shiertier / ccip_merge

Merge algorithm of ccip embeddings

ccip_merge

Merge algorithm of ccip embeddings

Install

git clone https://github.com/narugo1992/ccip_merge.git
cd ccip_merge
pip install -r requirements.txt

Best Solution

New update in 2024.4.28: Solved, the best solution is just aligning the norm of embeddings, and get their center point. Like this

import numpy as np


def ccip_merge_func(embs):
    lengths = np.linalg.norm(embs, axis=-1)
    embs = embs / lengths.reshape(-1, 1)
    ret_embedding = embs.mean(axis=0)
    return ret_embedding / np.linalg.norm(ret_embedding) * lengths.mean()

Run Experiment

Target: FIND OUT AN EMBEDDING CLOSEST TO THE GIVEN EMBEDDINGS

Go and change your code in ccip_merge/measure.py

def ccip_merge_func(embs):
    # TODO: just replace your function here!!!
    pass

You need to output a merge embedding (shape: (768, )) based on the input embeddings (shape: (n, 768)). And then the merged embedding will be used to calculate the distance between the input ones with imgutils.metrics.ccip_batch_differences function.

The current ccip_merge_func is a default one based on scipy.optimize.minimize. We think you must be able to write a better one.

After you complete these, just run

python -m ccip_merge.measure

100 randomly selected fixed character tags will be used as the test cases. And the result will be saved at test_result.csv.