142 - Multilabel classification using Keras

Code generated in the video can be downloaded from here: github.com/bns...

[@ameyparanjape3292]

I was curious about how do we go about transfer learning/fine-tuning in a multi-label classification problem (for example using pre-trained Keras/TF hub models)?

[@ruchirsrivastava4828]

May be late but you can have a look at kzbin.info/www/bejne/h3Orm6OgrtKkmK8

[@johniwuozor9015]

I love this video, so clear. Thanks a lot!

[@knonline]

Very intresting topic. Crisp and clear. Just wanted to understand more about: How do we understand actaully applicable label for the image? If the probability is more than 50% (0.5) should it be declared as present ? something like that ? Please reply when you get a chance. Thank you.

[@marawanahmed7829]

Great video. I am wondering about how we can do the same exercise but assuming we have three classes for every label, something like (yes, no, unknown), or (high, low, intermediate), thanks!

[@darshanhubbly9978]

Use CategoricalCrossentropy instead of BinaryCrossentropy mostly!

[@betulkara2245]

Amazing video, very clear and helpful. Thanks

[@DigitalSreeni]

You're welcome!

[@yaminadjoudi4357]

thank you, please if the images are in subfiles ? how can i gather it in just one folder with python ?

[@KazemAbdolpour]

Question: I want to classify a dataset that has 6 classes. My main goal is to be able to use this built model to identify or classify samples that contain combinations of 2, 3, 4, 5, 6 of these primary classes. Without training the model with combinations of 2, 3, 4, 5, 6. For example, when I give this model an example that actually contains 4 classes, the model should be able to determine the presence of these 4 classes more than the other classes. How can I do this? How will this dataset be labeled?

[@arslanmasood8355]

Thanks for the great tutorial. There are some amazing kaggle datasets in the biological domain for example Human protein atlas dataset. It would be great if you could use them to explain these kinds of problems

[@chevalharrichunder8953]

DigitalSreeni, thanks for this, i have learnt a lot. Just wanted to find out on how to do a multilabel for Images that have features with values from a scale of 0-5, your example was binary(0 or 1) or true and false.

[@DigitalSreeni]

Once way to do this is by binarizing your multi-class images so you only have one class at a time. At the end combine the results. Another way is to define a new architecture for this specific task.

[@chevalharrichunder8953]

@@DigitalSreeni thanks so much. Appreciate the feedback.

[@gpiano7410]

Currently I want to build my own model, but my question is how create dataset on multi label classification? specifically what is the best way to divide image dataset for each class labels? for my understanding in multi class problem is simply give it equal number of image per class, but on multi label?

[@shaelanderchauhan1963]

Your channel is really great but please add time stamps as well!

[@anthonieschaap1625]

The dataset doesn't have the .csv metadata file anymore, only text files for each year

[@junaidlatif2881]

Very good explanation... ❤

[@nataliabravo1102]

Thank you very much for this video! I was wondering how could I add data augmentation to a code like this one? I'm running a pretty similar model and my model is overfitting

[@aminadjoudi5163]

hi can i contact you please ?

[@ariouathanane]

amazing work. Thanks a lot

[@venkateshsadagopan2505]

I have seen you were using batchnorm after activation function. Is it right way to use batchnorm. I have seen people use batchnorm before activation function. Can you clarify on the same .

[@DigitalSreeni]

Batchnorm is used to normalize values for mean and variance. It should not make a big difference if you use it before or after normalization. In fact, it makes more sense to use it after activation but it is conventional to use it before. In this video I may have experimented with putting batch normalization before and after but forgot to change things before recording the video. But, it should not make much of a difference.

[@venkateshsadagopan2505]

@@DigitalSreeni Thanks for letting me know.

[@koteshwarraomaripudi1080]

good work. keep going. All the Best

[@DigitalSreeni]

Thanks a lot

[@thomasv860]

Thank you sreeni..

[@garyhuntress6871]

Very informative. "binary cross entropy......... please make peace with that." I literally laughed out loud :D

[@DigitalSreeni]

:)

[@dingowhiz481]

@@DigitalSreeni Thanks for the invaluable videos. I used'binary crossentropy' in one of the models and it output only two repeated probability values. when I used 'categorical crossentropy', the accuracy dropped and loss skyrocketed. What should I do next?

[@sanjaisrao484]

Sir how to handle class imbalance in multilabel classification please help.

[@fouried96]

Great video! Thank you!

[@mtahirrasheed5538]

Your lectures help me a lot to learn about neural networks. Can you make a video lecture to input two images to two pipelines of a neural network without using concatenation or share some link related to it. Thanks a lot

[@stefanAH97]

Amazing content, god explanation, learn too much, thnx a lot

[@DigitalSreeni]

Glad you liked it!

[@gueshdagnew7658]

Nice job, Sir. Kindly, can you share the CSV file here? Because I couldn't get the ready-made CSV file from the link you provide as they are providing separate txt files. Thank you.

[@DigitalSreeni]

Check now

[@DigitalSreeni]

It should be under other_files on my github page.

[@rephlanca]

Hi there! Thank you so much for the video. I'm trying to run this on my own set of training data and I get the error "ValueError: logits and labels must have the same shape ((None, 25) vs (None, 8))" when trying to fit the model. My X_train appears as (700, 200, 200, 3) and my y_train appears as a size of (700, 8). Do you think I did something wrong when setting up the data frame?

[@arsalanzaid05]

Make sure your last layer( output layer ) has same number of nodes as your target labels.

[@rephlanca]

@@arsalanzaid05 Thank you so much! That was the problem!

[@arsalanzaid05]

@@rephlanca Do you know how to calculate Confusion matrix, precision and recall for this same problem ? He did not perform any performance metrics apart from 'accuracy'.

[@oladosuoladimeji370]

Hi, Thanks for the tutorial, I am getting 'nan' as the validation loss any idea of what it means or what the error is?

[@ramuyk]

Hi i tried training for same dataset with same code, but my accuracy isnt going above 30%. Please let me know why and how to improve it.

[@DigitalSreeni]

Using same data set with same code and parameters should yield statistically similar results. Something else may be wrong, please check.

[@mhsiao13]

Same here... copying the code as is and running it give a final val_accuracy of ~0.2 ???

[@Macias096]

can you upload this meta data file? because now, posters are separated by year of its release, and meta data is also separated in txt files. And I would like this yours csv file.

[@chournsolidet5406]

By following the code from your github repo, I keep training with no raise of accuracy. loss: 0.2748 - accuracy: 0.2972 - val_loss: 0.2525 - val_accuracy: 0.1900 Could someone help me on this?

[@arsalanzaid05]

Remove dropout from sequential model, and train with large epochs

[@chournsolidet5406]

@@arsalanzaid05thanks, sir. Let me try this.

[@scienceaddicted3431]

Thank you for your video! very useful

[@DigitalSreeni]

You are welcome!

[@xuzhang7443]

What is the size of y?

[@javeriaehsan369]

how can i label normal nd abnormal from an ecg signal?

[@shan_alex]

Is it possible to do a multi label classification using bert

[@oladosuoladimeji370]

Thanks for the tutorial Can you is it to have feature selection/extraction for multilabel classification. For example: images of mixed dish (e.g. boiled egg, meat, rice etc in a plate) for multilabel classification. Thanks

[@syedfahadsaif3121]

hello, I checked on the website for the dataset but the dataset isnt available in a combined csv file, can you send it to me?

[@gauravjha8986]

amazing work. Thank you for it.

[@DigitalSreeni]

I’m glad you liked it.

[@bhuvaneshs.k638]

@@DigitalSreeni why are u using Binary cross entropy loss ??? This is Multi-label... Isn't it categorical cross entropy. I tried Binary Cross Entropy and I got very high accuracy but in Categorical I got reasonable results for Chexpert Dataset

[@slimhamdi3383]

Great Job Bro !

[@DigitalSreeni]

Thanks 🔥

[@HabtamuDesalegn]

Very helpful Thanks!! Can you do video on Image segmentation using Spectral Clustering? Cause I learn alot from your lecture on Image segmentation using Kmeans.

[@DigitalSreeni]

What do you mean my spectral clustering? Do you have spectral images with a full spectrum at every pixel? If so, you can consider each channel in the spectrum as a feature and train a machine learning algorithm (e.g. Random forest) to segment.

[@HabtamuDesalegn]

@@DigitalSreeni Thank you for your reply, If I can clarify any idea on image segmentation by graph partitioning like normalized cut , dominant sets,...

[@roofatbt738]

Thank you Man !

[@MS-mr3yg]

great work 👍

[@DigitalSreeni]

Thank you.

[@syedfahadsaif3121]

my accuracy is very low, even though I loaded the yearly json files and multi hot encoded the genres

[@AbdullahKhan-uw8mo]

thanks sir. but plz upload csv file as there is no csv file.

[@raspberrypi6669]

Thanks a lot sir!

[@shilpashivamallu9056]

Hello Sir, Do you have any videos for Multilabel classification using Metadata instead of Image classification

[@DigitalSreeni]

No, but the approach should be similar as long as you can get your data organized.

[@priyankadurairajan887]

Nice, I am using Multilabel classification on air pollution dataset in csv format. My dataset is imbalanced , how to solve it.

[@DigitalSreeni]

I'll try to do a video on imbalanced datasets. But for now, search for 'over sampling'. Many people get creative in over and under sampling and also augmenting under represented data.

[@priyankadurairajan887]

@@DigitalSreeni thanks for your suggestions

[@aminadjoudi5163]

@@DigitalSreeni did you dot it sir ?

[@arsalanzaid05]

Thank you, very helpful but could you please let me know how to solve precision, recall, F1 score, and Confusion matrix by your code for this problem ?

[@aminadjoudi5163]

hi do you know how ?

[@arsalanzaid05]

@@aminadjoudi5163 As far as I know, confusion matrix cannot be calculated for multilabels, as there are many labels so there would be many confusion matrix as well.

[@aminadjoudi5163]

@@arsalanzaid05 i work on multilbael classification of X-ray pathologies and i try to solve this and choose the right metrics for this, and solve the problem of unbalanced dataset, i think the binary cross entropy and Macro F1-score are good

[@arsalanzaid05]

@@aminadjoudi5163 can we have Zoom meeting if you don't mind? We can discuss the entire thing in the meeting ?

[@arsalanzaid05]

@@aminadjoudi5163 add me on Instagram - @iarsalanzaid LinkedIn - Arsalan Zaid We can discuss the timing over there

[@saraal7493]

je veux savoir ou son les attributs? et classes ?ou bien sont tous des attributs ???

[@natcolley4878]

Q1: @7:34 image size 200x200 I get. What is the 3 for? Q2: you turned a list of len=2000 into a numpy array. Could you use a generator here? Q3: @10:00 'I hate Windows' So do I! Why are you on it?!

[@DigitalSreeni]

Q1: The 3 represents the 3 channels, Red, Green and Blue. Q2: Not sure of your question. I turned a list to array to pass it through the model. You can use generator to generate new data using a set a rules you define. Q3: I have a love hate relationship with Windows. Every system has strengths and flaws and given everything I like to work with Windows. My order of preference for general work (including coding) is Windows, MAC and Linux. But I hate Windows for certain tasks like opening images or smart-search for files.

[@kaustubhkalyankar8795]

Hello sir, i copied ua code but it's giving me syntax error in for loop.

[@chaitanyakalagara3045]

How to do it using ImageDataGenrator. I have a huge dataset and when I try adding the images to array it is running out of RAM

[@DigitalSreeni]

Just define your ImageDataGenerator for training and test datasets and while training use fitgenerator instead of fit. Of course you define a batch size to make sure your system can handle the batches. The ImageDataGenerator comes with excellent documentation that explains these steps. Also, you can look up my video 128 to see how I implemented.

[@ArunKumar-sg6jf]

how to create own cnn model in keras

[@DigitalSreeni]

Please watch my videos on Deep Learning. I covered many topics, including putting together your own network.

[@kannanv9304]

So many take-aways from this too.......I wanted to ask you, How is your Dad now????

[@DigitalSreeni]

Dad is recovering, out of danger. Thanks for asking.

[@kannanv9304]

@@DigitalSreeni Good to hear that Ajarn-Sreeni.......Ajarn is the word we use in "Thai Language" for Professor......The root word is "Acharya" from Sanskrit........Your tutorial and Harsha Bogle's Cricbuzz talks are like same side of the coin for me........Can't miss one word of it or you may miss to grasp the vital cues.........Thank you so much for all these meticulously prepared tutorials........On Apeer, I am learning to annotate OME.TIFF files for a different application, by following your Videos.......

[@AIinAgriculture]

Respected sir, How to create bounding boxes for each class in an image? What type of models should I use for such tasks?