InceptionV3 | Inception Network

[@KOTESWARARAOMAKKENAPHD]

I am Full Time Research Scholar in VIT-AP University ,Your tutorial is beneficiary for my research

[@CodeWithAarohi]

Glad it helped!

[@siddharthmodi2740]

Your channel is an underrated gem💎 Thank you mam 💛

[@CodeWithAarohi]

Glad my video is helpful! Keep watching 😊

[@yahiarafik9965]

Thanks for the effort you put in, i had an image classification task for a small dataset and thought of using Inception for it, this was really helpful!

[@CodeWithAarohi]

Glad it helped!

[@arunagarnipudi4475]

Simple and awasome, Thanks for the video!

[@CodeWithAarohi]

Welcome

[@alimaan7581]

Thank you so much, this video was very helpful for me.

[@CodeWithAarohi]

Glad it was helpful!

[@AyushSharma-of9en]

awesome and thank you for clearing my all concept

[@CodeWithAarohi]

Glad it helped you

[@MrRabeaahmed]

thank you for your great explanation 👌👍

[@CodeWithAarohi]

Welcome

[@shilpsshilpa1285]

Could you please make a video on visualizations of feature maps and filters in CNN?? ur explanations are spot on. Thank you

[@CodeWithAarohi]

Sure , Will do soon

[@vijayak7308]

Excellent explanation, thank you very much Aarohi.

[@CodeWithAarohi]

Glad you liked it

[@prasadinipadwal3641]

superb explanation

[@CodeWithAarohi]

Thank you so much 🙂

[@rahmamahgoub4363]

Excellent explanation, Thanks

[@CodeWithAarohi]

welcome

[@radhikapatil8003]

Amazing explanation ❤️ Thanks lot👍 mam

[@CodeWithAarohi]

Welcome 😊

[@NakibAkash]

Can you please make a video on Squeezenet architecture?

[@trangle1506]

well presented. thanks a lot

[@CodeWithAarohi]

Most welcome!

[@photogenicgestures3199]

Arohi cana you please tell me how the image sizes are 149*149*32 and 147*147*32 and 147*147*64 after first conv of 3*3

[@satyajeetshashwat4115]

((299-3)/2) +1 =149 and 32 is the number of filters, just apply the formula : ((input - filter +2*padding)/stride)+ 1

[@zakariaelalaoui2813]

Good work, thanks

[@CodeWithAarohi]

Thanks for watching!

[@nabeelaashraf3518]

Well Explained

[@CodeWithAarohi]

Thanks!

[@littlespace28]

what is the reason for not implementing the upper path between Inception Block B and Reduction Block B (in which we have avg pool , Conv, FC, FC) and also the reason for not implementing the FC between Global average Pool and softmax

[@CodeWithAarohi]

Sorry I am not getting what exactly you were asking

[@littlespace28]

@@CodeWithAarohi What is the reason for not implementing the auxiliary classifier ?

[@KOTESWARARAOMAKKENAPHD]

good explanation madam

[@CodeWithAarohi]

Thank you so much 🙂

[@meyouanddata9338]

awesome :)

[@CodeWithAarohi]

Thankyou

[@yuchenzhao2663]

really appreciate your work, big thanks! I have one question, when we factorizating 5x5 to two 3x3, the number of parameters are reduced, but if we slide the input from previous activation using 3x3 filter, it would be more multiplications than using 5x5 filter, so is this factorization make the convolution less computation effecient even with fewer parameters?

[@CodeWithAarohi]

Thankyou for appreciating my work . Now coming to your question- See always output of previous layer become input to next layer and when output of previous layer comes to next CNN layer where filter size is 5*5 then the number of parameters would be more as compare to 2 cnn layers with 3*3 filter size. because of this there would be less computation with fewer parameters.

[@yuchenzhao2663]

@@CodeWithAarohi If we use 5x5 filter to convolve a for example 5x5 input to a 1x1 output, we will need 25 multiplications. However if we use two 3x3 filter instead, we will need to do 3x3x9 multiplications in first 3x3 filter convolution and 9 multiplications in second 3x3 filter convolution to do the same thing as the original 5x5 filter does. Is this suggest this factorization lead to more computational expensive? Thanks for repling. I understand the parameter is less after this factorization, but if the computational efficient is reduced why would we do this? Or if there are some balance between those two metrics?

[@rafabaranowski513]

@@yuchenzhao2663 If you have input picture 100x100x1 and mask 5x5 stride 1, result will be 96x96x1 which gives 96x96x25 = 230400 multiplications. Inf you have mask 3x3 stride 1 output size will be 98x98x1 which gives 98x98x9 = 86436 multiplications. It reduces number of parameters and also number of multiplications.

[@yuchenzhao2663]

@@rafabaranowski513 Got it! Thanks!

[@PLimbu-gh3vf]

can you please share the presentation

[@rijulsingh9803]

Great great great explanation and really intuitive Although I have one doubt, 5x5 convolutions were being used for feature extraction at a larger scale so how is that going to work out with two 3x3 convolutions?

[@anwaarkhalid3508]

Can you please share the link of the original paper?

[@KOTESWARARAOMAKKENAPHD]

please provide the lecture videos on topics Deep Dream ,Deep Art , Fooling Convolutional Neural Networks

[@CodeWithAarohi]

I will try.

[@KOTESWARARAOMAKKENAPHD]

@@CodeWithAarohi thank you for your response

[@ragibshariar7789]

Could you please give me a summary article of inception v3 architecture ??

[@CodeWithAarohi]

Sorry, I don't have such article

[@darshansr9025]

Ma'am great explaination! if you don't mind could you share the ppt for the whole CNN playlist. Thanks in Advance.

[@ahmedhusham7728]

Can you please explain what is "RMSProp Optimizer" in Inception-v3 and where it found?

[@CodeWithAarohi]

Sure, Will do a video on that soon

[@leelamanasachidella3396]

Mam can you please make a video on architecture description of Inception resnet v2

[@CodeWithAarohi]

Will try to cover this topic

[@paninilal8322]

Nice

[@CodeWithAarohi]

Thanks

[@Fatima-kj9ws]

Great

[@anjithaanju2584]

Can you make a video about Inception v4?

[@CodeWithAarohi]

Sure