Upload embeddings.py
Browse files- embeddings.py +53 -0
embeddings.py
ADDED
@@ -0,0 +1,53 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import datasets
|
2 |
+
from transformers import CLIPProcessor, CLIPModel
|
3 |
+
import torch
|
4 |
+
from PIL import Image
|
5 |
+
|
6 |
+
# Load the dataset
|
7 |
+
dataset = datasets.load_dataset("metmuseum/openaccess")
|
8 |
+
|
9 |
+
# Initialize the CLIP model and processor
|
10 |
+
model = CLIPModel.from_pretrained("openai/clip-vit-base-patch32")
|
11 |
+
processor = CLIPProcessor.from_pretrained("openai/clip-vit-base-patch32")
|
12 |
+
|
13 |
+
# Set the model to use the CPU or Apple's MPS if available
|
14 |
+
# Change this if you have a fancier computer (:
|
15 |
+
device = torch.device("mps" if torch.backends.mps.is_available() else "cpu")
|
16 |
+
model.to(device)
|
17 |
+
|
18 |
+
# Map to store embeddings with Object ID as key
|
19 |
+
embeddings_map = {}
|
20 |
+
|
21 |
+
# Function to create embeddings for a PIL Jpeg image
|
22 |
+
def create_embedding(image_pil):
|
23 |
+
try:
|
24 |
+
inputs = processor(images=image_pil, return_tensors="pt", padding=True).to(device)
|
25 |
+
with torch.no_grad():
|
26 |
+
embeddings = model.get_image_features(**inputs)
|
27 |
+
return embeddings
|
28 |
+
except Exception as e:
|
29 |
+
print(f"Error processing image: {e}")
|
30 |
+
return None
|
31 |
+
|
32 |
+
# Loop through the dataset and process the images, and add them to a map
|
33 |
+
# Optionally, you could add more keys here
|
34 |
+
# Or, just add the embeddings to the full dataset
|
35 |
+
for item in dataset['train']:
|
36 |
+
object_id = item['Object ID']
|
37 |
+
image_pil = item['jpg']
|
38 |
+
if image_pil:
|
39 |
+
embedding = create_embedding(image_pil)
|
40 |
+
if embedding is not None:
|
41 |
+
embeddings_map[object_id] = embedding.cpu().numpy()
|
42 |
+
|
43 |
+
# Convert embeddings map to a new dataset
|
44 |
+
# Note: I changed to to [embeddings][0] because the examples seemed to like [embeddings] format better
|
45 |
+
# than [[[embeddings]]]
|
46 |
+
# ...perhaps that's incorrect?
|
47 |
+
embedding_dataset = datasets.Dataset.from_dict({
|
48 |
+
'Object ID': list(embeddings_map.keys()),
|
49 |
+
'Embedding': [embedding.tolist() for embedding in embeddings_map.values()][0]
|
50 |
+
})
|
51 |
+
|
52 |
+
# Save the new dataset to disk
|
53 |
+
embedding_dataset.save_to_disk('metmuseum_embeddings')
|