File size: 7,978 Bytes
71d3180
 
 
1cc6474
 
 
 
 
 
 
 
 
 
 
71d3180
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6c19590
71d3180
6c19590
71d3180
 
 
 
6c19590
 
71d3180
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6c19590
 
 
71d3180
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6c19590
 
71d3180
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
"""
Example usage script for LLM2Vec4CXR model.
This demonstrates how to load and use the model for chest X-ray report analysis.

Prerequisites:
1. Install the LLM2Vec4CXR package:
   pip install git+https://github.com/lukeingawesome/llm2vec4cxr.git
   
   Or clone and install in development mode:
   git clone https://github.com/lukeingawesome/llm2vec4cxr.git
   cd llm2vec4cxr
   pip install -e .

2. The model will be automatically downloaded from Hugging Face when first used.
"""

import torch
import torch.nn.functional as F
from llm2vec_wrapper import LLM2VecWrapper as LLM2Vec

def load_llm2vec4cxr_model(model_name_or_path="lukeingawesome/llm2vec4cxr"):
    """
    Load the LLM2Vec4CXR model with proper configuration.
    
    Args:
        model_name_or_path (str): Hugging Face model path or local path
    
    Returns:
        tuple: (model, tokenizer)
    """
    # Load model with the specific configuration used for LLM2Vec4CXR
    model = LLM2Vec.from_pretrained(
        base_model_name_or_path=model_name_or_path,
        enable_bidirectional=True,
        pooling_mode="latent_attention",  # This is the key modification
        max_length=512,
        torch_dtype=torch.bfloat16,
    )
    
    # Configure tokenizer
    tokenizer = model.tokenizer
    tokenizer.padding_side = 'left'
    
    return model, tokenizer

def tokenize_with_separator(texts, tokenizer, max_length=512):
    """
    Tokenize texts with special handling for separator-based splitting.
    This is useful for instruction-following tasks.
    
    Args:
        texts (list): List of texts to tokenize
        tokenizer: The tokenizer to use
        max_length (int): Maximum sequence length
    
    Returns:
        dict: Tokenized inputs with attention masks and embed masks
    """
    texts_2 = []
    original_texts = []
    separator = '!@#$%^&*()'
    
    for text in texts:
        parts = text.split(separator)
        texts_2.append(parts[1] if len(parts) > 1 else "")
        original_texts.append("".join(parts))

    # Tokenize original texts
    tokenized = tokenizer(
        original_texts,
        return_tensors="pt",
        padding=True,
        truncation=True,
        max_length=max_length,
    )
    
    # Create embedding masks for the separated parts
    embed_mask = None
    for t_i, t in enumerate(texts_2):
        ids = tokenizer(
            [t],
            return_tensors="pt",
            padding=True,
            truncation=True,
            max_length=max_length,
            add_special_tokens=False,
        )
        
        e_m = torch.zeros_like(tokenized["attention_mask"][t_i])
        if len(ids["input_ids"][0]) > 0:
            e_m[-len(ids["input_ids"][0]):] = torch.ones(len(ids["input_ids"][0]))
            
        if embed_mask is None:
            embed_mask = e_m.unsqueeze(0)
        else:
            embed_mask = torch.cat((embed_mask, e_m.unsqueeze(0)), dim=0)

    tokenized["embed_mask"] = embed_mask
    return tokenized

def compute_similarities(model, tokenizer, texts, device):
    """
    Compute similarity scores between the first text and all other texts.
    
    Args:
        model: The LLM2Vec model
        tokenizer: The tokenizer
        texts (list): List of texts to compare (first text is the reference)
        device: The device to run computations on
    
    Returns:
        tuple: (embeddings, similarities)
    """
    with torch.no_grad():
        # Use separator-based tokenization if texts contain the separator
        if any('!@#$%^&*()' in text for text in texts):
            tokenized = tokenize_with_separator(texts, tokenizer, 512)
        else:
            tokenized = tokenizer(
                texts,
                return_tensors="pt",
                padding=True,
                truncation=True,
                max_length=512,
            )
        
        tokenized = tokenized.to(device)
        if hasattr(tokenized, 'to'):
            tokenized = tokenized.to(torch.bfloat16)
        else:
            # Convert each tensor in the dict
            for key in tokenized:
                if torch.is_tensor(tokenized[key]):
                    tokenized[key] = tokenized[key].to(torch.bfloat16)
        
        embeddings = model(tokenized)
        
        # Compute cosine similarities between first embedding and all others
        similarities = F.cosine_similarity(embeddings[0], embeddings[1:], dim=1)
    
    return embeddings, similarities

def main():
    """
    Example usage of the LLM2Vec4CXR model for chest X-ray report analysis.
    """
    # Set device
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    print(f"Using device: {device}")
    
    # Load the model
    print("Loading LLM2Vec4CXR model...")
    model, tokenizer = load_llm2vec4cxr_model()
    model = model.to(device).to(torch.bfloat16)
    model.eval()
    
    # Example 1: Basic text embedding using built-in method
    print("\n" + "="*60)
    print("Example 1: Basic Text Embedding (Built-in Method)")
    print("="*60)
    
    report = "There is a small increase in the left-sided effusion. There continues to be volume loss at both bases."
    
    # Use the convenient built-in method
    embedding = model.encode_text(report)
    
    print(f"Report: {report}")
    print(f"Embedding shape: {embedding.shape}")
    print(f"Embedding norm: {torch.norm(embedding).item():.4f}")
    
    # Example 2: Instruction-based similarity comparison
    print("\n" + "="*60)
    print("Example 2: Instruction-based Similarity Comparison")
    print("="*60)
    
    separator = '!@#$%^&*()'
    instruction = 'Determine the change or the status of the pleural effusion.'
    report = 'There is a small increase in the left-sided effusion. There continues to be volume loss at both bases.'
    text = instruction + separator + report
    
    comparison_options = [
        'No pleural effusion',
        'Pleural effusion',
        'Effusion is seen in the right',
        'Effusion is seen in the left',
        'Pleural effusion is improving',
        'Pleural effusion is stable',
        'Pleural effusion is worsening'
    ]
    
    all_texts = [text] + comparison_options
    
    # Use built-in method for instruction-based encoding
    embeddings = model.encode_with_instruction(all_texts)
    similarities = F.cosine_similarity(embeddings[0], embeddings[1:], dim=1)
    
    print(f"Original text: {report}")
    print(f"Instruction: {instruction}")
    print("\nSimilarity Scores:")
    print("-" * 50)
    
    for option, score in zip(comparison_options, similarities):
        print(f"{option:<35} | {score.item():.4f}")
    
    # Find the most similar option
    best_match_idx = torch.argmax(similarities).item()
    print(f"\nBest match: {comparison_options[best_match_idx]} (score: {similarities[best_match_idx].item():.4f})")
    
    # Example 3: Multiple report comparison
    print("\n" + "="*60)
    print("Example 3: Multiple Report Comparison")
    print("="*60)
    
    reports = [
        "No acute cardiopulmonary abnormality.",
        "Small bilateral pleural effusions.",
        "Large left pleural effusion with compressive atelectasis.",
        "Interval improvement in bilateral pleural effusions.",
        "Worsening bilateral pleural effusions."
    ]
    
    print("Computing embeddings for multiple reports...")
    # Use built-in method for multiple texts
    embeddings = model.encode_text(reports)
    
    # Compute pairwise similarities
    similarity_matrix = F.cosine_similarity(
        embeddings.unsqueeze(1), 
        embeddings.unsqueeze(0), 
        dim=2
    )
    
    print("\nPairwise Similarity Matrix:")
    print("-" * 30)
    for i, report1 in enumerate(reports):
        print(f"Report {i+1}: {report1[:30]}...")
        for j, report2 in enumerate(reports):
            print(f"  vs Report {j+1}: {similarity_matrix[i][j].item():.4f}")
        print()

if __name__ == "__main__":
    main()