gistfile1.txt

import os
import json
import numpy as np
import coremltools as ct
from transformers import CLIPTokenizer

# 1. Load labels
with open("stats.txt", "r") as f:
    categories = [line.strip() for line in f if line.strip()]

# 2. Load CLIP-compatible tokenizer
tokenizer = CLIPTokenizer.from_pretrained("openai/clip-vit-base-patch32")

# 3. Process embeddings 
model_path = os.path.join(
    "mobileclip_blt_text.mlpackage",
    "Data",
    "com.apple.CoreML",
    "model.mlmodel"
)

# 4. Validate model path
if not os.path.isfile(model_path):
    raise FileNotFoundError(f"❌ model.mlmodel not found at: {model_path}")

print(f"📦 Loading model: {model_path}")
model = ct.models.MLModel(model_path)

entries = []

# Encode text and extract embeddings
for text in categories:
    try:
        # Tokenize text using CLIP tokenizer
        encoded = tokenizer(
            text,
            return_tensors="np",
            padding="max_length",
            truncation=True,
            max_length=77
        )

        # Rank-2 shape (1, 77) and int32 type
        input_ids = encoded["input_ids"].astype("int32")

        # CoreML model expects the tokens under "text"
        input_data = {
            "text": input_ids
        }

        # Run prediction using CoreML model
        output = model.predict(input_data)
        print(f"✅ Output keys for '{text}': {output.keys()}")

        # Extract the embedding output from the model
        embedding_key = list(output.keys())[0]
        embedding = output[embedding_key]
        embedding = embedding.flatten().tolist() if hasattr(embedding, "flatten") else list(embedding)
        norm = np.linalg.norm(embedding)
        normalized = (np.array(embedding) / norm).tolist()
    
        # Save normalized embedding with label
        entries.append({
            "label": text,
            "embedding": normalized
        })

    except Exception as e:
        print(f"⚠️ Failed to process '{text}': {e}")
        continue

output_filename = "text_embeddings.json"
with open(output_filename, "w") as f:
    json.dump(entries, f, indent=2)

print(f"✅ Saved embeddings to {output_filename}")