Spaces:
Running
Running
moving share link under input field on the sparse reps tab
Browse files
app.py
CHANGED
@@ -9,56 +9,9 @@ import os
|
|
9 |
|
10 |
# Add this CSS at the top of your file, after the imports
|
11 |
css = """
|
12 |
-
/*
|
13 |
-
.
|
14 |
-
|
15 |
-
top: 20px !important;
|
16 |
-
right: 20px !important;
|
17 |
-
z-index: 1000 !important;
|
18 |
-
background: #4CAF50 !important;
|
19 |
-
color: white !important;
|
20 |
-
border-radius: 8px !important;
|
21 |
-
padding: 8px 16px !important;
|
22 |
-
font-weight: bold !important;
|
23 |
-
box-shadow: 0 2px 10px rgba(0,0,0,0.2) !important;
|
24 |
-
}
|
25 |
-
|
26 |
-
.share-button:hover {
|
27 |
-
background: #45a049 !important;
|
28 |
-
transform: translateY(-1px) !important;
|
29 |
-
}
|
30 |
-
|
31 |
-
/* Alternative positions - uncomment the one you want instead */
|
32 |
-
|
33 |
-
/* Top-left corner */
|
34 |
-
/*
|
35 |
-
.share-button {
|
36 |
-
position: fixed !important;
|
37 |
-
top: 20px !important;
|
38 |
-
left: 20px !important;
|
39 |
-
z-index: 1000 !important;
|
40 |
-
}
|
41 |
-
*/
|
42 |
-
|
43 |
-
/* Bottom-right corner (mobile-friendly) */
|
44 |
-
/*
|
45 |
-
.share-button {
|
46 |
-
position: fixed !important;
|
47 |
-
bottom: 20px !important;
|
48 |
-
right: 20px !important;
|
49 |
-
z-index: 1000 !important;
|
50 |
-
}
|
51 |
-
*/
|
52 |
-
|
53 |
-
/* Bottom-center */
|
54 |
-
/*
|
55 |
-
.share-button {
|
56 |
-
position: fixed !important;
|
57 |
-
bottom: 20px !important;
|
58 |
-
left: 50% !important;
|
59 |
-
transform: translateX(-50%) !important;
|
60 |
-
z-index: 1000 !important;
|
61 |
-
}
|
62 |
*/
|
63 |
"""
|
64 |
|
@@ -130,11 +83,11 @@ def create_lexical_bow_mask(input_ids_batch, vocab_size, tokenizer):
|
|
130 |
tokenizer.unk_token_id
|
131 |
]:
|
132 |
meaningful_token_ids.append(token_id)
|
133 |
-
|
134 |
if meaningful_token_ids:
|
135 |
# Apply mask to the current row in the batch
|
136 |
bow_masks[i, list(set(meaningful_token_ids))] = 1
|
137 |
-
|
138 |
return bow_masks
|
139 |
|
140 |
|
@@ -185,7 +138,7 @@ def get_splade_cocondenser_representation(text):
|
|
185 |
|
186 |
info_output = f"" # Line 1
|
187 |
info_output += f"Total non-zero terms in vector: {len(indices)}\n" # Line 2 (and onwards for sparsity)
|
188 |
-
|
189 |
|
190 |
return formatted_output, info_output
|
191 |
|
@@ -243,7 +196,7 @@ def get_splade_lexical_representation(text):
|
|
243 |
|
244 |
info_output = f"" # Line 1
|
245 |
info_output += f"Total non-zero terms in vector: {len(indices)}\n" # Line 2 (and onwards for sparsity)
|
246 |
-
|
247 |
|
248 |
return formatted_output, info_output
|
249 |
|
@@ -260,11 +213,11 @@ def get_splade_doc_representation(text):
|
|
260 |
binary_bow_vector = create_lexical_bow_mask(
|
261 |
inputs['input_ids'], vocab_size, tokenizer_splade_doc
|
262 |
).squeeze() # Squeeze back for single output
|
263 |
-
|
264 |
indices = torch.nonzero(binary_bow_vector).squeeze().cpu().tolist()
|
265 |
if not isinstance(indices, list):
|
266 |
indices = [indices] if indices else []
|
267 |
-
|
268 |
values = [1.0] * len(indices) # All values are 1 for binary representation
|
269 |
token_weights = dict(zip(indices, values))
|
270 |
|
@@ -338,12 +291,12 @@ def get_splade_lexical_vector(text):
|
|
338 |
torch.log(1 + torch.relu(output.logits)) * inputs['attention_mask'].unsqueeze(-1),
|
339 |
dim=1
|
340 |
)[0].squeeze()
|
341 |
-
|
342 |
vocab_size = tokenizer_splade_lexical.vocab_size
|
343 |
bow_mask = create_lexical_bow_mask(
|
344 |
inputs['input_ids'], vocab_size, tokenizer_splade_lexical
|
345 |
).squeeze()
|
346 |
-
|
347 |
splade_vector = splade_vector * bow_mask
|
348 |
return splade_vector
|
349 |
return None
|
@@ -377,7 +330,7 @@ def format_sparse_vector_output(splade_vector, tokenizer, is_binary=False):
|
|
377 |
values = [1.0] * len(indices)
|
378 |
else:
|
379 |
values = splade_vector[indices].cpu().tolist()
|
380 |
-
|
381 |
token_weights = dict(zip(indices, values))
|
382 |
|
383 |
meaningful_tokens = {}
|
@@ -408,8 +361,8 @@ def format_sparse_vector_output(splade_vector, tokenizer, is_binary=False):
|
|
408 |
|
409 |
# This is the line that will now always be split into two
|
410 |
info_output = f"Total non-zero terms: {len(indices)}\n" # Line 1
|
411 |
-
|
412 |
-
|
413 |
return formatted_output, info_output
|
414 |
|
415 |
|
@@ -451,7 +404,7 @@ def calculate_dot_product_and_representations_independent(query_model_choice, do
|
|
451 |
# Combine output into a single string for the Markdown component
|
452 |
full_output = f"### Dot Product Score: {dot_product:.6f}\n\n"
|
453 |
full_output += "---\n\n"
|
454 |
-
|
455 |
# Query Representation
|
456 |
full_output += f"Query Representation ({query_model_name_display}):\n\n"
|
457 |
full_output += query_main_rep_str + "\n\n" + query_info_str # Added an extra newline for better spacing
|
@@ -460,7 +413,7 @@ def calculate_dot_product_and_representations_independent(query_model_choice, do
|
|
460 |
# Document Representation
|
461 |
full_output += f"Document Representation ({doc_model_name_display}):\n\n"
|
462 |
full_output += doc_main_rep_str + "\n\n" + doc_info_str # Added an extra newline for better spacing
|
463 |
-
|
464 |
return full_output
|
465 |
|
466 |
|
@@ -488,7 +441,13 @@ with gr.Blocks(title="SPLADE Demos", css=css) as demo:
|
|
488 |
label="Enter your query or document text here:",
|
489 |
placeholder="e.g., Why is Padua the nicest city in Italy?"
|
490 |
)
|
491 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
492 |
info_output_display = gr.Markdown(
|
493 |
value="",
|
494 |
label="Vector Information",
|
@@ -496,7 +455,7 @@ with gr.Blocks(title="SPLADE Demos", css=css) as demo:
|
|
496 |
)
|
497 |
with gr.Column(scale=2): # Right column for the main representation output
|
498 |
main_representation_output = gr.Markdown()
|
499 |
-
|
500 |
# Connect the interface elements
|
501 |
model_radio.change(
|
502 |
fn=predict_representation_explorer,
|
@@ -508,15 +467,16 @@ with gr.Blocks(title="SPLADE Demos", css=css) as demo:
|
|
508 |
inputs=[model_radio, input_text],
|
509 |
outputs=[main_representation_output, info_output_display]
|
510 |
)
|
511 |
-
|
512 |
# Initial call to populate on load (optional, but good for demo)
|
513 |
demo.load(
|
514 |
fn=lambda: predict_representation_explorer(model_radio.value, input_text.value),
|
515 |
outputs=[main_representation_output, info_output_display]
|
516 |
)
|
517 |
|
518 |
-
with gr.TabItem("Compare Encoders"): #
|
519 |
-
|
|
|
520 |
# Define the common model choices for cleaner code
|
521 |
model_choices = [
|
522 |
"MLM encoder (SPLADE-cocondenser-distil)",
|
@@ -549,7 +509,7 @@ with gr.Blocks(title="SPLADE Demos", css=css) as demo:
|
|
549 |
)
|
550 |
],
|
551 |
outputs=gr.Markdown(),
|
552 |
-
allow_flagging="never"
|
553 |
)
|
554 |
|
555 |
demo.launch()
|
|
|
9 |
|
10 |
# Add this CSS at the top of your file, after the imports
|
11 |
css = """
|
12 |
+
/* The global fixed positioning for the share button is no longer needed
|
13 |
+
because we'll place gr.ShareButton directly in the UI.
|
14 |
+
You can remove or comment out any previous .share-button CSS if it was there.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
15 |
*/
|
16 |
"""
|
17 |
|
|
|
83 |
tokenizer.unk_token_id
|
84 |
]:
|
85 |
meaningful_token_ids.append(token_id)
|
86 |
+
|
87 |
if meaningful_token_ids:
|
88 |
# Apply mask to the current row in the batch
|
89 |
bow_masks[i, list(set(meaningful_token_ids))] = 1
|
90 |
+
|
91 |
return bow_masks
|
92 |
|
93 |
|
|
|
138 |
|
139 |
info_output = f"" # Line 1
|
140 |
info_output += f"Total non-zero terms in vector: {len(indices)}\n" # Line 2 (and onwards for sparsity)
|
141 |
+
|
142 |
|
143 |
return formatted_output, info_output
|
144 |
|
|
|
196 |
|
197 |
info_output = f"" # Line 1
|
198 |
info_output += f"Total non-zero terms in vector: {len(indices)}\n" # Line 2 (and onwards for sparsity)
|
199 |
+
|
200 |
|
201 |
return formatted_output, info_output
|
202 |
|
|
|
213 |
binary_bow_vector = create_lexical_bow_mask(
|
214 |
inputs['input_ids'], vocab_size, tokenizer_splade_doc
|
215 |
).squeeze() # Squeeze back for single output
|
216 |
+
|
217 |
indices = torch.nonzero(binary_bow_vector).squeeze().cpu().tolist()
|
218 |
if not isinstance(indices, list):
|
219 |
indices = [indices] if indices else []
|
220 |
+
|
221 |
values = [1.0] * len(indices) # All values are 1 for binary representation
|
222 |
token_weights = dict(zip(indices, values))
|
223 |
|
|
|
291 |
torch.log(1 + torch.relu(output.logits)) * inputs['attention_mask'].unsqueeze(-1),
|
292 |
dim=1
|
293 |
)[0].squeeze()
|
294 |
+
|
295 |
vocab_size = tokenizer_splade_lexical.vocab_size
|
296 |
bow_mask = create_lexical_bow_mask(
|
297 |
inputs['input_ids'], vocab_size, tokenizer_splade_lexical
|
298 |
).squeeze()
|
299 |
+
|
300 |
splade_vector = splade_vector * bow_mask
|
301 |
return splade_vector
|
302 |
return None
|
|
|
330 |
values = [1.0] * len(indices)
|
331 |
else:
|
332 |
values = splade_vector[indices].cpu().tolist()
|
333 |
+
|
334 |
token_weights = dict(zip(indices, values))
|
335 |
|
336 |
meaningful_tokens = {}
|
|
|
361 |
|
362 |
# This is the line that will now always be split into two
|
363 |
info_output = f"Total non-zero terms: {len(indices)}\n" # Line 1
|
364 |
+
|
365 |
+
|
366 |
return formatted_output, info_output
|
367 |
|
368 |
|
|
|
404 |
# Combine output into a single string for the Markdown component
|
405 |
full_output = f"### Dot Product Score: {dot_product:.6f}\n\n"
|
406 |
full_output += "---\n\n"
|
407 |
+
|
408 |
# Query Representation
|
409 |
full_output += f"Query Representation ({query_model_name_display}):\n\n"
|
410 |
full_output += query_main_rep_str + "\n\n" + query_info_str # Added an extra newline for better spacing
|
|
|
413 |
# Document Representation
|
414 |
full_output += f"Document Representation ({doc_model_name_display}):\n\n"
|
415 |
full_output += doc_main_rep_str + "\n\n" + doc_info_str # Added an extra newline for better spacing
|
416 |
+
|
417 |
return full_output
|
418 |
|
419 |
|
|
|
441 |
label="Enter your query or document text here:",
|
442 |
placeholder="e.g., Why is Padua the nicest city in Italy?"
|
443 |
)
|
444 |
+
# --- NEW: Place the gr.ShareButton here ---
|
445 |
+
gr.ShareButton(
|
446 |
+
value="Share My Sparse Representation",
|
447 |
+
components=[input_text, model_radio], # You can specify components to share
|
448 |
+
visible=True # Make sure it's visible
|
449 |
+
)
|
450 |
+
# --- End New ---
|
451 |
info_output_display = gr.Markdown(
|
452 |
value="",
|
453 |
label="Vector Information",
|
|
|
455 |
)
|
456 |
with gr.Column(scale=2): # Right column for the main representation output
|
457 |
main_representation_output = gr.Markdown()
|
458 |
+
|
459 |
# Connect the interface elements
|
460 |
model_radio.change(
|
461 |
fn=predict_representation_explorer,
|
|
|
467 |
inputs=[model_radio, input_text],
|
468 |
outputs=[main_representation_output, info_output_display]
|
469 |
)
|
470 |
+
|
471 |
# Initial call to populate on load (optional, but good for demo)
|
472 |
demo.load(
|
473 |
fn=lambda: predict_representation_explorer(model_radio.value, input_text.value),
|
474 |
outputs=[main_representation_output, info_output_display]
|
475 |
)
|
476 |
|
477 |
+
with gr.TabItem("Compare Encoders"): # Reverted to original gr.Interface setup
|
478 |
+
gr.Markdown("### Calculate Dot Product Similarity Between Encoded Query and Document")
|
479 |
+
|
480 |
# Define the common model choices for cleaner code
|
481 |
model_choices = [
|
482 |
"MLM encoder (SPLADE-cocondenser-distil)",
|
|
|
509 |
)
|
510 |
],
|
511 |
outputs=gr.Markdown(),
|
512 |
+
allow_flagging="never" # Keep this to keep the share button at the bottom of THIS interface
|
513 |
)
|
514 |
|
515 |
demo.launch()
|