What is/was your first thought on Few-Shot Learning? Update 03/22/2022: AttributeError in training part is fixed. Check the Colab code link in the description.
Thanks for the detailed explanation. 👍 The video is really helpful. 🙂 Soon, I will be starting my master’s project to classify soil classes in Germany, during excavation by an autonomous excavator using Few-Shot Learning technique.
@@keertiprem5718 Appreciate the kind words. 🙂 Good luck with your project. It sounds very interesting. I would love to read your project when it’s done.
Your video and the resource has been tremendously helpful. Thanks a lot. I have some certain questions, can you please share your email address? Thank you.
Thanks for the kind words. Somehow I missed your comment. I would prefer you to comment here if you have any questions so that others can benefit from our conversation. Or, here is my Facebook page: facebook.com/LCVPage
Thanks for the great explanation. question : if we are doing episodic training and creating a large number of few-shot task to train the prototypical network so it mean it also require a large amount of labeled data .how we can say that few-shot learning need less amount of labeled data ? Please guide
Hi, Wonderful video !. Couple of questions please !. 1. Does the data should be balanced for every class ?. If yes then how it is few shot ?. 2. In each episode, there should be image of all classes in query and support set ? Thanks
Yes, you can.But I'm skeptical about the performance of ML algorithms using meta-learning on complex NLP tasks. The recent papers are written on neural networks or transformers. But you can definitely try.
Great video! I have one question. Can you use few-shot learning also in combination with 1d CNN's? I have some acceleration data I want to classify but my "traditional" CNN's have a very hard time doing that because I (intentionally) don't use a lot of data. I don't want to use a classical machine learning algorithms like random forest and few-shot learning looks very promising so far
Thank you. Yes, you can use it for your case. I have only seen people using FSL in computer vision and NLP. But I don't see why you can't use it in other domains, people may already have. ✌️
Hi, Can u provide an example for few shot learning, for object detection. May I know the time it takes to complete the learning? Is it faster than training the model?
great!, I have one question. i want to use your model with my personal data, but im faced with a problem. i want to orgainze the train_set and test_set by my personal data, not Omniglot data, how can i modify that part..? plz give me your wisdom thanks.
Thank you. I’m sharing two links here, the first one is for PyTorch dataloader for the few-shot learning and the second one is for TensorFlow. You may have to modify it for your case. Hope it helps. ✌️ github.com/sicara/easy-few-shot-learning/blob/master/easyfsl/samplers/task_sampler.py github.com/schatty/matching-networks-tf/blob/master/matchnet/data/mini_imagenet.py
Hi, Thanks for the great video but I don't understand one thing, we are training the network with 5 images/support set, 5 images/query set, and 40,000 tasks so that means we are using 400,000 images but omniglot dataset consists of 1623 images each class so for 5 classes the total will be 8115 images, how are we having the 40,000 tasks or the images can be repetitive in the task? Also, I have a dataset that consists of 100 good images and just 4 or 5 bad images so should I use 2-way-1-shot approach?
Hello, Thank you. Yes, there are image and class repetitions in training tasks. But the meta training classes/images are different than the meta testing classes/images. Btw the Omniglot data has 1623 alphabets from 50 different languages, and each alphabet has 20 images. So total number of images is 32460. Could you please explain a little more what do you mean by you have good images and bad images?
@@LightsCameraVision I have 2 kinds of vehicle parts each part has 2 types good part and faulty part, the problem is that for faulty part I have just 5 images and good part are around 700. Therefore, I am looking for some solutions which require less training data. Do you think few shot will work in this case?
5 is very low. Of course, the best solution is to get more images if possible. Otherwise, you can look for data that is similar to what you are trying to do then do the meta training on that, and then meta test on your data. You can also look at zero-shot learning.
Hey one question: A lot of the times I see the terms "episodic learning" and "meta-learning" used interchangebly in the context of few-shot learning. are there any substantial differences between those terms or are they identical in this context?
Appreciate the kind words. ✌️ Yes, you can. Meta learning algorithms are model agnostic in general. There are some work on this. Check these out. arxiv.org/pdf/2203.01482.pdf edoc.ub.uni-muenchen.de/24557/1/Probst_Philipp.pdf ieeetv.ieee.org/video/meta-algorithms-in-machine-learning
This tutorial is very helpful, but I still have some problems. I tried to run it on Colab, but it seems that there are some problems when using TaskSampler and I cannot figure it out, can you check it out?
At a high level, the goal of both triplet loss is same as metric learning. We use them to learn a representation function. But they are doing it in different ways. Triplet loss is usually used in self-supervised learning where it learns by comparing represented vectors. Whereas metric learning algorithms learn a function that maps instances to a new space and later on a test instance is projected on that space and classified based on the closest distance to a learned class. Other types of meta-learning algorithms are different than construction learning (triplet loss).
Hey! It's a great video. I'm trying to classify some house images with this method. My images are in google drive. It'll be great if you can make a video on how to use custom data from drive!
Hello sir, I tried to run your code in Google Colab with omniglot dataset, it worked fine, but I couldn't make it work with different dataset (FashionMNIST or other dataset) it shows error. I know a bit of python and new to FSL, will you please guide me to understand your code properly to run with other dataset.
Hello, I'm a little busy right now with some deadlines. I'll look into the issue soon. Please comment the error you are getting when you tried it with Fashion MNIST. Thanks!
Hello, your video is really useful! But i have a question, im doing a proyect about image classification of problems in Wind Turbines! I have 3 classes, with 3 diferentes problems like impactLightnings/pitting/fissure in each classes i have 7 images, its ok?
![[Few-shot learning][2.2] Prototypical Networks: intuition, algorithm, pytorch code](http://i.ytimg.com/vi/rHGPfl0pvLY/mqdefault.jpg)
What is/was your first thought on Few-Shot Learning?
Update 03/22/2022: AttributeError in training part is fixed. Check the Colab code link in the description.
Thanks for the detailed explanation. 👍
The video is really helpful. 🙂
Soon, I will be starting my master’s project to classify soil classes in Germany, during excavation by an autonomous excavator using Few-Shot Learning technique.
@@keertiprem5718 Appreciate the kind words. 🙂 Good luck with your project. It sounds very interesting. I would love to read your project when it’s done.
Your video and the resource has been tremendously helpful. Thanks a lot.
I have some certain questions, can you please share your email address? Thank you.
Thanks for the kind words. Somehow I missed your comment. I would prefer you to comment here if you have any questions so that others can benefit from our conversation. Or, here is my Facebook page: facebook.com/LCVPage
Genius.. The way you explain this is epic
Finally found an algorithm with intelligence than usual object detection algorithms. Thanks a ton.
great explanation👌, thanks a lot. Hope to see more videos about few-shot learning
Appreciate the kind words. Thank you. ✌️
Superb....Thank you for such a great knowledge sharing video.
Appreciate the kind words. 🙂✌️
Great! you have made my day. Thanks a lot for this wonderful video!
I’m glad it helped you. Thanks for the kind words. ✌️
thank you so much, it was perfect.
Appreciate the kind words. Thank you. ✌️
really fantastic video. but how can I use my own dataset? (load into modal and train-set split). please give the details or example. thank you..
Thank you sir for making such a wonderful video.
Appreciate the kind words. Thanks for watching. ✌️
Thanks for the great explanation.
question : if we are doing episodic training and creating a large number of few-shot task to train the prototypical network so it mean it also require a large amount of labeled data .how we can say that few-shot learning need less amount of labeled data ? Please guide
Hi, Wonderful video !. Couple of questions please !.
1. Does the data should be balanced for every class ?. If yes then how it is few shot ?.
2. In each episode, there should be image of all classes in query and support set ?
Thanks
can I apply Nlp with meta learning without use deep learning , I mean machine learning algorithm ,nlp, meta learning?
Yes, you can.But I'm skeptical about the performance of ML algorithms using meta-learning on complex NLP tasks. The recent papers are written on neural networks or transformers. But you can definitely try.
Thanks for this video, I need help for object detection code of few shot learning
Great! Thank you ! can you please make a video for Few-Shot with Graph Neural Network? :)
had a query, can i use augmentation like this
[Main iman]: [90 degree rotation, 180 degree rotation, flip by two axes, top hat bottom hat]?
Great video! I have one question. Can you use few-shot learning also in combination with 1d CNN's? I have some acceleration data I want to classify but my "traditional" CNN's have a very hard time doing that because I (intentionally) don't use a lot of data.
I don't want to use a classical machine learning algorithms like random forest and few-shot learning looks very promising so far
Thank you. Yes, you can use it for your case. I have only seen people using FSL in computer vision and NLP. But I don't see why you can't use it in other domains, people may already have. ✌️
Hi, Can u provide an example for few shot learning, for object detection. May I know the time it takes to complete the learning? Is it faster than training the model?
Is the model able to retain the knowledge that it gained from past or the model able to predict the new dataset image only
great!, I have one question. i want to use your model with my personal data, but im faced with a problem. i want to orgainze the train_set and test_set by my personal data, not Omniglot data, how can i modify that part..? plz give me your wisdom thanks.
Thank you. I’m sharing two links here, the first one is for PyTorch dataloader for the few-shot learning and the second one is for TensorFlow. You may have to modify it for your case. Hope it helps. ✌️
github.com/sicara/easy-few-shot-learning/blob/master/easyfsl/samplers/task_sampler.py
github.com/schatty/matching-networks-tf/blob/master/matchnet/data/mini_imagenet.py
Hi,
Thanks for the great video but I don't understand one thing, we are training the network with 5 images/support set, 5 images/query set, and 40,000 tasks so that means we are using 400,000 images but omniglot dataset consists of 1623 images each class so for 5 classes the total will be 8115 images, how are we having the 40,000 tasks or the images can be repetitive in the task?
Also, I have a dataset that consists of 100 good images and just 4 or 5 bad images so should I use 2-way-1-shot approach?
Hello, Thank you. Yes, there are image and class repetitions in training tasks. But the meta training classes/images are different than the meta testing classes/images. Btw the Omniglot data has 1623 alphabets from 50 different languages, and each alphabet has 20 images. So total number of images is 32460.
Could you please explain a little more what do you mean by you have good images and bad images?
@@LightsCameraVision I have 2 kinds of vehicle parts each part has 2 types good part and faulty part, the problem is that for faulty part I have just 5 images and good part are around 700. Therefore, I am looking for some solutions which require less training data.
Do you think few shot will work in this case?
5 is very low. Of course, the best solution is to get more images if possible.
Otherwise, you can look for data that is similar to what you are trying to do then do the meta training on that, and then meta test on your data. You can also look at zero-shot learning.
Hey one question: A lot of the times I see the terms "episodic learning" and "meta-learning" used interchangebly in the context of few-shot learning. are there any substantial differences between those terms or are they identical in this context?
Thanks for explaining, great video, can I apply random forest with meta-learning?
Appreciate the kind words. ✌️
Yes, you can. Meta learning algorithms are model agnostic in general. There are some work on this. Check these out.
arxiv.org/pdf/2203.01482.pdf
edoc.ub.uni-muenchen.de/24557/1/Probst_Philipp.pdf
ieeetv.ieee.org/video/meta-algorithms-in-machine-learning
This tutorial is very helpful, but I still have some problems.
I tried to run it on Colab, but it seems that there are some problems when using TaskSampler and I cannot figure it out, can you check it out?
I'm glad that you found it helpful. Thanks for pointing out the error. I have fixed it. Check out the updated Colab link in the video description. ✌️
@@LightsCameraVision Thank you!!
How to run EasyFSL code on my own customize dataset??
is meta-learning different from triplet loss?
At a high level, the goal of both triplet loss is same as metric learning. We use them to learn a representation function. But they are doing it in different ways. Triplet loss is usually used in self-supervised learning where it learns by comparing represented vectors. Whereas metric learning algorithms learn a function that maps instances to a new space and later on a test instance is projected on that space and classified based on the closest distance to a learned class. Other types of meta-learning algorithms are different than construction learning (triplet loss).
Hey! It's a great video. I'm trying to classify some house images with this method. My images are in google drive. It'll be great if you can make a video on how to use custom data from drive!
Thank you. I’ll try to make one or share some resources in a few days.
@@LightsCameraVision That'll be great :)
Hello sir, I tried to run your code in Google Colab with omniglot dataset, it worked fine, but I couldn't make it work with different dataset (FashionMNIST or other dataset) it shows error. I know a bit of python and new to FSL, will you please guide me to understand your code properly to run with other dataset.
Hello, I'm a little busy right now with some deadlines. I'll look into the issue soon. Please comment the error you are getting when you tried it with Fashion MNIST. Thanks!
Hello, your video is really useful! But i have a question, im doing a proyect about image classification of problems in Wind Turbines! I have 3 classes, with 3 diferentes problems like impactLightnings/pitting/fissure in each classes i have 7 images, its ok?