Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -244,11 +244,170 @@ def predict_representation_explorer(model_choice, text):
|
|
244 |
else:
|
245 |
return "Please select a model."
|
246 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
247 |
|
248 |
# --- Gradio Interface Setup with Tabs ---
|
249 |
with gr.Blocks(title="SPLADE Demos") as demo:
|
250 |
-
gr.Markdown("# 🌌 SPLADE Demos: Sparse Representation Explorer") # Updated title
|
251 |
-
gr.Markdown("Explore different SPLADE models and their sparse representation types.") # Updated description
|
252 |
|
253 |
with gr.Tabs():
|
254 |
with gr.TabItem("Sparse Representation Explorer"):
|
@@ -275,5 +434,35 @@ with gr.Blocks(title="SPLADE Demos") as demo:
|
|
275 |
allow_flagging="never",
|
276 |
# live=True # Setting live=True might be slow for complex models on every keystroke
|
277 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
278 |
|
279 |
-
demo.launch()
|
|
|
244 |
else:
|
245 |
return "Please select a model."
|
246 |
|
247 |
+
# --- NEW: Core Representation Functions (Return RAW TENSORS - for Dot Product Tab) ---
|
248 |
+
def get_splade_cocondenser_vector(text):
|
249 |
+
if tokenizer_splade is None or model_splade is None:
|
250 |
+
return None
|
251 |
+
|
252 |
+
inputs = tokenizer_splade(text, return_tensors="pt", padding=True, truncation=True)
|
253 |
+
inputs = {k: v.to(model_splade.device) for k, v in inputs.items()}
|
254 |
+
|
255 |
+
with torch.no_grad():
|
256 |
+
output = model_splade(**inputs)
|
257 |
+
|
258 |
+
if hasattr(output, 'logits'):
|
259 |
+
splade_vector = torch.max(
|
260 |
+
torch.log(1 + torch.relu(output.logits)) * inputs['attention_mask'].unsqueeze(-1),
|
261 |
+
dim=1
|
262 |
+
)[0].squeeze()
|
263 |
+
return splade_vector
|
264 |
+
return None
|
265 |
+
|
266 |
+
def get_splade_lexical_vector(text):
|
267 |
+
if tokenizer_splade_lexical is None or model_splade_lexical is None:
|
268 |
+
return None
|
269 |
+
|
270 |
+
inputs = tokenizer_splade_lexical(text, return_tensors="pt", padding=True, truncation=True)
|
271 |
+
inputs = {k: v.to(model_splade_lexical.device) for k, v in inputs.items()}
|
272 |
+
|
273 |
+
with torch.no_grad():
|
274 |
+
output = model_splade_lexical(**inputs)
|
275 |
+
|
276 |
+
if hasattr(output, 'logits'):
|
277 |
+
splade_vector = torch.max(
|
278 |
+
torch.log(1 + torch.relu(output.logits)) * inputs['attention_mask'].unsqueeze(-1),
|
279 |
+
dim=1
|
280 |
+
)[0].squeeze()
|
281 |
+
|
282 |
+
vocab_size = tokenizer_splade_lexical.vocab_size
|
283 |
+
bow_mask = create_lexical_bow_mask(
|
284 |
+
inputs['input_ids'], vocab_size, tokenizer_splade_lexical
|
285 |
+
).squeeze()
|
286 |
+
|
287 |
+
splade_vector = splade_vector * bow_mask
|
288 |
+
return splade_vector
|
289 |
+
return None
|
290 |
+
|
291 |
+
def get_splade_doc_vector(text):
|
292 |
+
if tokenizer_splade_doc is None or model_splade_doc is None:
|
293 |
+
return None
|
294 |
+
|
295 |
+
inputs = tokenizer_splade_doc(text, return_tensors="pt", padding=True, truncation=True)
|
296 |
+
inputs = {k: v.to(model_splade_doc.device) for k, v in inputs.items()}
|
297 |
+
|
298 |
+
with torch.no_grad():
|
299 |
+
output = model_splade_doc(**inputs)
|
300 |
+
|
301 |
+
if hasattr(output, "logits"):
|
302 |
+
vocab_size = tokenizer_splade_doc.vocab_size
|
303 |
+
binary_splade_vector = create_lexical_bow_mask(
|
304 |
+
inputs['input_ids'], vocab_size, tokenizer_splade_doc
|
305 |
+
).squeeze()
|
306 |
+
return binary_splade_vector
|
307 |
+
return None
|
308 |
+
|
309 |
+
|
310 |
+
# --- NEW: Function to get formatted representation from a raw vector and tokenizer ---
|
311 |
+
def format_sparse_vector_output(splade_vector, tokenizer, is_binary=False):
|
312 |
+
if splade_vector is None:
|
313 |
+
return "Failed to generate vector."
|
314 |
+
|
315 |
+
indices = torch.nonzero(splade_vector).squeeze().cpu().tolist()
|
316 |
+
if not isinstance(indices, list):
|
317 |
+
indices = [indices] if indices else []
|
318 |
+
|
319 |
+
if is_binary:
|
320 |
+
values = [1.0] * len(indices)
|
321 |
+
else:
|
322 |
+
values = splade_vector[indices].cpu().tolist()
|
323 |
+
|
324 |
+
token_weights = dict(zip(indices, values))
|
325 |
+
|
326 |
+
meaningful_tokens = {}
|
327 |
+
for token_id, weight in token_weights.items():
|
328 |
+
decoded_token = tokenizer.decode([token_id])
|
329 |
+
if decoded_token not in ["[CLS]", "[SEP]", "[PAD]", "[UNK]"] and len(decoded_token.strip()) > 0:
|
330 |
+
meaningful_tokens[decoded_token] = weight
|
331 |
+
|
332 |
+
if is_binary:
|
333 |
+
sorted_representation = sorted(meaningful_tokens.items(), key=lambda item: item[0]) # Sort alphabetically for binary
|
334 |
+
else:
|
335 |
+
sorted_representation = sorted(meaningful_tokens.items(), key=lambda item: item[1], reverse=True)
|
336 |
+
|
337 |
+
formatted_output = ""
|
338 |
+
if not sorted_representation:
|
339 |
+
formatted_output += "No significant terms found.\n"
|
340 |
+
else:
|
341 |
+
for i, (term, weight) in enumerate(sorted_representation):
|
342 |
+
if i >= 50 and is_binary: # Limit display for very long binary lists
|
343 |
+
formatted_output += f"...and {len(sorted_representation) - 50} more terms.\n"
|
344 |
+
break
|
345 |
+
if is_binary:
|
346 |
+
formatted_output += f"- **{term}**\n"
|
347 |
+
else:
|
348 |
+
formatted_output += f"- **{term}**: {weight:.4f}\n"
|
349 |
+
|
350 |
+
formatted_output += f"\nTotal non-zero terms: {len(indices)}\n"
|
351 |
+
formatted_output += f"Sparsity: {1 - (len(indices) / tokenizer.vocab_size):.2%}\n"
|
352 |
+
|
353 |
+
return formatted_output
|
354 |
+
|
355 |
+
|
356 |
+
# --- NEW: Dot Product Calculation Function for the new tab ---
|
357 |
+
def calculate_dot_product_and_representations(model_choice, query_text, doc_text):
|
358 |
+
query_vector = None
|
359 |
+
doc_vector = None
|
360 |
+
query_rep_str = ""
|
361 |
+
doc_rep_str = ""
|
362 |
+
|
363 |
+
selected_tokenizer = None
|
364 |
+
|
365 |
+
if model_choice == "SPLADE-cocondenser-distil (weighting and expansion)":
|
366 |
+
query_vector = get_splade_cocondenser_vector(query_text)
|
367 |
+
doc_vector = get_splade_cocondenser_vector(doc_text)
|
368 |
+
selected_tokenizer = tokenizer_splade
|
369 |
+
query_rep_str = "Query SPLADE-cocondenser-distil Representation (Weighting and Expansion):\n"
|
370 |
+
doc_rep_str = "Document SPLADE-cocondenser-distil Representation (Weighting and Expansion):\n"
|
371 |
+
is_binary = False
|
372 |
+
elif model_choice == "SPLADE-v3-Lexical (weighting)":
|
373 |
+
query_vector = get_splade_lexical_vector(query_text)
|
374 |
+
doc_vector = get_splade_lexical_vector(doc_text)
|
375 |
+
selected_tokenizer = tokenizer_splade_lexical
|
376 |
+
query_rep_str = "Query SPLADE-v3-Lexical Representation (Weighting):\n"
|
377 |
+
doc_rep_str = "Document SPLADE-v3-Lexical Representation (Weighting):\n"
|
378 |
+
is_binary = False
|
379 |
+
elif model_choice == "SPLADE-v3-Doc (binary)":
|
380 |
+
query_vector = get_splade_doc_vector(query_text)
|
381 |
+
doc_vector = get_splade_doc_vector(doc_text)
|
382 |
+
selected_tokenizer = tokenizer_splade_doc
|
383 |
+
query_rep_str = "Query SPLADE-v3-Doc Representation (Binary):\n"
|
384 |
+
doc_rep_str = "Document SPLADE-v3-Doc Representation (Binary):\n"
|
385 |
+
is_binary = True
|
386 |
+
else:
|
387 |
+
return "Please select a model.", "", ""
|
388 |
+
|
389 |
+
if query_vector is None or doc_vector is None:
|
390 |
+
return "Failed to generate one or both vectors. Please check model loading.", "", ""
|
391 |
+
|
392 |
+
# Calculate dot product
|
393 |
+
dot_product = float(torch.dot(query_vector.cpu(), doc_vector.cpu()).item())
|
394 |
+
|
395 |
+
# Format representations
|
396 |
+
query_rep_str += format_sparse_vector_output(query_vector, selected_tokenizer, is_binary)
|
397 |
+
doc_rep_str += format_sparse_vector_output(doc_vector, selected_tokenizer, is_binary)
|
398 |
+
|
399 |
+
# Combine output
|
400 |
+
full_output = f"### Dot Product Score: {dot_product:.6f}\n\n"
|
401 |
+
full_output += "---\n\n"
|
402 |
+
full_output += f"{query_rep_str}\n\n---\n\n{doc_rep_str}"
|
403 |
+
|
404 |
+
return full_output
|
405 |
+
|
406 |
|
407 |
# --- Gradio Interface Setup with Tabs ---
|
408 |
with gr.Blocks(title="SPLADE Demos") as demo:
|
409 |
+
gr.Markdown("# 🌌 SPLADE Demos: Sparse Representation Explorer and Retriever") # Updated title
|
410 |
+
gr.Markdown("Explore different SPLADE models and their sparse representation types, and calculate similarity between query and document representations.") # Updated description
|
411 |
|
412 |
with gr.Tabs():
|
413 |
with gr.TabItem("Sparse Representation Explorer"):
|
|
|
434 |
allow_flagging="never",
|
435 |
# live=True # Setting live=True might be slow for complex models on every keystroke
|
436 |
)
|
437 |
+
|
438 |
+
with gr.TabItem("Query-Document Dot Product Calculator"): # NEW TAB
|
439 |
+
gr.Markdown("### Calculate Dot Product Similarity between Query and Document")
|
440 |
+
gr.Markdown("Select a SPLADE model to encode both your query and document, then see their sparse representations and their similarity score.")
|
441 |
+
gr.Interface(
|
442 |
+
fn=calculate_dot_product_and_representations,
|
443 |
+
inputs=[
|
444 |
+
gr.Radio(
|
445 |
+
[
|
446 |
+
"SPLADE-cocondenser-distil (weighting and expansion)",
|
447 |
+
"SPLADE-v3-Lexical (weighting)",
|
448 |
+
"SPLADE-v3-Doc (binary)"
|
449 |
+
],
|
450 |
+
label="Choose Encoding Model",
|
451 |
+
value="SPLADE-cocondenser-distil (weighting and expansion)"
|
452 |
+
),
|
453 |
+
gr.Textbox(
|
454 |
+
lines=3,
|
455 |
+
label="Enter Query Text:",
|
456 |
+
placeholder="e.g., best pizza in Naples"
|
457 |
+
),
|
458 |
+
gr.Textbox(
|
459 |
+
lines=5,
|
460 |
+
label="Enter Document Text:",
|
461 |
+
placeholder="e.g., Naples is famous for its delicious pizza, known for its soft, chewy crust and fresh ingredients."
|
462 |
+
)
|
463 |
+
],
|
464 |
+
outputs=gr.Markdown(),
|
465 |
+
allow_flagging="never"
|
466 |
+
)
|
467 |
|
468 |
+
demo.launch()
|