[MUSIC] We saw how deep features were great for
image classification. They allowed us to get quite high
accuracy, even with just a little bit of training data on a particular
classification task by first finding features from this deep learning model
that won the 2012 ImageNet competition. So, that was pretty exciting. But if you remember
the motivation at the beginning, kind of the sample use case at
the beginning of this module, we're talking about finding
images of products that I like. You know, I like shoes, I'm gonna find
shoes that's similar to my shoes. I want to find images that are similar
to in the product set I care about. So instead of doing an image
classification task, now we're gonna take those deep features
and do an image retrieval task. So let's go ahead and do that, okay? So we're going to name our notebook, so I'm gonna talk about Deep Features for
Image Retrieval. And I'm going to hide the header and
hide the toolbar, so we can go ahead and
get a little bit more space. And I'm gonna fire up graphlab create and I'm also going to load the CIFAR-10 dataset, which is the same
dataset that we loaded earlier when we talked about deep
features for classification. So in this case I'm just going to
load the training set because we're just going to do retrieval
in the training set. So let's start from there. So I'm gonna call it image_train and I'm going to use graphlab.Sframe and then as input I'm going to
give it the image_train_data/. And just as a little reminder,
if I look at this image_train data set, .head, in this s frame,
we have the image id, the actual image,
the label which we're not gonna use for retrieval, we're not gonna
use that label at all. The image array of the pixels,
which we're also not gonna use, but it already has the deep
features precomputed. In the notebook on deep features for
classification, we looked at those lines that you need
to compute that column, deep features. Here, I'm just gonna use
those deep features. [MUSIC]