File size: 1,500 Bytes
4f732a3
1a9bde9
 
 
 
4f732a3
a4a1c67
4f732a3
 
a4a1c67
4f732a3
 
1a9bde9
 
 
 
 
 
 
 
af2f1d5
 
1a9bde9
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
---
title: CLIP Model Evaluation
emoji: 📊🚀
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.31.0
app_file: app.py
pinned: false
license: apache-2.0
---

# 📊 CLIP Model Evaluation Space

This Space provides an interactive interface to evaluate the performance of various CLIP (Contrastive Language-Image Pre-Training) models on standard image-text retrieval benchmarks.

It calculates Recall@K (R@1, R@5, R@10) metrics for both:
*   **Image Retrieval (Text-to-Image):** Given a text query, how well does the model retrieve the correct image?
*   **Text Retrieval (Image-to-Text):** Given an image query, how well does the model retrieve the correct text description?

A higher Recall@1 means the model is better at placing the correct item at the very top of the results.

## 🚀 How to Use

1.  **Select a CLIP Model:** Choose a pre-trained CLIP model from the dropdown menu.
2.  **Select a Dataset:** Choose the dataset you want to evaluate on (e.g., "mscoco", "flickr").
3.  **Number of Samples:** Specify the number of image-text pairs from the dataset to use for the evaluation. Using fewer samples will be faster but less representative.
4.  **Click "Evaluate Model":** The evaluation will run, and the Recall@K metrics will be displayed.

## 🛠️ Under the Hood

This Space uses the `evaluate` library from Hugging Face and a custom metric script (`clipmodel_eval.py`) to perform the CLIP model evaluations. The models and datasets are loaded from the Hugging Face Hub.