YOLOv1 from Scratch

[@AladdinPersson]

Here's the outline for the video: 0:00 - Introduction 0:24 - Understanding YOLO 08:25 - Architecture and Implementation 32:00 - Loss Function and Implementation 58:53 - Dataset and Implementation 1:17:50 - Training setup & evaluation 1:40:58 - Thoughts and ending

[@venkatesanr9455]

Highly helpful and awesome

[@omarabubakr6524]

why didn't you explain the utils file?

[@PaAGadirajuSanjayVarma]

Plz give this man a noble proze

[@deeps-n5y]

*Nobel

[@iiVEVO]

A noble nobel prize*

[@vijayabhaskar-j]

This series was super helpful, can you please continue this by making one for Yolo v3, v4, SSD, and RetinaNet? That will make this content more unique because none of the channels that explains all these architectures and your explanations are great!

[@jertdw3646]

I'm confused on how i'm supposed to load the images up for training. Did you get that part?

[@Glitch40417]

​​@@jertdw3646on't know if you got it or not, actually there's a train.csv file. Instead of 8examples.csv or 100examples.csv we can use that file.

[@MohamedAli-dk6cb]

One of the greatest deep learning videos I have ever seen online. You are amazing Aladdin, please keep going with the same style. The connections you make between the theory and the implementation is beyond PhD level. Wish I can give you more than one like.

[@asiskumarroy4470]

I dont know how do I express my gratitude to you.Thanks a lot brother.

[@caidexiao9839]

Thanks a lot for you kindness to provide the yolov1 video. By the end of the video, you got mAP close to 1.0 with only 8 training images. I guess you used weights of a well trained model. With more than 10,000 images and more than 20 hours on Kaggle 's free GPU, my mAP was about 0.7, but my validation mAP was less than 0.2. Nobody mentioned the over fitting issue of yolo v1 model training.

[@satvik4225]

mine is coming 0.0 always

[@_nttai]

I was lost somewhere in the loss but still watch the whole thing. Great video. Thank you

[@nguyenthehoang9148]

By far, your series is one of the best content about computer vision on KZbin. It's very helpful when people explain how things work under the hood, like the very well-known courses by Andrew Ng. If you make a paid course for this kind of content, I'll definitely buy it.

[@keshavaggarwal5835]

Best Channel ever. Cleared all doubts about YOLO. I was able to implement this in tensorflow by following your guide with ease. Thanks a lot bro.

[@AladdinPersson]

Awesome to hear it! Leave a link to Github and people could use that if they are also doing it for TF?:)

[@Skybender153]

Link for the tensorflow repo would be appreciated Keshav

[@_adi_1900]

This channels going to blow up now. Great stuff!

[@AladdinPersson]

🙏 🙏

[@thanhquocbaonguyen8379]

massively thank you for implementing this in pytorch and explain every bits in detail. it was really helpful for my university project. i have watched your tutorials at least 3 times. thank you!

[@abireo2285]

PhDs are 100% learning how to code here :)

[@Anonymous-nz8wd]

GOD DAMN! I was searching for this for a really long time but you did it, bro. Fantastic.

[@eminemhc5763]

Only 3.5K subscribers ??? One of the most underrated channel in KZbin Keep posting quality video like this bro , soon you will reach 100K+ subs , congrats in advance Thanks for the quality content :)

[@AladdinPersson]

Appreciate the kinds words 🙏 🙏

[@rampanda2361]

The savior, Been looking at codes of other people for few days, Could not understand it better as those were codes only with no explanation what so ever. Thank you very much.

[@haldiramsharma4601]

Best channel ever!! All because of you, I learned to implement everything from scatch!! Thank you very much

[@sangrammishra4396]

I love the way he explained and always maimtain simplicity in explaining the code, thanks aladdin

[@crazynandu]

Great Video as usual . Looking forward to see RCNNs (mask , faster , fast , ..) from scratch from you !! Similar to Transformers you did, you can do one from scratch and other using the torchvision's implementation .Kudos !!

[@Тима-щ2ю]

What an amount of work! I don't often see people in the internet that are so dedicated to deep learning!

[@thetensordude]

Most underrated channel!!!

[@vanglequy7844]

Let's look at it upside down then!

[@sachavanweeren9578]

I can imagine this video took a lot of time to prepare, the result is great and super helpful. Thank you very much. Respect!

[@张子诚-z3b]

I'm a beginner of object detection, You videos help me a lot. I really like your style of code.

[@ai4popugai]

The most clear explanation that I have ever found, thank you!!

[@WiktorJurek]

This is insanely valuable. Thank you very much, dude.

[@ИльяЯгупов-н4я]

Thank you so much for this video, it's so helpful! Especially the concept in first 9 minutes. I read a lot of sources, but here it's the only place where it is clearly explauned. And more precisely the part where we are looking for a cell with midpoint of bounding box! Thank you so much for a great Explanation!

[@krzysztofmajchrzak1881]

I want to thank so much! It is literally a live saver for me! Your channel is underrated!

[@nikolayandcards]

So glad I came across your channel (Props to Python Engineer). Very valuable content. Thanks for sharing and you have gained a new loyal subscriber/fan lol.

[@AladdinPersson]

Welcome 😁

[@vamsibalijepally3431]

def test(S=7, B=2,C=20): model = Yolov1(in_channels=3,split_size=S,num_boxes = B,num_classes=C) x = torch.randn((2,3,448,448)) print(model(x).shape) this will throw help if got same error like me __init__() missing 1 required positional argument: 'kernel_size'

[@pranavkushare6788]

Yeah i'm getting the same error. Have you found any solution and reason ?

[@chinmay996]

@@pranavkushare6788 if you still have not solved the problem, check your parameters in CNNBlock inside _create_conv_layers method.

[@TheDroidMate]

Amazing video series, thanks! Extra kudos for the OS you're using 💜

[@정래혁-c8y]

This video was so helpful. Thank you!

[@abireo2285]

This is the best deep learning coding video I have ever seen.

[@Epistemophilos]

Is there a mistake in the network diagram in the paper? Surely the 64 7x7 filters in the first layer result in 64 channels, not 192? What am I missing? If it is a mistake (seems highly unlikely), then the question is if there are really 192 filters, or 64.

[@chocorramo99]

64 kernels and there are 3 channels, 192 resulting channels. lol kinda late.

[@Epistemophilos]

@@chocorramo99 Linear algebra is timeless! Thanks :D

[@horvathbalazs1480]

Hi, I really appreciate your work and patience to make this video, however I would like to ask the following: The loss function is created based on the original paper, but the loss for bounding box midpoint coordinates (x,y) are not included because we calculate just the sqrt of width, height of boxes. Am I right?

[@horvathbalazs1480]

Okay, sorry for the silly question. I just noticed that we should not get the squared root of x,y so that's why we skip here: box_predictions[..., 2:4] = torch.sign(box_predictions[..., 2:4]) * torch.sqrt( torch.abs(box_predictions[..., 2:4] + 1e-6) ) box_targets[..., 2:4] = torch.sqrt(box_targets[..., 2:4])

[@soorkie]

Hi, can you do a similar one with Graph Convolutional Networks? Your videos are very usefull ❤️

[@vishalm2338]

Thanks a ton Aladdin for making this video. I truly loved it. Also, Would like to see Retinanet implementation . It would be really fun to watch too. Kudos to you!!

[@zachhua7704]

Hi Aladdin, thanks for the great tutorial. I got a question at 1:13:09, in the paper, authors say the width and height of each bounding box are relative to the whole image, while you say they are relative to the cell. Is that a mistake?

[@jaylenzhang4198]

My understanding of this λ_noob-associated loss function is that it is used to penalize false negatives. This λ_noob-associated loss function includes all grid cells that do not contain any objects but have confidence scores larger than 0. Since there will be a lot of these false negatives, the author adds the coefficient λ_noob to lower their ratio in the overall loss function.

[@Duli998]

I think there might be a bug in the code but I'm not entirely sure, I'd hope you could check it. 1. The original paper defines the predicted width and height to be relative to the whole image. In your code, this is already done as you load the data (you mentioned that annotations are normalized w.r.t. the image size) so why multiply by self.S (number of cells)? Say for instance, given an image of width 600 and the bounding box of width 100 the normalized width of that bounding box is 1/6, given a grid of 7 cells it would ultimately result in your width_cell being = 1/6 * 7 = 7/6. This seems wrong since the width and height should just fall to be 0-1 w.r.t. the image size. I'm wondering if I'm missing something here? 2. The x_cell, y_cell AND width_cell, height_cell go through a different set of transformations, the mid points are multiplied by number of cells and then the cell number is subtracted, the width and height on the other are just multiplied by number of cells in a grid (which itself seems odd and not as according to the paper). This means that (mid-x, mid-y) AND (width, height) are not in the same space which results in a wrong IoU being computed due to the fact that you can't retrieve top left and bottom right corners of the bounding box accurately. Say, let iou_1 = iou(bbox_1, bbox_2) and iou_2 = iou(bbox_normalized_1, bbox_normalized_2) where bbox_normalized_1 = normalize(bbox_1) and where normalize() is the function that transforms (mid-x, mid-y, width, height) with respect to the grid as you did in "Dataset and Implementation" section. If this was correct then you'd expect the iou_1 == iou_2 but that is not the case. Therefore, in the loss when you compute your IoUs to get true box (from 2 candidate boxes) yours numbers might not be correct (although you probably still get which predicted box is the responsible one correctly). 3. In your loss you seem to be missing one important bit. The paper states that the confidence (Section 2) is defined as Pr(Object) * IoU_truth_pred. However, you only compute the loss between predicted confidence and true confidence without the 2nd term ( * IoU_truth_pred). I believe that for a toy example this is fine and you still get valid predictions. I'd appreciate if you could address the above, I'm just learning Yolo and tried to implement it from the paper and found your example which have been extremely helpful. Cheers!

[@poojanpanchal3721]

Great Video!! never seen anyone implementing a complete YOLO algorithm from scratch.

[@AladdinPersson]

...and I understand why :3

[@1chimaruGin0_0]

Great work as always! This video help me a lot to understand my confusion about yolo loss. Could you do some video on Anchors and Focal loss?

[@AladdinPersson]

I'll revisit object detection at some point and try to implement more state of the art architectures and will look into it :)

[@shantambajpai8064]

Dude, this is AMAZING !

[@SamtapesGamer]

Amazing!! Thank you very much for all these lessons! It would help me a lot if you could make videos implementing Kalman Filter and DeepSort from scratch, for object tracking

[@josephherrera639]

Do you mind showing how to plot the images with their bounding boxes (and how that can be applied to testing on new data)? Also, do all images have a maximum of 2 objects to localize?

[@vaibhavsingh1049]

I think there's a mistake in how you rescale the width and height of the bounding box to be greater than 1 because in the paper it stated as follows: "We normalize the bounding box width and height by the image width and height so that they fall between 0 and 1. We parametrize the bounding box x and y coordinates to be offsets of a particular grid cell location so they are also bounded between 0 and 1". See that all the values for the box lie between 0 and 1, x and y relative to the cell, and w and h relative to the entire image size. If I'm wrong please correct me.

[@정현호-u9j]

I agree with this dude

[@NamNguyen-fn5td]

@@정현호-u9j I think in this video he does this because Yolo model make image to 7x7x30 . So does he have to do it so that the label fits the image size ?

[@bhavyashah8674]

Hii @Aladdin Persson. Amazing video. I just have a doubt. While calculating iou for true_label and pred_labels, should we not add the width and height that we clipped when creating true_labels? That is, in case of the example you gave of [0.95, 0.55, 0.5, 1.5], shouldn't we convert 0.95 to 0.95(as the cell we chose is in 0th index along the width) and 0.55 to 1.55(as the cell we chose is in 1st index along the height). This is because we are doing geometric operations like converting x_centre and y_centre to xmin, ymin, xmax and ymax and on not doing the conversion I mentioned, instead of getting the xmin, ymin, xmax and ymax of the bounding box we get some other coordinates instead. Also could you please create the same using Tensorflow?

[@zukofire6424]

Thanks! I don't understand the code regarding the bounding boxes though... Could you do a deep dive into the bounding boxes calculations AND show how to test on a new image?

[@bradleyadjileye1202]

Absolutely wonderful, thank you very much for such a fantastic job !

[@francomozo6096]

Thank you man!!!! Great video! Gave me a really good understanding on Yolo, will subscribe

[@markgazol5404]

Very clear and helpful! Thanks for the videos. I've got one question, though, Can you please explain what is the label for the images with no objects? During the training should it be like [0, 0, 0, 0, 0] or smth?

[@mahdiamrollahi8456]

Hello. Why the target and prediction are in different shapes?

[@sumitbali9194]

Your videos are a great help to data science beginners. Keep up the good work 👍

[@ilikeBrothers]

Просто топчик! Огромное спасибо за столь подробное разъяснение ещё и с кодом.

[@ignaciofalchini8264]

you are awesome bro, really nice job, best YOLOv1 video in existence, thanks a lot

[@R0Ck50LiD-b5z]

Hi, do you have any details on how you prepared the dataset?

[@yantinghuang7491]

Great video! Will you make "from scratch" series video for Siamese network?

[@AladdinPersson]

I'll look into it! Any specific paper?

[@yantinghuang7491]

@@AladdinPersson Thanks Aladdin! This one should be a good reference: Hermans, Alexander, Lucas Beyer, and Bastian Leibe. "In defense of the triplet loss for person re-identification." arXiv preprint arXiv:1703.07737 (2017).

[@santoshwaddi6201]

Very nicely explained in detail.... Great work

[@jeroenritmeester73]

How does the very first layer of the DarkNet with out_channels = 64 produce 192 feature maps? I understand that 3*64 = 192 but I don't really see how that applies. Similarly, the second step has a convlution of 3x3x192, but there are 256 feature maps afterwards.

[@DanielPietsch-o6r]

I am also confused about that part. In my understanding it should be 7x7x3 and then 192 total kernels, right?

[@danlan4132]

Thank you very much!!!! Excellent video!!!! By the way, do you have any tutorials for oriented bounding box detection?

[@mizhou1409]

Great job, very helpful for a new beginner.

[@Wh1teD]

Very informative video and I think I understood the algo but there is one doubt I have: the code you wrote would only work with this specific dataset? If I would want to use a different dataset, would I need to rewrite the bigger part of the code (i. e. the loss function, the training code)?

[@changliu3367]

Awesome video. Pretty helpful! Thanks a lot.

[@anshulgoyal1095]

Works well on Colab GPU. Just need to change the addresses of file references.

[@siddhantjain2591]

Awesome as always! Could you do some video on EfficientNets sometime, that would be great !

[@majtales]

@27:05 why flatten again? Isn't it already flattened in the forward method of the class? Also, do we really need to flatten? @51:22 The MSELoss documentation says it sums over all dimensions by default. Also how did you work around that division by zero?@1:33:15

[@RicardoRodriguez-nn5jw]

Hey man i just found your channel, really good videos. I just saw that you are doing also a tensorflow playlist, are you planning to make maybe a yolo3,4 on tensorflow like this one from pytorch? Maybe common implementations, yolo or mtcnn, pcn? Looking forward to it! Greeeeets

[@nerdyguy7270]

Hi, this is awesome and really helpful. I was going through the yolov1 paper and found that the height and the width are relative to the whole image and not to the cell. Is that correct?

[@qichongxia2110]

very helpful! thank you !

[@vikramsandu6054]

Your name is Aladdin but you are a genie to us. Thanks for this video.

[@larafischer420]

muito boa essa série de vídeos! Vc pode passar as referências q vc usa pra montar esses notes? Tenho dificuldade em encontrar materiais pra estudar

[@buat_simple_saja]

Thank you man, your video help me a lot

[@GursewakSinghDhiman]

You are doing an amazing job. Thanks alot

[@ZXCOLA-z7s]

That’s totally awesome!

[@radoslavstavrev5636]

You are amazing Aladdin, is it possible to run the demo on a video for demonstration purposes?

[@haideralishuvo4781]

FInally , Most waited video , Will have a look asap

[@jitmanewtyagi565]

Broooooo, thanks for this man.

[@dominicyang-y8b]

您好，貌似在数据集方面有一定问题，您直接使用resize方法可能会造成图像的失真，我认为在图像中添加灰条的方式更加合理一些

[@pixarlyVII]

I have a question. At 39:41 you, from utils, import intersection_over_union. I thought that dataset.py, loss.py, ..., utils.py where empty python files. Why did you imported a function from utils.py if in the tutorial we dont code anything in this file? I've followed the tutorial and Im stucked at 59:50 bc my code cant import name "intersection_over_union" from "utils".

[@pixarlyVII]

Nada, soy gilipollas. Me he copiado el archivo utils.py de lo que has subido a GitHub y ya va. It would be interesting to code that part (utils.py) too in the tutorial.

[@NamNguyen-fn5td]

Hi. I have question at 1:12:29. Why "x_cell, y_cell = self.S * x - j, self.S * y - i" minus j and i ? What does this mean?

[@NamNguyen-fn5td]

at 50:27 if you not flatten box_predictions and box_target in MSEloss, it is the same result as flatten

[@leochang3915]

Thank you , you really help me a lot!

[@edelweiss7680]

Hi there !! Great video ! Smth I do not understand : how this YOLO is a "real time" detector with a such large network (if one put 4096 instead of 496 in the FC part) ?

[@proxyme3628]

Regarding the loss for the confidence (FOR OBJECT LOSS part in loss.py), the label Ci should be IoU? In the code, it is (torch.flatten(exists_box * target[..., 20:21]), but because exists_box is target[..., 20:21], it is just a square of target[..., 20:21]? The original v1 paper said "Formally we define confidence as Pr(Object)   IOUtruth pred . If no object exists in that cell, the confidence scores should be zero. Otherwise we want the confidence score to equal the intersection over union (IOU) between the predicted box and the ground truth", which suggests the Ci_hat is to be calculated from IoU.

[@yanxu4968]

one question about the unit test or integration test. There is quite some code for Yolo, I am not at all confident that my code is working correctly even following this great video. So, do you guys spent time on unit test/integration test in real-world scenarios? If yes, how to do it?

[@venkatesanr9455]

Kindly do other versions and SSD that will be helpful. Shall I know your future plans? Like Computer Vision or others

[@AladdinPersson]

Right now I don't have any clear plans, will see what interests me and what people wanna see

[@venkatesanr9455]

@@AladdinPersson Also include different yolo versions that will be exciting

[@lucaluca5154]

Hi Aladdin, im new to Yolo and it is until now that i found your channel. I wanna ask why nothing happened after I run the train file, it started then it finished but no picture showed up. Hope you can reply bro.

[@alanjohnstone8766]

Wonderful! A minor quibble, wand h are proportions of the main image not the cell in the original paper.

[@piyushjaininventor]

You are right. its mentioned in paper.

[@PaAGadirajuSanjayVarma]

I am glad I found your channel

[@nova2577]

Appreciate your effort!!

[@wuke4231]

thank you for your video!😘

[@淮都先生]

many thanks!!

[@mbgdemon]

This project is immensely helpful. I am working on doing it myself, however, and I have noticed some issues with your implementation: - You implement target confidence wrong. It's supposed to be iou_maxes, not exists_box (but to be fair the paper does not make this very clear). - The reduction="sum" for MSELoss is a bad idea. It links learning rate to batch size, which you do not want. Change it to reduction="mean" and use the LR specified in the paper.

[@AladdinPersson]

Could you do a PR? Not exactly sure how iou_maxes would look like

[@mbgdemon]

@@AladdinPersson Sure, I will make sure my version works properly before I suggest any code changes though. On target confidence, I am referring to this sentence of the paper: "Otherwise we want the confidence score to equal the intersection over union (IOU) between the predicted box and the ground truth" and the surrounding section. They are saying that C hat for a responsible predictor is not 1 but rather equal to the current IOU of the predicted box. I admit this is weird, since it trains one output against the success of another output, but it's what they do. Not sure yet if it is important for performance. These are small details that I only noticed because I'm trying to do the full training; I couldn't have gotten this far without your video!

[@mbgdemon]

@@AladdinPersson I submitted a writeup on your issues page

[@apunbhagwan4473]

He is simply Great

[@patloeber]

Amazing effort!

[@AladdinPersson]

Thank you:)

[@pphuangyi]

Thanks!

[@shenbin2930]

When I use the code, the detection accuracy of the training set is very good, but the detection accuracy of the test set is almost equal to 0, which is obviously overfitting. In fact, the original code is to train an overfitting model, but I have modified some of the code. Why is it still overfitting? I have made the following modifications： nn.Dropout(0) -> nn.Dropout(0.5) WEIGHT_DECAY = 0->WEIGHT_DECAY = 2e-4 This question has bothered me for a long time. I would appreciate it if you could answer it.

[@FanFanlanguageworld1707]

How many images you trained with?

[@m4gh3]

I got the same results, I too am trying to understand what is going on Also I can overfit with a way smaller network

[@sb-tq3xw]

Amazing Work!!

[@canyi9103]

4:24, In paper the width and height are predicted relative to the whole image. they can not be larger than 1, but in your video, you said it can larger than 1. It seems not right

[@usmaniyaz1059]

Hi Aladdin! Your work is awesome. Hey, I have a query I am splitting my image 3000x 2000 into 1024x1024 patches along with bounding boxes. Now I want to get back the original size of the bounding box relative to the original image. Yolo 7X7 grid was somewhat analogy to that but still not able to figure out how to get the original bounding box. Any suggestions? This is just a preprocessing step. Kindly help

[@amartyabhattacharya]

One question that I have is, How can I get to know the coordinates of the grid cell of which the centers are a part of? Is it like (1,1) of the output prediction gives the prediction for grid cell having two endpoints as (0,0),(64,64) ? (448/7 = 64)

[@NityaStriker]

Hi. I’m unable to load the PascalVOC_YOLO dataset within a Colab notebook due to the dataset being private. Is there a way to use the dataset in a Colab notebook without downloading it on my laptop ?

[@AladdinPersson]

I'm not sure, I think you need to download it. Isn't there a way to upload the dataset to Colab so you can run it?

[@NityaStriker]

@@AladdinPersson There is, but my internet connection is not the fastest while having a small data cap which is why I usually use !wget within colab itself. In this case, both the !wget command and Kaggle’s command failed within Colab for the Kaggle file after which I wrote the above comment. Later, I copied the code from the get_data file, pasted it onto a cell, added a few lines of code for creating 8examples.csv and 100examples.csv, and ran it for the code to work.

[@BENHARARVIND]

Brother please help If the program detects i want an alarm But I don't know what to write in 'if' condition (what will be name of the detected image inside the box from the video) How can I know the name of the detected frame

[@КонстантинМогилевский-о2л]

Really great tutorial, but why do we do flatten in forward pass of the Yolov1 and also nn.Flatten() in the _create_fcs? Isn't it redundant?