Update app.py
Browse files
app.py
CHANGED
@@ -14,7 +14,7 @@ def get_relevance_score_and_excerpt(query, paragraph):
|
|
14 |
return "Please provide both a query and a document paragraph.", ""
|
15 |
|
16 |
# Tokenize the input
|
17 |
-
inputs = tokenizer(query, paragraph, return_tensors="pt", truncation=True, padding=True
|
18 |
|
19 |
with torch.no_grad():
|
20 |
output = model(**inputs, output_attentions=True) # Get attention scores
|
@@ -23,31 +23,37 @@ def get_relevance_score_and_excerpt(query, paragraph):
|
|
23 |
logit = output.logits.squeeze().item()
|
24 |
relevance_score = torch.sigmoid(torch.tensor(logit)).item()
|
25 |
|
26 |
-
# Extract attention scores (
|
27 |
attention = output.attentions[-1] # Shape: (batch_size, num_heads, seq_len, seq_len)
|
28 |
|
29 |
-
# Average across
|
30 |
-
attention_scores = attention.mean(dim=1).
|
31 |
-
|
32 |
-
#
|
33 |
query_tokens = tokenizer.tokenize(query)
|
34 |
paragraph_tokens = tokenizer.tokenize(paragraph)
|
35 |
|
36 |
-
#
|
37 |
-
|
38 |
-
para_start_idx = query_length
|
39 |
para_end_idx = len(inputs["input_ids"][0]) - 1 # Ignore final [SEP] token
|
40 |
|
41 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
42 |
|
43 |
# Get indices of top-k attended tokens while preserving order
|
44 |
-
top_k = min(5,
|
45 |
-
top_indices = para_attention_scores.
|
46 |
|
47 |
-
# Extract
|
48 |
-
highlighted_tokens = [paragraph_tokens[i] for i in top_indices]
|
49 |
|
50 |
-
#
|
51 |
excerpt = tokenizer.convert_tokens_to_string(highlighted_tokens)
|
52 |
|
53 |
return round(relevance_score, 4), excerpt
|
|
|
14 |
return "Please provide both a query and a document paragraph.", ""
|
15 |
|
16 |
# Tokenize the input
|
17 |
+
inputs = tokenizer(query, paragraph, return_tensors="pt", truncation=True, padding=True)
|
18 |
|
19 |
with torch.no_grad():
|
20 |
output = model(**inputs, output_attentions=True) # Get attention scores
|
|
|
23 |
logit = output.logits.squeeze().item()
|
24 |
relevance_score = torch.sigmoid(torch.tensor(logit)).item()
|
25 |
|
26 |
+
# Extract attention scores (last attention layer)
|
27 |
attention = output.attentions[-1] # Shape: (batch_size, num_heads, seq_len, seq_len)
|
28 |
|
29 |
+
# Average attention across heads and batch dimension
|
30 |
+
attention_scores = attention.mean(dim=1).mean(dim=0) # Shape: (seq_len, seq_len)
|
31 |
+
|
32 |
+
# Get tokenized query and paragraph separately
|
33 |
query_tokens = tokenizer.tokenize(query)
|
34 |
paragraph_tokens = tokenizer.tokenize(paragraph)
|
35 |
|
36 |
+
query_len = len(query_tokens) + 2 # +2 for special tokens [CLS] and [SEP]
|
37 |
+
para_start_idx = query_len
|
|
|
38 |
para_end_idx = len(inputs["input_ids"][0]) - 1 # Ignore final [SEP] token
|
39 |
|
40 |
+
# Handle potential indexing issues
|
41 |
+
if para_end_idx <= para_start_idx:
|
42 |
+
return round(relevance_score, 4), "No relevant tokens extracted."
|
43 |
+
|
44 |
+
para_attention_scores = attention_scores[para_start_idx:para_end_idx, para_start_idx:para_end_idx].mean(dim=0)
|
45 |
+
|
46 |
+
if para_attention_scores.numel() == 0:
|
47 |
+
return round(relevance_score, 4), "No relevant tokens extracted."
|
48 |
|
49 |
# Get indices of top-k attended tokens while preserving order
|
50 |
+
top_k = min(5, para_attention_scores.size(0)) # Ensure top-k does not exceed available tokens
|
51 |
+
top_indices = para_attention_scores.topk(top_k).indices.sort().values # Sort indices to preserve order
|
52 |
|
53 |
+
# Extract highlighted tokens from the paragraph
|
54 |
+
highlighted_tokens = [paragraph_tokens[i] for i in top_indices.tolist()]
|
55 |
|
56 |
+
# Convert tokens back to a readable string
|
57 |
excerpt = tokenizer.convert_tokens_to_string(highlighted_tokens)
|
58 |
|
59 |
return round(relevance_score, 4), excerpt
|