_msgs))) RuntimeError: Error(s) in loading state_dict for YOLOX: size mismatch for head.cls_preds.0.weight: copying a param with shape torch.Size([1, 128, 1, 1]) from checkpoint, the shape in current model is torch.Size([80, 128, 1, 1]). size mismatch for head.cls_preds.0.bias: copying a param with shape torch.Size([1]) from checkpoint, the shape in current model is torch.Size([80]). size mismatch for head.cls_preds.1.weight: copying a param with shape torch.Size([1, 128, 1, 1]) from checkpoint, the shape in current model is torch.Size([80, 128, 1, 1]). size mismatch for head.cls_preds.1.bias: copying a param with shape torch.Size([1]) from checkpoint, the shape in current model is torch.Size([80]). size mismatch for head.cls_preds.2.weight: copying a param with shape torch.Size([1, 128, 1, 1]) from checkpoint, the shape in current model is torch.Size([80, 128, 1, 1]). size mismatch for head.cls_preds.2.bias: copying a param with shape torch.Size([1]) from checkpoint, the shape in current model is torch.Size([80]).

[@tanjunjie9248]

Hi, I wanted to ask whether yolox will be able to be trained with images of sized 1024 and 1280? I heard that yolox can't do that due to its architecture. Do you know whats the biggest image size it can be trained with?

[@Alpha-hj2ss]

Great Tutorial. Thank you so much. Could you explain the inside detail of the network architecture of the model YOLOX in the code by tie with some of their features like Decoupled head, Anchor-free, Multi positives, and SimOTA …….. Thank you

[@neuralearn]

Thanks for the kind words :) Noted

[@ShivangiKeshri-nb7wr]

Can we use the loco dataset for the training the yolox, if not can you share how to convert that into coco or voc dataset so that i able to perform my activity

[@neuralearn]

Hello my dear Shivangi, we converted the dataset into pascal voc format in a previous video, which you can check out in this playlist:kzbin.info/www/bejne/i5CblYd6i7eeopI

[@tanjunjie9248]

Hi, thanks for this video! May I ask how you get the trainval.txt file? I only have images and their annotations for the train and validation dataset

[@neuralearn]

Hello my dear Tan, in this separate video, we explain how to prepare the dataset in detail :) kzbin.info/www/bejne/i5CblYd6i7eeopI This video is part of a tutorial series : kzbin.info/aero/PL2L83ZcMO-5PZN8Ixp5y4wi6atPwbreVl

[@tanjunjie9248]

@@neuralearn I have managed to make it this far into training my yolo model!!! thanks a lot. however during training a lot of my JPEG images ended prematurely!! Do you happen to have any idea why? This is what showed up 2022-06-08 09:03:58 | INFO | yolox.core.trainer:261 - epoch: 1/300, iter: 20/35, mem: 12252Mb, iter_time: 2.404s, data_time: 1.217s, total_loss: 7.1, iou_loss: 3.0, l1_loss: 0.0, conf_loss: 3.5, cls_loss: 0.7, lr: 8.163e-04, size: 672, ETA: 5:51:58 Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file 2022-06-08 09:04:25 | INFO | yolox.core.trainer:261 - epoch: 1/300, iter: 30/35, mem: 12252Mb, iter_time: 2.749s, data_time: 1.462s, total_loss: 6.1, iou_loss: 2.9, l1_loss: 0.0, conf_loss: 2.5, cls_loss: 0.7, lr: 1.837e-03, size: 768, ETA: 6:34:19 Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file 2022-06-08 09:04:36 | INFO | yolox.core.trainer:352 - Save weights to ./YOLOX_outputs/yolox_voc_s 2022-06-08 09:04:36 | INFO | yolox.core.trainer:203 - ---> start train epoch2 Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file 2022-06-08 09:05:02 | INFO | yolox.core.trainer:261 - epoch: 2/300, iter: 10/35, mem: 12252Mb, iter_time: 2.504s, data_time: 1.583s, total_loss: 6.0, iou_loss: 2.8, l1_loss: 0.0, conf_loss: 2.5, cls_loss: 0.7, lr: 2.500e-03, size: 512, ETA: 6:39:44 Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file Premature end of JPEG file

[@tanjunjie9248]

update: It seems that the premature end of the jpeg file doesnt really affect much (?) as im still able to train my model and it it able to detect what I wanted it to detect sucfessfully though im not sure if the accuracy wouldve been better if not for the premature ending of the pictures

[@divyasreereddy8071]

I have an error at the training part. Can you please help me… its very important and urgent!

[@neuralearn]

Hello my dear Divya, Which error exactly?

[@satyajitpanigrahy6162]

Hi, after running the train command , I am getting TypeError: int() argument must be a string, a bytes-like object or a number, not 'NoneType'

[@neuralearn]

Hello my dear Satyajit, please refer to this video, to see how to prepare your dataset : kzbin.info/www/bejne/i5CblYd6i7eeopI

[@maiductai94]

Hi, I have a question: How can I control or adjust the learning rate? Thank you

[@neuralearn]

Hello Tai, At this instance: kzbin.info/www/bejne/mJbCdWaMbrWrosk, you can see that we made an update to the -------------- max_epochs-------------- You can do same for ---------- warmup_lr ---------------- ---------- min_lr_ratio -------------- ---------- basic_lr_per_img ------ When you start training, make sure the configuration (kzbin.info/www/bejne/mJbCdWaMbrWrosk) you set has been respected.

[@ThamChunHwaAMB]

Hi, I saw there are a tensorboard file in the YOLOX_outputs. Do you know how to run that

[@neuralearn]

Hello my dear Tham. We shall add this to the video on testing :)

[@looiwenghoon]

Hi, while following along your tutorial, I faced this problem where it says "yolox.core.launch:98 - An error has been caught in function 'launch', process 'MainProcess' (549), thread 'MainThread' (139716376967040): Traceback (most recent call last):" while training the model, seems like there is an error for ImportError (cannot import name 'notf' from 'tensorboard.compat' (/usr/local/lib/python3.7/dist-packages/tensorboard/compat/__init__.py)) & AttributeError ('StreamToLoguru' object has no attribute 'isatty'). May I know how to solve this problem?

[@tanjunjie9248]

Yes!!! Previously it was working perfectly fine but suddenly this error keeps happening!

[@looiwenghoon]

@@tanjunjie9248 I saw the problem is jax library updates and I supposed the updates clash with colab environments. The solution is simply downgrade the jax and jaxlib library to 0.3.10 version and it will continue to work!

[@tanjunjie9248]

@@looiwenghoon Ah! Do you know how I can downgrade those 2? I'm a noob when it comes to thos!! Thanks for the help!!

[@looiwenghoon]

@@tanjunjie9248 Try: !pip install -U jax==0.3.10 !pip install -U jaxlib==0.3.10 The output should show: (For jax) Uninstalling jax-0.3.14: Successfully uninstalled jax-0.3.14 Successfully installed jax-0.3.10 (For jaxlib) Uninstalling jaxlib-0.3.14+cuda11.cudnn805: Successfully uninstalled jaxlib-0.3.14+cuda11.cudnn805 Successfully installed jaxlib-0.3.10 ** Side note: I tried 0.3.13 version and it works as well

[@tanjunjie9248]

@@looiwenghoon Ok thanks so much, I thought I had clicked on something which messed up the google colab.

[@ardnhj523]

Hi sir, Ive followed your tutorial from the very first video and during the custom training, i have this following error. "AttributeError: 'NoneType' object has no attribute 'text'". I'm stressed and this is my first time doing it, and i followed everything closely but still got error. pls help 🙏

[@neuralearn]

Hello Ardnh, send screenshots to [e-mail] :)

[@ThamChunHwaAMB]

Can I know why the 20 epochs is used to train model?

[@neuralearn]

Hello my dear Tham, sorry for the late response :(. For the dataset we were using, after 20 epochs, we got reasonable results. Though you could train for more epochs, while making sure to stop before the model starts overfitting.

[@Chelsea-ix8bj]

Hello, I have followed your steps and could successfully train my model. It's just that some pictures I have too many bounding boxes to count. Is there any way that could be done to find precision, recall, FT FP on the predicted photo in google colab? Such as comparing the ground truth photo and final photo. Is it possible to make it run as a code on google colab? I am not very good at coding ):

[@neuralearn]

Hello my dear Chelsea. We are proud of your achievement :) Unfortunately these metrics are yet to be supported by yolox. We shall be doing videos on yolov5, which supports this. So stay tuned :)

[@Chelsea-ix8bj]

@@neuralearn ahh so I won't be able to do it even if I tweak some codes and try to run it?

[@neuralearn]

My dear Chelsea, you can obviously add those metrics to the existing yolox code base. My response was based on the fact that you said, 'I am not very good at coding'

[@Chelsea-ix8bj]

@@neuralearn Ahh ok, I am doing a project with the assistance of a supervisor so he was going to aid me with it. Thanks for replying!

[@mouhammadabdallahaw1131]

Amazing tuto. Thank you very much. I have one question. Sorry for my english...I wanna do a training for about 1000 epoch but on colab. With the free version if it stop before the end, is there any way to continue the training where it left ?

[@neuralearn]

Hello my dear Mouhammad :) Yes, you can continue training based on the models state before the runtime was disconnected. You can do this very easily when you integrate Wandb to the training process, since the model will be saved to wandb and you can retrieve it in wandb artifacts. Check out this video on integrating wandb with yolox(kzbin.info/www/bejne/g3-sequOpM-emdU)

[@mouhammadabdallahaw1131]

thank you very much i will check it. With my team we are going to participate in a competition of fighter uav. I am thinking of using Nvidia Jetson Xavier NX as flight computer, its main role will be to analyze the images and detect other uav, makes decisions and send the necessary commands to the flight controller to track the detected drone. We need at least like 20 fps. What version of Yolo could you suggest? Accuracy and speed are important, but we will be more focused on speed because we will have to send the video processed by the flight computer to the jury in real time

[@mouhammadabdallahaw1131]

@@neuralearn do you have any email or social media account where I can contact you ?

[@neuralearn]

Sure. [e-mail]

[@neuralearn]

Since you are using an nvidia gpu, you can make use of tensorrt directly or using tensorrt with onnxruntime. This will help you get high speeds even with high accuracy (larger) yolox models.

[@marouanekhoukh2438]

Google colab link needs Authorization?

[@neuralearn]

Yes it does

[@saurabhbelote6746]

thanks a lot :)

[@neuralearn]

You're welcome my dear Saurabh

[@ThamChunHwaAMB]

Did we need to run the testing datasets

[@neuralearn]

Hello Tham, I don't seem to understand your question. Please what do you mean by 'run the testing datasets'

[@ThamChunHwaAMB]

@@neuralearn Because the datasets we have split into trainval and test one right

[@neuralearn]

Yes, you need the test dataset, to evaluate model performance. Without evaluation, you may deploy an overfitted model!

[@ThamChunHwaAMB]

@@neuralearn how to run the coding for the test datasets? Izit change the train.py to eval.py?

[@ThamChunHwaAMB]

@@neuralearn I hope that you can have a video tutorial about how to evaluate using testing datasets and compare their performance. Thanks

[@SathishKumar-rr1nq]

Super

[@neuralearn]

:) Feel free to check out our other videos

[@fkeyvan]

Going too fast through the data creation part. Also the accent is hard to understand? what is EER? You mean here? DATER? DEEMO?

[@neuralearn]

Hello my dear Farhad, here is the video on data creation: kzbin.info/www/bejne/i5CblYd6i7eeopI

[@fkeyvan]

@@neuralearn thank you

[@neuralearn]

The pleasure is ours :)

[@fkeyvan]

@@neuralearn nice job with remo

[@fkeyvan]

I get the error: "AssertionError: database connection isn't set to UTC" after python -m remo_app

[@datasaver-l4f]

you did not show how to create a dataset total waste of 1 hour

[@neuralearn]

Hello my dear friend, this video is part of a playlist: kzbin.info/aero/PL2L83ZcMO-5PZN8Ixp5y4wi6atPwbreVl We create the dataset in this video:kzbin.info/www/bejne/i5CblYd6i7eeopI