Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Cluster images by fc10 output - display images closest to cluster center #17

Open
etrain opened this issue Dec 10, 2013 · 5 comments
Open
Assignees
Labels

Comments

@etrain
Copy link
Collaborator

etrain commented Dec 10, 2013

Cluster images by fc10 output - display images closest to cluster center.

@ghost ghost assigned etrain Dec 10, 2013
@etrain
Copy link
Collaborator Author

etrain commented Dec 10, 2013

ModelStats now has a top_k_images_by_cluster property - a matrix whose entries A[i,j] correspond to the image id of the j'th ranked image in cluster i, according to (euclidean) distance to cluster center. We are clustering FC10 outputs of all the images.

@etrain
Copy link
Collaborator Author

etrain commented Dec 12, 2013

Done (thanks josh!) - not currently very useful. Planning to see if we can get better clusterings if time. If not - worth showing in the paper anyway.

@etrain etrain closed this as completed Dec 12, 2013
@etrain
Copy link
Collaborator Author

etrain commented Dec 12, 2013

actually, reopening as bug

@etrain etrain reopened this Dec 12, 2013
@etrain
Copy link
Collaborator Author

etrain commented Dec 13, 2013

We are now clustering by "probs" - once #17 is fixed, this will hopefully look better.

I have also added a property called "cluster_centers" to model stats that contains the centroids for the clusters. This could be useful for display somehow.

@JoshRosen
Copy link
Collaborator

Even after fixing #19, the final clusterings still don't seem very intuitive:

image

I might try using cosine similarity instead of Euclidean distance to see if that gives us more interesting results.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
Projects
None yet
Development

No branches or pull requests

2 participants