OpenClip-B32-KNN-Captioner / make_captions_367m.py
ChristophSchuhmann's picture
Upload make_captions_367m.py with huggingface_hub
6ba8811
raw
history blame
No virus
2.15 kB
import faiss
import numpy as np
from autofaiss import build_index
import open_clip
import torch
import pandas as pd
import pickle
import numpy as np
from PIL import Image
import glob
import os
df = pd.read_parquet("laioncocoknn367.parquet")
print(df)
texts = df['caption'].tolist()
model, _, transform = open_clip.create_model_and_transforms("hf-hub:laion/CLIP-ViT-B-32-256x256-DataComp-s34B-b86K")
def normalized(a, axis=-1, order=2):
import numpy as np # pylint: disable=import-outside-toplevel
l2 = np.atleast_1d(np.linalg.norm(a, order, axis))
l2[l2 == 0] = 1
return a / np.expand_dims(l2, axis)
index = faiss.read_index("laioncocoknn367.index", faiss.IO_FLAG_MMAP | faiss.IO_FLAG_READ_ONLY)
files =glob.glob("/images/*.jpg")
mediatype = ["movie still HQ depth-of-field", "painting","drawing","realistic photo, photograph","CGI - computer graphics - 3D", "powerpoint slide - text - ebook", "pixelart, pixelated retro video game, low resolution", "ASCII", "cartoon", "stockphoto", "beautiful anime still, no text", "meme", "selfie", "beautiful artwork" , "wallpaper, HD, 4k"]
tokenizer = open_clip.get_tokenizer('ViT-B-32')
with torch.no_grad(), torch.cuda.amp.autocast():
text_features = normalized(model.encode_text(tokenizer(mediatype)).cpu().detach().numpy())
for im in files:
image = Image.open(im)
tensor_image = transform(image).unsqueeze(0) # Adds a batch dimension
with torch.no_grad():
image_features = normalized(model.encode_image(tensor_image).cpu().detach().numpy())
mediatypepredictions = np.matmul(image_features, text_features.T) # Transpose text_features
# or equivalently
mediatypepredictions = image_features @ text_features.T # Transpose text_features
max_index = np.argmax(mediatypepredictions)
query_vector = image_features
k =1
distances, indices = index.search(query_vector, k)
results = list(zip(distances[0], indices[0]))
text=""
for r in results:
text +=texts[r[1]].replace("_"," ")+" , " #+" - "+ str(r[0])+" , "
text += mediatype[max_index]
print(im, text)
try:
image.save(f"./output/{text}.jpg")
except:
pass