Week 4 (July 18th - July 22nd)

July 19th:

In our meeting my Professor advised that I create a full table of CLIP ViT-L/14, ViT-L/14-336px, ViT-B/32, … (i.e. all the open CLIP models published) as well as start creating a directory of dataset for Politicans from the internet.

For our project we decided to have a Politicians dataset which included:

• 100 mayor from the 100 most populous setings
• All 436 U.S. House of Representatives members
• All 100 U.S. Senators

Basically, I went on Wikipedia and bgean downloading all images of politicians and putting them in a folder on the server. In order to download all these images at once I created a python script which pulls images from webpages.

Here are some example photos that where downloaded (images downloaded where in smallest format):

July 21st:

As we brainstormed more we came up with two rough draft ideas 1) Similar to Winoground where we have 2 pictures and 2 captions – and we compute the same metrics as Winoground – This is a picture of a congressperson from Texas – This is a picture of a congressperson from Minnesota 2) Design a different tests like the ones you were doing before with words. Input 1 image – and 6 captions – [This is a picture of a congressperson] [This is a picture of a US representative in the Senate] [This is a picture of a Professor] [This is a picture of an assistant] [This is a picture of a receptionist]

July 22nd:

Today I started going through all the CLIP models and running the Winoground dataset on them. Here is a summary of my results for each different model:

Upcoming

• Start uploading all Politicans datasets to the scratch folder
• Keep Brainstorming Tests for Politician Dataset
• Start Tests on Politician Dataset