ann-unsplash-25k / main.py
ashvardanian's picture
Add: multilingual index (#5)
909641a
raw
history blame
2.84 kB
#!/usr/bin/env python3
from os import PathLike, listdir, remove
from os.path import isfile, join, exists
from mimetypes import guess_type
from base64 import b64encode
import pandas as pd
import numpy as np
from PIL import Image
from PIL import ImageFile
from tqdm import tqdm
from uform import get_model
from usearch.index import Index, MetricKind
from usearch.io import save_matrix, load_matrix
ImageFile.LOAD_TRUNCATED_IMAGES = True
def is_image(path: PathLike) -> bool:
if not isfile(path):
return False
try:
Image.open(path)
return True
except Exception:
return False
def image_to_data(path: PathLike) -> str:
"""Convert a file (specified by a path) into a data URI."""
if not exists(path):
raise FileNotFoundError
mime, _ = guess_type(path)
with open(path, 'rb') as fp:
data = fp.read()
data64 = b64encode(data).decode('utf-8')
return f'data:{mime}/jpg;base64,{data64}'
def trim_extension(filename: str) -> str:
return filename.rsplit('.', 1)[0]
names = sorted(f for f in listdir('images') if is_image(join('images', f)))
names = [trim_extension(f) for f in names]
table = pd.read_table('images.tsv') if exists(
'images.tsv') else pd.read_table('images.csv')
table = table[table['photo_id'].isin(names)]
table = table.sort_values('photo_id')
table.reset_index()
table.to_csv('images.csv', index=False)
names = list(set(table['photo_id']).intersection(names))
names_to_delete = [f for f in listdir(
'images') if trim_extension(f) not in names]
names = list(table['photo_id'])
if len(names_to_delete) > 0:
print(f'Plans to delete: {len(names_to_delete)} images without metadata')
for name in names_to_delete:
remove(join('images', name))
if not exists('images.fbin'):
model = get_model('unum-cloud/uform-vl-multilingual')
vectors = []
for name in tqdm(names, desc='Vectorizing images'):
image = Image.open(join('images', name + '.jpg'))
image_data = model.preprocess_image(image)
image_embedding = model.encode_image(image_data).detach().numpy()
vectors.append(image_embedding)
image_mat = np.vstack(vectors)
save_matrix(image_mat, 'images.fbin')
if not exists('images.txt'):
datas = []
for name in tqdm(names, desc='Encoding images'):
data = image_to_data(join('images', name + '.jpg'))
datas.append(data)
with open('images.txt', 'w') as f:
f.write('\n'.join(datas))
if not exists('images.usearch'):
image_mat = load_matrix('images.fbin')
count = image_mat.shape[0]
ndim = image_mat.shape[1]
index = Index(ndim=ndim, metric=MetricKind.Cos)
for idx in tqdm(range(count), desc='Indexing vectors'):
index.add(idx, image_mat[idx, :].flatten())
index.save('images.usearch')