TensorFlow Tutorial #08 Transfer Learning

How to use the pre-trained Inception model on the CIFAR-10 data-set using Transfer Learning.

This tutorial does NOT work with TensorFlow 2 and later versions, and it would take too much time and effort to update it. But Tutorial #10 is similar and it does work in TensorFlow 2.

[hvasslabs]

There is now an easier way of doing Transfer Learning in TensorFlow which is explained in Tutorial #10. This tutorial is still useful for understanding the concept, but Tutorial #10 is easier to use in practice.

[burobernd]

Dude, this is an amazing tutorial. Very concise. Thanks!

[jokmenen_]

Very good video! I never thought of using PCA for understanding the encodings.

[gw]

Hi Magnus, your tutorials are excellent, I just vote for your pull request, a Big Thank to you!

[bilginaksoy]

This tutorial is the best tutorial about the transfer learning. Thanks.

[mertsefatarhan]

I could run this code before. Then I tried running it in conda environment and dang! AttributeError: module has no attribute '_VARSCOPE_KEY'. I know it took you 3 full days to make it work, before (read your answer in github). So, the best choice is using Tutorial 10. I admire your determination. Thank you Hvass.

[lankanathaekanayake]

why we can't change fc layer like(2048, 2) and train directly with softmax classification? instead of saving to file and using PCA

[amelhouzi]

hi this code is writhing with python3 . plz what is the difference if I use python2.7 and tensorflow only CPU. Thanks.

[tusharsinghal]

Hi, The tutorials are gr8 and accurate.I have a question, the dataset that you are using is of cifar-10 and when i looked into the data, it is in encoded form. Can we use inception model to use our custom data with custom class name. I tried using bazel build but i am not able to do so.

[directvoffshore]

I am working on object detection API using tensorflow. I trained the API using to detect custom object. But after training, the API detects only the custom object and not the objects for which it is already trained. Is there any way to add more classes to an already existing model. model has 90 classes, i want to add 2 more classes so that it becomes 92.

[darkcrusader]

Hello Magnus!
I have a dataset of MRI scans of the brain. Would it make sense to use the inception model to train this dataset? None of the labels in the inception dataset are related to MRI scan images, so my first guess is that it wouldn't help. Nevertheless, I wanted your take on this.
Thanks for the awesome videos!

[hajeressefi]

Great tutorial! I have one question... If I want to use this model to classify an image on my computer, what do I do after calling the transfer_values() function explained in 19:14 ? The result of this function is the transfer values of my input image. How do I feed my classifier those transfer values?

[pkprobot]

Thanks Magnus for putting together a wonderful tutorial on Tensorflow. My question is, what is the *performance of transfer learning* with respect to train the system from scratch. I know from scratch it will take a monster server. But let's say if I am doing *transfer learning* for one class with 2000 images *without GPU* . How much *~time* will it take? Waiting for your response.

[BestMusiCollection]

what if I want to perform transfer learning in my own model ?

[ashitguy]

Can I train the inception model on characters like 0 to 9 and A-Z?? I want to use it for a license plate detection software.
Is there any other model which would be apt for this?

[remariorichards]

Good day, question what exactly are transfer values? are they like the various soft-max classified values.

[JHauris]

Magnus, I am getting a kernel error when I try to run the pca.
transfer_values_reduced =
says "Kernel died and will restart automatically" and I need to rerun the notebook. Do you k now what is causing this? I tried going from 3000 to 300 samples and got same error. Thanks, Jon

[thisiscrispin]

Amazing tutorial! Does this still work with the new version of tensorflow? (>1.0)

[rebeenali]

have used transfer learning for unsupervised learning?

[chushekwai]

If i have very larger dataset, does this implementation have memory problem? It seems that all the transfer value feature vector is loaded into the main memory.