-
Notifications
You must be signed in to change notification settings - Fork 0
/
image_evaluator.py
43 lines (34 loc) · 1.44 KB
/
image_evaluator.py
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
33
34
35
36
37
38
39
40
41
42
43
import numpy as np
from PIL import Image
from laion_aesthetics import MLP, normalizer, init_laion
import clip
import torch
import sys
## device for the clip and aesthetic model
device = "mps" # mps -> mac m chips, can also be "cuda" or "cpu" depending on torch installation
aesthetic_model, vit_model, preprocess = init_laion(device)
### Process the prompt, only needed once!
prompt = sys.argv[2] #e.g. "sunset, bright colors"
text_inputs = clip.tokenize(prompt).to(device)
with torch.no_grad():
text_features = vit_model.encode_text(text_inputs)
###
# Load the image from the file path
image_path = sys.argv[1]#"63b51fd5-5c75-4d02-9846-01a4736889e7.jpeg"
pil_image = Image.open(image_path)
# Display the image
pil_image.show()
## process image features
image = preprocess(pil_image).unsqueeze(0).to(device)
with torch.no_grad():
image_features = vit_model.encode_image(image)
im_emb_arr = normalizer(image_features.cpu().detach().numpy())
# aesthetic model prediction
prediction = aesthetic_model(torch.from_numpy(im_emb_arr).to(device).type(torch.float))
# cosine similarity between image features and text features
similarity = torch.cosine_similarity(text_features, image_features, dim=-1).mean()
#print("similarity", similarity)
aesthetic_eval_laion = prediction.item()
print("aesthetic_eval_laion", aesthetic_eval_laion)
similarity = torch.cosine_similarity(text_features, image_features, dim=-1).mean()
print("similarity", similarity.item())