customize metrics to .fit

[2533245542]

is there a way to provide customized metrics to .fit function?

import pandas as pd
from sklearn import datasets
from openTSNE import TSNE
from scipy.spatial.distance import euclidean
import numpy as np
import matplotlib.pyplot as plt

# a. use the iris dataset
iris = datasets.load_iris()
x, y = iris["data"], iris["target"]
year = np.arange(x.shape[0]).reshape(-1, 1)
data_with_time = np.hstack([x, year])


# b. use my dataset
# embeddings = np.load("asdf.npy")
# metadata = pd.read_csv("asdf.tsv", sep="\t")
# year = metadata['year'].values.reshape(-1, 1)
# data_with_time = np.hstack([embeddings, year])

def custom_metric(v1, v2):
    feature_dist = euclidean(v1[:-1], v2[:-1])  # non-time dis
    time_dist = abs(v1[-1] - v2[-1])  # time dis
    return feature_dist + time_dist

embedding = TSNE(
    metric=custom_metric, 
    perplexity=30, 
    n_iter=500, 
    random_state=42
).fit(data_with_time)

this script seems to be work with sklearn's iris dataset but after switching to my dataset, it says error

  File "anaconda3/envs/tsne/lib/python3.10/site-packages/openTSNE/nearest_neighbors.py", line 247, in build
    self.index = AnnoyIndex(data.shape[1], annoy_metric)
TypeError: argument 2 must be str, not function

[dkobak]

I think it is possible to use custom metric with PyNNDescent, but you will probably need to do it outside of openTSNE, and then pass the kNN graph into openTSNE as PrecomputedNeighbors.

[2533245542]

I am able to install opentsne in a editable state with pip install -e ..

Do you think if there is a way to modify the source code directly so I can customize the metrics in a more straightforward way?