Slides from the TensorFlow meetup at eBay NYC 06/07/2016 based on my blog https://medium.com/@st553/using-transfer-learning-to-classify-images-with-tensorflow-b0f3142b9366
9. How to train?
Google Inception-v3
Pre-trained on ImageNet
http://download.tensorflow.org/models/image/imagenet/inception-2015-12-05.tgz
Convolutional Neural Nets
Source: https://research.googleblog.com/2016/03/train-your-own-image-classifier-with.html
10. 1,000,000 images, 1,000 categories
Business office Internet site President
Image Source: http://image-net.org/explore
11. Why use a pre-trained model?
It’s faster (it’s pre-trained)
It’s cheaper (no need for GPU farm)
It generalizes (avoid overfitting)
Image Source: http://image-net.org/explore
16. Performance
About 1 second per image
Gilt has ~40,000 images
~12 hours
Embarrassingly Parallel -> Partition across threads / machines
Cache pool_3 features (numpy array)
17. Applying pool_3 features
Nearest Neighbor Search
Brute Force
Cosine Distance
Checkout https://github.com/spotify/annoy
Train a classifier
Nearest Neighbor Classifier
18. You don’t need to use TensorFlow for everything
Reading files / Work Queues
Image resizing
Serializing weights