Data augmentation to address overfitting | Deep Learning Tutorial 26 (Tensorflow, Keras & Python)

When we don't have enough training samples to cover diverse cases in image classification, often CNN might overfit. To address this we use a technique called data augmentation in deep learning. Data augmentation is used to generate new training samples from current training set using various transformations such as scaling, rotation, contrast change etc. In this video, we will classify flower images and see how our cnn model overfits. After that we will use data augmentation to generate new training samples and see how model performance improves.

#dataaugmentation #dataaugmentationdeeplearning #addressoverfitting #cnn #deeplearningtutorial

DISCLAIMER: All opinions expressed in this video are of my own and not that of my employers'.

[ayoolafakoya]

You are a gifted teacher. I easily understand any topic you teach. Thanks

[andrewzolensky]

You are hands down the best teacher I've found on youtube for deep learning and coding. I've spent hours and hours trying to figure this stuff out, and you just make it so simple and elegant. Thank you good sir.

[acrux]

nice video, I've struggling in data augmentation for a long time T_T. Now I found how to use cnn with data augmentation, I hope I can create my own deep learning model using this method. thanks bro

[sandiproy]

Your teaching makes things so simple. Thank you, sir.

[vikashdas]

Video by video this tutorial is getting more awesome, excellent teaching and very calm explanation by our Guruji

[leamon]

Thanks sir, looking forward to more in-depth content in the upcoming videos.

[raghavendrac]

use layers.RandomFlip(), layers.RandomRotation(), instead of using . This fixes the issue of module not found

[eyerusalemtsehaye]

I am getting an error while scaling the train and test set.
Unable to allocate 35.7 GiB for an array with shape (3258, 700, 700, 3) and data type float64
How can I solve it?

[codinghighlightswithsadra]

great job, I've found your channel recently it's helped me a lot, thank you so much🍀

[pabloreynoso]

Awesome content man!! I have liked all your tutorials and I started to follow you on youtube!!👌🏽

[pallavipriyadarshini]

sir..i am a big fan of urs..ur codes are very simple and easily understandable..sir please do some videos on image segmentation in medical imaging.

[hardikvegad]

Sir can you please explain RNN and LSTM just like you explained ANN....Please Sir :)

[NguyenNhan-ygcb]

you should save this video for us, so I can learn it later lol :)) Bye thanks so much

[wussboi]

Hi sir, for the stack of Conv2D layers...are there any general rule-of-thumb for selecting the number of filters? 
Whether its recommended to go in an increasing order (like you did), decreasing order or constant?

[MarcusVinicius-curd]

Can someone explain me this: After the data aug, the model.evaluate function returns 070~0.75 accuracy, but when using classification_report(y_test, y_pred_classes), i get 0.51 as accuracy. Wasn't it supposed to be the same as model.evaluate()?

[primeprime]

Thank you sir. I have one doubt, what % of the image dataset should be augmented to generate a good model?

[yccdavis]

It was just so helpful!!! Thank you so much!!!

[kazifahimlateef]

Sir, in data_augmentation section, while running this cell, it says "img_height" is not defined!!
How to get rid from this problem?
Thnks for the great tutorial

[anirbandas]

Hi, during the first process of trainning the model, i used as loss function and the accuracy was terrible (~20-30%)

but when i used (as shown in video), the accuracy was so much better (~99%)

can someone explain what is the difference between these two losses? (irrespective of having the same name)

[siddharthsingh]

The part were image is getting zoomed in and out is not working for me and while plotting it is not showing and difference. Is anyone facing the same issue ? Sometimes it is working sometime it doesn't.