TensorFlow Tutorial 13 - Data Augmentation

[@AladdinPersson]

I was inspired and learned the basics of TensorFlow after I completed the TensorFlow specialization on coursera. Personally I think these videos I created give a similar understanding but if you wanna check it out you can. Below you'll find both affiliate and non-affiliate links, the pricing for you is the same but a small commission goes back to the channel if you buy it through the affiliate link which helps me create more future videos. affiliate: bit.ly/3JyvdVK non-affiliate: bit.ly/3qtrK39

[@antonisvenianakis1047]

This channel is going to be a gold mine, thanks a lot..!

[@apocalypt0723]

this playlist is so good

[@indieelectroworks]

Very useful hands-on tutorial. Thanks man!

[@eduardolopes7678]

Man, you are the best ! Thanks so much.

[@라봉한-h8f]

in your code there's a typo ds_test is mapped from ds_train. i guess it's not what you've meant to do. Github repo is same too.

[@sameepshrestha7285]

so when we call our augmentation function here do we get an increase in the dataset or the same image is augmented differently

[@AladdinPersson]

Every image is augmented sequentially and "on the fly", so we increase our dataset although we do it implicitly. Let's say we use a random rotation of -30 to 30 degrees of the original image, then it's difficult to say how much we've increased our dataset. We do not store anything to disk but every iteration the image will be randomly rotated between [-30, 30] degrees, and in theory this would be infinitely many different images.

[@sameepshrestha7285]

@@AladdinPersson oh thanks since i saw only one image going through the function at a time and the function returning single image, i was a bit confused

[@mitchellphiri5054]

I think Francois Chollet said the augmentation is done on the GPU if you include it in your model. He posted the pros on his Twitter

[@AladdinPersson]

I actually read it just after I made the video, but I'm still unsure as to the efficiency compared to the first method I showed. I'm thinking if it's running on the GPU we're wasting computation that would otherwise be spent training our model, because prefetching and doing augmentation on the CPU is done while the model is running on the GPU and in this way we should never be waiting for input to the model.

[@superaluis]

@@AladdinPersson Jeremy Howard from fastai has some discussion about it on the first chapters of his Fastbook. I think it's just faster when the preprocessing is done on GPU, but considering your GPU has enough memory. However you have a very good point. I'd love to see an article or video discussion about this topic.

[@Klinsmann02]

Hello. Thank you very much for the quality of your videos, it really helps! Question: when you include data augmentation in your model, does it apply on the test set too, so during evaluation? EDIT: just found the answer on the tensorflow website: "Data augmentation is inactive at test time so input images will only be augmented during calls to model.fit (not model.evaluate or model.predict)".

[@aldonin21]

So does your function have a probability of 50% to horizontally flip an image? Or it will 100% of the times horizontally flip the images in batch for all the epochs? If so, then how to write my own custom function where i specify all the augmentation I'd like to add and the PROBABILITY, so that each epoch i can have different augmented version of my original image.

[@EugeneGolovachev]

Is there a way to enlarge my dataset with augmentation and tf.data API on fly? Or I need to save new images to disk first and then create dataset?

[@AladdinPersson]

The way I showed in the video is data Augmentation on the fly. This is the recommended way rather than saving to disk

[@EugeneGolovachev]

@@AladdinPersson I meant if we want to create 10 new images from 1 on disk. Can we do this with map, or we need to construct generator only?

[@riis08]

@Aladdin Persson: Amazing videos, thanks for these amazing videos. I have gone through lot of tensorflow courses (tensorflow specialization from deep learning, udacity, tensorflow courses), but your videos are the best. Already passing to my friends, hope they will pass it forward, and hope you will have much bigger audience, in future.

[@AladdinPersson]

That means a lot to hear, thank you so much :)

[@imanefahim8557]

@Aladdin Persson Is there a way to visualize the data augmentation in python jupyter please?

[@라봉한-h8f]

And there's one question. we are mapping data with augmentation fuction we made. and then in our model using augmentation layer. so aren't we augment two times?

[@abdulbareehussein787]

when i evaluted my model , i become good accuracy but a big loss . is this overfitting . Acc =0.90 and loss = 9.0

[@ramezanifard]

Hi. Thanks for the video. In the video, you used tensorflow 2.3.1 for data augmentation. The highest version of tensorflow on my Pycharm is 2.1.1 (for gpu). Would you mind explaining how you managed to install 2.3.1 for Pycharm? I appreciate that. Thanks.

[@RajatSharma-ct6ie]

Hi Aladdin, thanks for the wonderful tutorials. I have another requirement in which in have multiple inputs: [ds_1, ds_2] and output as y1 , so ds_train = ([ds_1, ds_2], y1). Also, the augmentation is applied only on ds_1. Where the count of ds_1 & ds_2 is equal. How to pass it in Model.fit ()? Any references or suggestions?

[@wolfisraging]

Great video, and Wow Wow :), I didn't know that we can set the augmentation in the form tf.keras.Sequential, this is much more elegant and simple... like "torchvision.transforms.Compose" in pytorch. Though I wonder what would happen if we evaluate our model... cuz now the augmentation is the part of the model, so I doubt that every time the prediction is made ... the model will be applying that augmentation on inference inputs too, which is not supposed to happen. Any thoughts?

[@AladdinPersson]

From what I've read it's disabled during evaluation of our model

[@wolfisraging]

@@AladdinPersson Awesome then!!!