TensorFlow Tutorial #02 Convolutional Neural Network 

ONE WORD! EXCELLENT! Far superior to Udacity's DL course which only provides confusion. Same with the rest, I care about concepts a lot as implementation is too basic once you get the fundamentals of what you are implementing. Excellent work sir!

One of the highly underrated conv neural net videos on the internet.

Udacity DL student here and this is gold! thinking about posting it in the forums

Thank you for your effort into creating this tutorial series for tensorflow!

I was having the trouble of my life understanding the TF business but thanks to you I now understand. Thank you sincerely

Wow, your work is great. Helper function you made gave us the intuition of Convolution Neural Networks. I wanted to see filtered images by convolution layers. And you showed me them. Thank you!

You made the most intuitive insight on CNN. Well documented! Thanks for working on this.

Best CNN tutorials I've ever seen so far. So good!

I've learned more from a single one of your tutorials than hours upon hours digging through Tensorflow documentation. I honestly wish you could have been there for the development of Tensorflow making design decisions. I'm sure it'd be much more intuitive to use and better documented.

Great and in-depth descriptive video on Covnets!

Best tutorial I ever watched on this topic.Thank you so much.

hands down the best tutorials out there. Thanks a lot!

This tutorial > any paid Deep learning and CNN tutorial ever! Best Cnn tutorial on youtube!

you are a true savior for me man, wish the best always come for you.

You explain the basics so well making the whole experience almost painless. The discontinued libraries make getting started a pain but I guess nothing stands still in this AI world. Well done sir and a donation will be on the way to my favorite charity.

Amazing! Thanks a lot! You, Andrej Karpathy's cs231n class and Michael Nielsen's book are helping me way too much! Thanks again for this incredible tutorial.

This is simply excellent. Very professional. Thank you so much

this is wonderfully exlained.. thank you so much

Great tutorial in understanding MNIST.

I fell in love with that code because it's so well commented hahaha, thanks so much for the tutorial!

Best tutorial on the internet!

This is definitely the best explanation I have ever seen !!! Thank you so much for the effort !!!

Took me a week now I'm not getting frustrated by those errors and it feels good to get rid of them

Awesome tutorial for self-learners. Thank you

Thank You. This is the best tutorial

Thanks a lot these video series are magnificent! Helping me a lot for my Udacity deep learning foundations nanodegree where they took the videos from the free Udacity DL course everyone is complaining about...

Excellent, explanation of using Tensor flow for CNN .. thanks

Thank you. Very beautiful explanation.

THANK YOU. So clearly described.

Thank you so much for these videos !

Great Great Lectures... I have learned a lot

The best explanation of CNN. Thank you

Amazing tutorial, thanks!

You Sir are a Genius...!Thank you for this video.

Great tutorial!

Super good explanation.

Very very nice. I'm not sure that I can understand the rest of these tutorials. At least I understand the first tutorial and I think I must leave a comment here and say thank you very much. This tutorial is really detailed

EXCELLENT...EXCELLENT...EXCELLENTTTTT...explanation! sir thank you so muchhhh .

YOU ARE A LEGEND! THANK YOU FOR THIS AWESOME SERIES! I WILL DONATE MONEY TO OUR LOCAL CHARITY!

Very well done. This helped me enormously.

Excellent tutorial. Thanks.

Thanks for the useful tutorial

Great Tutorial. Thanks

The thing I want to know, I get it with in 4 min... So man you earn subscription with in 3min... Thanks for your good information...keep going...

I need help on how can I use the same CNN for regression problem, like for given input image output should edge detected . What kind of cost function I need to use and how should be predict . Thanks

I want to try my own data set images as training and testing. Any idea to do that?

Thankyou so much... Really awesome code.

Your tutorial are great

Amazing! Thank you so much for your video. I think Google should buy this tutorial for MNIST! :D

how can i create my oun training dataset? howcan i feed the images to network?

How would you get 14*14 output from a 5*5 pixel slide on 28*28 pixel ? - i got the same 28 pixels snce after each calculation one pixels is formed and one is neglected.

Hello. Thank you for the tutorial - I wish I started my learning with it. Just wanted to ask you if you know any good references for the specifics on how to take the trained model and implement it. What I mean is once the models was trained, we get sets of filters for each of the layers and apply them to the input image. You do briefly explain how this is done, but I can not find a good tutorial that does in depth - I am rather interested in implementing the model (let's say in hardware) instead of designing or training it. For example, in layer 2, when the first pixel of each input is convoluted with its corresponding filter we then use Relu on each result or add them up first? what happens if the value is > 255? Another question is zero padding - should it be done on the inputs or outputs? Any info would help. Thank you again.

Question: in the 2nd convolutional layer, are the filters being used in parallel for each of the input feature maps equal to each other?

what a f* great tutorialllllllllllllll. Thank you so much.

layer_conv2,weights_conv2 = \ new_conv_layer() Since the output of a convolutional layer is a 4d tensor..how can u assign it to 2 variables ? plz reply asap

how will I get the output for the predicted input?

Den bedste tutorial med teori og praktiske lektioner. Mange tak. Fra Ballerup

thank you ,that is so useful

Amazing！thanks sir!

Why is padding set to "SAME" in the maxpool unit? Doesn't this prevent the dimensionality to decrease?

why did you set parameters in filters with randomly values ? we want to extract features for example when we want to detect straight line the filter will set as straight lines of ones and -1s and so on

Awesome, respect man!

How to implement next_batch on your own dataset??

how do i train this model on my data set.??

i have one doubt regarding the model, there are models like QuocNet, AlexNet, Inception, which models is included in this tutorial

Where are these filters used in the Convolutional layers derived from? They appear to have arbitrary shapes inside them, is that the case or do they have some purpose?

Hi! Regarding why not use relu in the last layer, is it because it will overlook some features?

Hi, why choose not to use a hidden layer? What the the use cases for using hidden layers?

If you're working with a dataset like, say, CIFAR-10, would you have to account for the multiple channels?

Great tutorial is there is any for LSTM

You are a blessing :)

This is the best lecture series on tensorflow I have come across. Can you please cover capsule nets in the upcoing tutorials?

In the flattern part,why is there 128 features?

weights should be the number of dimensions and not the filter_size right? For example, in my data the no. of dimension is 28*28 , but the filter size is 3*3 and no of filter is 32 shape = [ number of dimensions in input, number of dimensionsin input, num_channels, no. of filter] shape = [28, 28, 1, 32 ]

Your videos are awesome.... Thanks for that... i have one question that if i passed the document which have all the two,three and four digit number then how should i predict that all number if i passed that document image??

one word : best

really awesome

Excellent

Hi, how to get the code for this tutorial? Please let me know.

What is the type of the model? (LeNet-5 , AlexNet, VGGNet, GoogLeNet or ResNet)

Good tutorial sir, am wondering where can i find the solution for these exercices.

How do I load my own data? I scanned my own letters, resized to 32X32 and saved them as csv file with 1024 numbers in each line. Thanks in advance.

Awesome tutorial. But, have a question reg. how the last convolution layer gets mapped to the fully connected layer(s) and then that gets mapped to the output. I completely understood how the filters (features) get applied on the input picture to obtain the intermediate images. But, the last layer, where 36 filters get applied, how does these filters are applied to the fully connected layers? Is it just more filters get applied on these intermediate images (the output of the last convolution layer)? If yes then what is the difference? And how does the last layer in the fully connected layer gets mapped to the output image - i.e., how does the CNN knows to match the images to the proper image? I am a bit confused there.

Thank you sir :)

thank u sir for ur great video !!! i just wanna know how can i test that on one image that i will give it to ..plz anny help

cls_pred = np.zeros(shape=num_test, dtype=np.int32) getting below error at the above-mentioned line, Kindly help ValueError: sequence too large; cannot be greater than 32

I'm just curious why only 36 output channels? Why not more or less?

original input image is of 28*28 pixels. then we have convolution layer1 with 16 filters of 5*5 and stride is 1 (according to the explanation from video). So the output of the first convolution layer supposed to be 14*14*16. Can anyone explain me why this is 14*14*16.? Even if I take the stride as 2, it should be 12*12*16....

Why did you use -1 while reshaping x_image

Will you be doing capsule networks ?

Excellent video! And I want to ask, what does weights mean in plot_conv_weights function?

Who gives the filters for the very first convolutional layer of the very 1st sample batch??

Are there any resources out there to make your own data set just like MNIST or CIFAR 10. I am new to the field. Will appreciate your help.

thanks

awesome！ awesome！awesome！

I M Getting Error in loading Input data @ 36:19 Error : - cannot find reference input_data PLz Some1 Help me out

how did the 28x28 pixel end with 14x14 through convolutional layer 1? What I expect is a 24x24 output.

Hi! Thanks for this tutorial.. cleared many doubts. one thing I want to ask is that u said after convolution layer we use rectifier to remove negative values and make them zeros, but I am here in doubt that from where do those negative values come in the first place?. I would be really thankful if u would clear that..

thanks so much

from mnist import MNIST data = MNIST(data_dir="data/MNIST/") this part is not working having some problem with __init()__ data_dir is invalid plzz helo