Transfer Learning for Computer Vision and Keras (9.3)

[@pankeshpatel6835]

Thank you so much for sharing these videos, Prof. Heaton!!

[@MrStrangerinthewind]

Thank you Jeff, your videos is amazing. I think your videos are super good and fun. ha would like to see your future fun projects

[@AYMENCHANNEL4U]

Thank you so much prof, i have one problem when i put my path i got error. i do not how i can fix it

[@michaels8297]

Hi Jeff, thank you for your videos. I would love to hear your opinion on the fast.ai library and if it is necessary to do transfer learning on a GPU or if it is fine to do on my mac laptop or something like a linode server that uses CPU. When I say "do" i mean train the model on jupyter notebook from the GPU, linode server, or mac. Lets say training set is roughly 1000 total photos of 3 classes.

[@bestbuddy777]

Thank you so much sir you are great inspiration for me keep doing well sir :)

[@mujdatvolkan]

Thank you so much for these materials. I have a question. Why did you set the first 20 layers as untrainable and the last 20 layers as trainable? I was expecting the top 2 layer would be trainable and the rest would be untrainable.

[@MrShooter00000000]

Good question! I had the same one

[@sidesplitter9497]

@@MrShooter00000000 its because the model he transfered only had 20 layers. He then added 2 layers on top so he is only setting the top 2 layers as trainable

[@Tinigala]

Hi Professor, I have an error that occurs when I try to retrain the model. My kernel dies everytime at the first epoch. I've reduced the epochs to 5 and the same problem happens. What can I do ?

[@shivvratpandey4362]

Isnt Flatten() layer necessary??

[@aa-xn5hc]

Great educational material, and also funny!

[@F4-Free_Fire_For_Fun]

Great this man's video was recommended for me by a AI Bot

[@lakshay510]

Awesome Explanation, but what to do if input image is of less size like in CIFAR-10 we have 32x32 dimension but these nets are trained for imagenet which takes 224x224 so how can we use Transfer Learning for CIFAR-10?????

[@alibahari4217]

Yea he is really really awesome.. You don't get it exactly.. There is a part which is "img.resize((IMAGE_WIDTH,IMAGE_HEIGHT),Image.ANTIALIAS)" that changes any picture in any size into the appropriate size.. As a result you can input any dataset you want to this network

[@lakshay510]

Ali Bahari it’s been 7 months since I asked this question and now I think that using other method will just add padding to the image and result won’t be correct. We can’t use 224X224 for 32x32 anyhow. Either we have to find a model which is trained on 32X32 or we can use an existing architecture to retrain the whole model.

[@AmeerulIslam]

it's hard to follow you sir. ! a lot of building blocks are unexplained!