jrheiner/thesis-clip-geoloc-country

Model Card for Thesis-CLIP-geoloc-continent

CLIP-ViT model fine-tuned for image geolocation. Optimized for queries at country-level.

Model Details

Model Description

• Developed by: jrheiner
• Model type: CLIP-ViT
• Language(s) (NLP): English
• License: Creative Commons Attribution Non Commercial 4.0
• Finetuned from model: openai/clip-vit-large-patch14-336

Model Sources

• Repository: https://github.com/jrheiner/thesis-appendix
• Demo: Image Geolocation Demo Space

How to Get Started with the Model

from PIL import Image
import requests
from transformers import CLIPProcessor, CLIPModel

model = CLIPModel.from_pretrained("jrheiner/thesis-clip-geoloc-continent")
processor = CLIPProcessor.from_pretrained("jrheiner/thesis-clip-geoloc-continent")

url = "https://huggingface.co./spaces/jrheiner/thesis-demo/resolve/main/kerger-test-images/Oceania_Australia_-32.947127313081_151.47903359833_kerger.jpg"
image = Image.open(requests.get(url, stream=True).raw)
choices = ["Botswana", "Eswatini", "Ghana", "Kenya", "Lesotho", "Nigeria", "Senegal", "South Africa", "Rwanda", "Uganda", "Tanzania", "Madagascar", "Djibouti", "Mali", "Libya", "Morocco", "Somalia", "Tunisia", "Egypt", "Réunion", "Bangladesh", "Bhutan", "Cambodia", "China", "India", "Indonesia", "Israel", "Japan", "Jordan", "Kyrgyzstan", "Laos", "Malaysia", "Mongolia", "Nepal", "Palestine", "Philippines", "Singapore", "South Korea", "Sri Lanka", "Taiwan", "Thailand", "United Arab Emirates", "Vietnam", "Afghanistan", "Azerbaijan", "Cyprus", "Iran", "Syria", "Tajikistan", "Turkey", "Russia", "Pakistan", "Hong Kong", "Albania", "Andorra", "Austria", "Belgium", "Bulgaria", "Croatia", "Czechia", "Denmark", "Estonia", "Finland", "France", "Germany", "Greece", "Hungary", "Iceland", "Ireland", "Italy", "Latvia", "Lithuania", "Luxembourg", "Montenegro", "Netherlands", "North Macedonia", "Norway", "Poland", "Portugal", "Romania", "Russia", "Serbia", "Slovakia", "Slovenia", "Spain", "Sweden", "Switzerland", "Ukraine", "United Kingdom", "Bosnia and Herzegovina", "Cyprus", "Turkey", "Greenland", "Faroe Islands", "Canada", "Dominican Republic", "Guatemala", "Mexico", "United States", "Bahamas", "Cuba", "Panama", "Puerto Rico", "Bermuda", "Greenland", "Australia", "New Zealand", "Fiji", "Papua New Guinea", "Solomon Islands", "Vanuatu", "Argentina", "Bolivia", "Brazil", "Chile", "Colombia", "Ecuador", "Paraguay", "Peru", "Uruguay"]
inputs = processor(text=choices, images=image, return_tensors="pt", padding=True)
outputs = model(**inputs)
logits_per_image = outputs.logits_per_image # this is the image-text similarity score
probs = logits_per_image.softmax(dim=1) # we can take the softmax to get the label probabilities

Training Details

The model was fine-tuned on 177 270 images (29 545 per continent) sourced from Mapillary.