Spaces:
Running
on
T4
Running
on
T4
Update app.py
Browse files
app.py
CHANGED
|
@@ -30,14 +30,12 @@ STYLE = """
|
|
| 30 |
.prose table {
|
| 31 |
margin-bottom: 0!important;
|
| 32 |
}
|
| 33 |
-
|
| 34 |
.prose td, th {
|
| 35 |
padding-left: 2px;
|
| 36 |
padding-right: 2px;
|
| 37 |
padding-top: 0;
|
| 38 |
padding-bottom: 0;
|
| 39 |
}
|
| 40 |
-
|
| 41 |
.tree {
|
| 42 |
padding: 0px;
|
| 43 |
margin: 0!important;
|
|
@@ -48,13 +46,11 @@ STYLE = """
|
|
| 48 |
text-align: center;
|
| 49 |
display:inline-block;
|
| 50 |
}
|
| 51 |
-
|
| 52 |
#root {
|
| 53 |
display: inline-grid!important;
|
| 54 |
width:auto!important;
|
| 55 |
min-width: 220px;
|
| 56 |
}
|
| 57 |
-
|
| 58 |
.tree ul {
|
| 59 |
padding-left: 20px;
|
| 60 |
position: relative;
|
|
@@ -75,7 +71,6 @@ STYLE = """
|
|
| 75 |
justify-content: start;
|
| 76 |
align-items: center;
|
| 77 |
}
|
| 78 |
-
|
| 79 |
.tree li::before, .tree li::after {
|
| 80 |
content: "";
|
| 81 |
position: absolute;
|
|
@@ -96,7 +91,6 @@ STYLE = """
|
|
| 96 |
.tree li:only-child::after, li:only-child::before {
|
| 97 |
display: none;
|
| 98 |
}
|
| 99 |
-
|
| 100 |
.tree li:first-child::before, .tree li:last-child::after {
|
| 101 |
border: 0 none;
|
| 102 |
}
|
|
@@ -111,7 +105,6 @@ STYLE = """
|
|
| 111 |
-webkit-border-radius: 5px 0 0 0;
|
| 112 |
-moz-border-radius: 5px 0 0 0;
|
| 113 |
}
|
| 114 |
-
|
| 115 |
.tree ul ul::before {
|
| 116 |
content: "";
|
| 117 |
position: absolute;
|
|
@@ -124,7 +117,6 @@ STYLE = """
|
|
| 124 |
.tree ul:has(> li:only-child)::before {
|
| 125 |
width:40px;
|
| 126 |
}
|
| 127 |
-
|
| 128 |
a:before {
|
| 129 |
border-right: 1px solid var(--body-text-color);
|
| 130 |
border-bottom: 1px solid var(--body-text-color);
|
|
@@ -138,8 +130,6 @@ a:before {
|
|
| 138 |
margin-left: 6px;
|
| 139 |
transform: rotate(315deg);
|
| 140 |
}
|
| 141 |
-
|
| 142 |
-
|
| 143 |
.tree li a {
|
| 144 |
border: 1px solid var(--body-text-color);
|
| 145 |
padding: 5px;
|
|
@@ -155,7 +145,6 @@ a:before {
|
|
| 155 |
.tree li a span {
|
| 156 |
padding: 5px;
|
| 157 |
font-size: 12px;
|
| 158 |
-
text-transform: uppercase;
|
| 159 |
letter-spacing: 1px;
|
| 160 |
font-weight: 500;
|
| 161 |
}
|
|
@@ -166,7 +155,7 @@ a:before {
|
|
| 166 |
.tree li a:hover+ul li::after, .tree li a:hover+ul li::before, .tree li a:hover+ul::before, .tree li a:hover+ul ul::before {
|
| 167 |
border-color: #7c2d12;
|
| 168 |
}
|
| 169 |
-
.chosen {
|
| 170 |
background-color: #ea580c;
|
| 171 |
width:auto!important;
|
| 172 |
}
|
|
@@ -206,33 +195,37 @@ def generate_markdown_table(
|
|
| 206 |
def generate_nodes(token_ix, node, step):
|
| 207 |
"""Recursively generate HTML for the tree nodes."""
|
| 208 |
token = tokenizer.decode([token_ix])
|
| 209 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 210 |
if node.table is not None:
|
| 211 |
html_content += node.table
|
| 212 |
html_content += "</a>"
|
|
|
|
| 213 |
if len(node.children.keys()) > 0:
|
| 214 |
html_content += "<ul> "
|
| 215 |
for token_ix, subnode in node.children.items():
|
| 216 |
html_content += generate_nodes(token_ix, subnode, step=step + 1)
|
| 217 |
html_content += "</ul>"
|
| 218 |
html_content += "</li>"
|
|
|
|
| 219 |
return html_content
|
| 220 |
|
| 221 |
|
| 222 |
def generate_html(start_sentence, original_tree):
|
| 223 |
-
|
| 224 |
html_output = f"""<div class="custom-container">
|
| 225 |
<div class="tree">
|
| 226 |
-
<ul>
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
html_output +=
|
| 230 |
-
|
| 231 |
-
html_output += generate_nodes(token_ix, subnode, step=1)
|
| 232 |
-
html_output += "</ul>"
|
| 233 |
-
|
| 234 |
html_output += """
|
| 235 |
-
</ul>
|
| 236 |
</div>
|
| 237 |
</body>
|
| 238 |
"""
|
|
@@ -246,11 +239,14 @@ from dataclasses import dataclass
|
|
| 246 |
|
| 247 |
@dataclass
|
| 248 |
class BeamNode:
|
|
|
|
| 249 |
cumulative_score: float
|
| 250 |
children_score_divider: float
|
| 251 |
table: str
|
| 252 |
current_sentence: str
|
| 253 |
children: Dict[int, "BeamNode"]
|
|
|
|
|
|
|
| 254 |
|
| 255 |
|
| 256 |
def generate_beams(start_sentence, scores, sequences, length_penalty):
|
|
@@ -258,13 +254,19 @@ def generate_beams(start_sentence, scores, sequences, length_penalty):
|
|
| 258 |
input_length = len(tokenizer([start_sentence], return_tensors="pt"))
|
| 259 |
original_tree = BeamNode(
|
| 260 |
cumulative_score=0,
|
|
|
|
| 261 |
table=None,
|
| 262 |
current_sentence=start_sentence,
|
| 263 |
children={},
|
| 264 |
children_score_divider=((input_length + 1) ** length_penalty),
|
|
|
|
|
|
|
| 265 |
)
|
| 266 |
n_beams = len(scores[0])
|
| 267 |
beam_trees = [original_tree] * n_beams
|
|
|
|
|
|
|
|
|
|
| 268 |
for step, step_scores in enumerate(scores):
|
| 269 |
(
|
| 270 |
top_token_indexes,
|
|
@@ -273,8 +275,13 @@ def generate_beams(start_sentence, scores, sequences, length_penalty):
|
|
| 273 |
current_completions,
|
| 274 |
top_tokens,
|
| 275 |
) = ([], [], [], [], [])
|
| 276 |
-
for beam_ix in range(n_beams):
|
| 277 |
current_beam = beam_trees[beam_ix]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 278 |
# Get top cumulative scores for the current beam
|
| 279 |
current_top_token_indexes = list(
|
| 280 |
np.array(scores[step][beam_ix].argsort()[-n_beams:])[::-1]
|
|
@@ -337,14 +344,31 @@ def generate_beams(start_sentence, scores, sequences, length_penalty):
|
|
| 337 |
+ scores[step][source_beam_ix][current_token_choice_ix].numpy()
|
| 338 |
)
|
| 339 |
beam_trees[source_beam_ix].children[current_token_choice_ix] = BeamNode(
|
|
|
|
| 340 |
table=None,
|
| 341 |
children={},
|
| 342 |
current_sentence=beam_trees[source_beam_ix].current_sentence
|
| 343 |
+ current_token_choice,
|
| 344 |
cumulative_score=cumulative_score,
|
|
|
|
|
|
|
| 345 |
children_score_divider=((input_length + step + 1) ** length_penalty),
|
|
|
|
|
|
|
|
|
|
|
|
|
| 346 |
)
|
| 347 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 348 |
# Reassign all beams at once
|
| 349 |
beam_trees = [
|
| 350 |
beam_trees[int(top_df_selected.iloc[beam_ix]["beam_index"])]
|
|
@@ -355,6 +379,7 @@ def generate_beams(start_sentence, scores, sequences, length_penalty):
|
|
| 355 |
for beam_ix in range(n_beams):
|
| 356 |
current_token_choice_ix = top_df_selected.iloc[beam_ix]["token_index"]
|
| 357 |
beam_trees[beam_ix] = beam_trees[beam_ix].children[current_token_choice_ix]
|
|
|
|
| 358 |
|
| 359 |
return original_tree
|
| 360 |
|
|
@@ -373,9 +398,10 @@ def get_beam_search_html(input_text, number_steps, number_beams, length_penalty)
|
|
| 373 |
do_sample=False,
|
| 374 |
)
|
| 375 |
markdown = "Output sequences:"
|
|
|
|
| 376 |
decoded_sequences = tokenizer.batch_decode(outputs.sequences)
|
| 377 |
for i, sequence in enumerate(decoded_sequences):
|
| 378 |
-
markdown += f"\n- {clean(sequence.replace('<s> ', ''))} (score {outputs.sequences_scores[i]:.2f})"
|
| 379 |
|
| 380 |
original_tree = generate_beams(
|
| 381 |
input_text,
|
|
@@ -393,7 +419,8 @@ with gr.Blocks(
|
|
| 393 |
),
|
| 394 |
css=STYLE,
|
| 395 |
) as demo:
|
| 396 |
-
gr.Markdown(
|
|
|
|
| 397 |
|
| 398 |
Play with the parameters below to understand how beam search decoding works!
|
| 399 |
|
|
@@ -402,15 +429,29 @@ Play with the parameters below to understand how beam search decoding works!
|
|
| 402 |
- **Number of steps**: the number of tokens to generate
|
| 403 |
- **Number of beams**: the number of beams to use
|
| 404 |
- **Length penalty**: the length penalty to apply to outputs. `length_penalty` > 0.0 promotes longer sequences, while `length_penalty` < 0.0 encourages shorter sequences.
|
| 405 |
-
"""
|
| 406 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 407 |
with gr.Row():
|
| 408 |
-
steps = gr.Slider(
|
| 409 |
-
|
| 410 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 411 |
button = gr.Button()
|
| 412 |
out_html = gr.Markdown()
|
| 413 |
out_markdown = gr.Markdown()
|
| 414 |
-
button.click(
|
|
|
|
|
|
|
|
|
|
|
|
|
| 415 |
|
| 416 |
demo.launch()
|
|
|
|
| 30 |
.prose table {
|
| 31 |
margin-bottom: 0!important;
|
| 32 |
}
|
|
|
|
| 33 |
.prose td, th {
|
| 34 |
padding-left: 2px;
|
| 35 |
padding-right: 2px;
|
| 36 |
padding-top: 0;
|
| 37 |
padding-bottom: 0;
|
| 38 |
}
|
|
|
|
| 39 |
.tree {
|
| 40 |
padding: 0px;
|
| 41 |
margin: 0!important;
|
|
|
|
| 46 |
text-align: center;
|
| 47 |
display:inline-block;
|
| 48 |
}
|
|
|
|
| 49 |
#root {
|
| 50 |
display: inline-grid!important;
|
| 51 |
width:auto!important;
|
| 52 |
min-width: 220px;
|
| 53 |
}
|
|
|
|
| 54 |
.tree ul {
|
| 55 |
padding-left: 20px;
|
| 56 |
position: relative;
|
|
|
|
| 71 |
justify-content: start;
|
| 72 |
align-items: center;
|
| 73 |
}
|
|
|
|
| 74 |
.tree li::before, .tree li::after {
|
| 75 |
content: "";
|
| 76 |
position: absolute;
|
|
|
|
| 91 |
.tree li:only-child::after, li:only-child::before {
|
| 92 |
display: none;
|
| 93 |
}
|
|
|
|
| 94 |
.tree li:first-child::before, .tree li:last-child::after {
|
| 95 |
border: 0 none;
|
| 96 |
}
|
|
|
|
| 105 |
-webkit-border-radius: 5px 0 0 0;
|
| 106 |
-moz-border-radius: 5px 0 0 0;
|
| 107 |
}
|
|
|
|
| 108 |
.tree ul ul::before {
|
| 109 |
content: "";
|
| 110 |
position: absolute;
|
|
|
|
| 117 |
.tree ul:has(> li:only-child)::before {
|
| 118 |
width:40px;
|
| 119 |
}
|
|
|
|
| 120 |
a:before {
|
| 121 |
border-right: 1px solid var(--body-text-color);
|
| 122 |
border-bottom: 1px solid var(--body-text-color);
|
|
|
|
| 130 |
margin-left: 6px;
|
| 131 |
transform: rotate(315deg);
|
| 132 |
}
|
|
|
|
|
|
|
| 133 |
.tree li a {
|
| 134 |
border: 1px solid var(--body-text-color);
|
| 135 |
padding: 5px;
|
|
|
|
| 145 |
.tree li a span {
|
| 146 |
padding: 5px;
|
| 147 |
font-size: 12px;
|
|
|
|
| 148 |
letter-spacing: 1px;
|
| 149 |
font-weight: 500;
|
| 150 |
}
|
|
|
|
| 155 |
.tree li a:hover+ul li::after, .tree li a:hover+ul li::before, .tree li a:hover+ul::before, .tree li a:hover+ul ul::before {
|
| 156 |
border-color: #7c2d12;
|
| 157 |
}
|
| 158 |
+
.end-of-text, .chosen {
|
| 159 |
background-color: #ea580c;
|
| 160 |
width:auto!important;
|
| 161 |
}
|
|
|
|
| 195 |
def generate_nodes(token_ix, node, step):
|
| 196 |
"""Recursively generate HTML for the tree nodes."""
|
| 197 |
token = tokenizer.decode([token_ix])
|
| 198 |
+
|
| 199 |
+
if node.is_final:
|
| 200 |
+
return f"<li> <a href='#' class='end-of-text'> <span> <b>{token_ix}:<br>{clean(token)}</b> <br> Total score: {node.total_score:.2f} </span> </a> </li>"
|
| 201 |
+
|
| 202 |
+
html_content = (
|
| 203 |
+
f"<li> <a href='#'> <span> <b>{token_ix}:<br>{clean(token)}</b> </span>"
|
| 204 |
+
)
|
| 205 |
if node.table is not None:
|
| 206 |
html_content += node.table
|
| 207 |
html_content += "</a>"
|
| 208 |
+
|
| 209 |
if len(node.children.keys()) > 0:
|
| 210 |
html_content += "<ul> "
|
| 211 |
for token_ix, subnode in node.children.items():
|
| 212 |
html_content += generate_nodes(token_ix, subnode, step=step + 1)
|
| 213 |
html_content += "</ul>"
|
| 214 |
html_content += "</li>"
|
| 215 |
+
|
| 216 |
return html_content
|
| 217 |
|
| 218 |
|
| 219 |
def generate_html(start_sentence, original_tree):
|
|
|
|
| 220 |
html_output = f"""<div class="custom-container">
|
| 221 |
<div class="tree">
|
| 222 |
+
<ul> <li> <a href='#' id='root'> <span> <b>{start_sentence}</b> </span> {original_tree.table} </a>"""
|
| 223 |
+
html_output += "<ul> "
|
| 224 |
+
for token_ix, subnode in original_tree.children.items():
|
| 225 |
+
html_output += generate_nodes(token_ix, subnode, step=1)
|
| 226 |
+
html_output += "</ul>"
|
|
|
|
|
|
|
|
|
|
| 227 |
html_output += """
|
| 228 |
+
</li> </ul>
|
| 229 |
</div>
|
| 230 |
</body>
|
| 231 |
"""
|
|
|
|
| 239 |
|
| 240 |
@dataclass
|
| 241 |
class BeamNode:
|
| 242 |
+
current_token_ix: int
|
| 243 |
cumulative_score: float
|
| 244 |
children_score_divider: float
|
| 245 |
table: str
|
| 246 |
current_sentence: str
|
| 247 |
children: Dict[int, "BeamNode"]
|
| 248 |
+
total_score: float
|
| 249 |
+
is_final: bool
|
| 250 |
|
| 251 |
|
| 252 |
def generate_beams(start_sentence, scores, sequences, length_penalty):
|
|
|
|
| 254 |
input_length = len(tokenizer([start_sentence], return_tensors="pt"))
|
| 255 |
original_tree = BeamNode(
|
| 256 |
cumulative_score=0,
|
| 257 |
+
current_token_ix=None,
|
| 258 |
table=None,
|
| 259 |
current_sentence=start_sentence,
|
| 260 |
children={},
|
| 261 |
children_score_divider=((input_length + 1) ** length_penalty),
|
| 262 |
+
total_score=None,
|
| 263 |
+
is_final=False,
|
| 264 |
)
|
| 265 |
n_beams = len(scores[0])
|
| 266 |
beam_trees = [original_tree] * n_beams
|
| 267 |
+
|
| 268 |
+
candidate_nodes = []
|
| 269 |
+
|
| 270 |
for step, step_scores in enumerate(scores):
|
| 271 |
(
|
| 272 |
top_token_indexes,
|
|
|
|
| 275 |
current_completions,
|
| 276 |
top_tokens,
|
| 277 |
) = ([], [], [], [], [])
|
| 278 |
+
for beam_ix in range(n_beams): # Get possible descendants for each beam
|
| 279 |
current_beam = beam_trees[beam_ix]
|
| 280 |
+
|
| 281 |
+
# skip if the beam is already final
|
| 282 |
+
if current_beam.is_final:
|
| 283 |
+
continue
|
| 284 |
+
|
| 285 |
# Get top cumulative scores for the current beam
|
| 286 |
current_top_token_indexes = list(
|
| 287 |
np.array(scores[step][beam_ix].argsort()[-n_beams:])[::-1]
|
|
|
|
| 344 |
+ scores[step][source_beam_ix][current_token_choice_ix].numpy()
|
| 345 |
)
|
| 346 |
beam_trees[source_beam_ix].children[current_token_choice_ix] = BeamNode(
|
| 347 |
+
current_token_ix=current_token_choice_ix,
|
| 348 |
table=None,
|
| 349 |
children={},
|
| 350 |
current_sentence=beam_trees[source_beam_ix].current_sentence
|
| 351 |
+ current_token_choice,
|
| 352 |
cumulative_score=cumulative_score,
|
| 353 |
+
total_score=cumulative_score
|
| 354 |
+
/ ((input_length + step - 1) ** length_penalty),
|
| 355 |
children_score_divider=((input_length + step + 1) ** length_penalty),
|
| 356 |
+
is_final=(
|
| 357 |
+
step == len(scores) - 1
|
| 358 |
+
or current_token_choice_ix == tokenizer.eos_token_id
|
| 359 |
+
),
|
| 360 |
)
|
| 361 |
|
| 362 |
+
# Check this child should be selected as a top beam.
|
| 363 |
+
# Is it a final step or an EOS token?
|
| 364 |
+
if (
|
| 365 |
+
step == len(scores) - 1
|
| 366 |
+
or current_token_choice_ix == tokenizer.eos_token_id
|
| 367 |
+
):
|
| 368 |
+
candidate_nodes.append(
|
| 369 |
+
beam_trees[source_beam_ix].children[current_token_choice_ix]
|
| 370 |
+
)
|
| 371 |
+
|
| 372 |
# Reassign all beams at once
|
| 373 |
beam_trees = [
|
| 374 |
beam_trees[int(top_df_selected.iloc[beam_ix]["beam_index"])]
|
|
|
|
| 379 |
for beam_ix in range(n_beams):
|
| 380 |
current_token_choice_ix = top_df_selected.iloc[beam_ix]["token_index"]
|
| 381 |
beam_trees[beam_ix] = beam_trees[beam_ix].children[current_token_choice_ix]
|
| 382 |
+
print("Final nodes", candidate_nodes)
|
| 383 |
|
| 384 |
return original_tree
|
| 385 |
|
|
|
|
| 398 |
do_sample=False,
|
| 399 |
)
|
| 400 |
markdown = "Output sequences:"
|
| 401 |
+
# Sequences are padded anyway so you can batch decode them
|
| 402 |
decoded_sequences = tokenizer.batch_decode(outputs.sequences)
|
| 403 |
for i, sequence in enumerate(decoded_sequences):
|
| 404 |
+
markdown += f"\n- '{clean(sequence.replace('<s> ', ''))}' (score {outputs.sequences_scores[i]:.2f})"
|
| 405 |
|
| 406 |
original_tree = generate_beams(
|
| 407 |
input_text,
|
|
|
|
| 419 |
),
|
| 420 |
css=STYLE,
|
| 421 |
) as demo:
|
| 422 |
+
gr.Markdown(
|
| 423 |
+
"""# Beam search visualizer
|
| 424 |
|
| 425 |
Play with the parameters below to understand how beam search decoding works!
|
| 426 |
|
|
|
|
| 429 |
- **Number of steps**: the number of tokens to generate
|
| 430 |
- **Number of beams**: the number of beams to use
|
| 431 |
- **Length penalty**: the length penalty to apply to outputs. `length_penalty` > 0.0 promotes longer sequences, while `length_penalty` < 0.0 encourages shorter sequences.
|
| 432 |
+
"""
|
| 433 |
+
)
|
| 434 |
+
text = gr.Textbox(
|
| 435 |
+
label="Sentence to decode from",
|
| 436 |
+
value="Conclusion: thanks a lot. This article was originally published on",
|
| 437 |
+
)
|
| 438 |
with gr.Row():
|
| 439 |
+
steps = gr.Slider(
|
| 440 |
+
label="Number of steps", minimum=1, maximum=8, step=1, value=4
|
| 441 |
+
)
|
| 442 |
+
beams = gr.Slider(
|
| 443 |
+
label="Number of beams", minimum=2, maximum=4, step=1, value=3
|
| 444 |
+
)
|
| 445 |
+
length_penalty = gr.Slider(
|
| 446 |
+
label="Length penalty", minimum=-4, maximum=4, step=0.5, value=1
|
| 447 |
+
)
|
| 448 |
button = gr.Button()
|
| 449 |
out_html = gr.Markdown()
|
| 450 |
out_markdown = gr.Markdown()
|
| 451 |
+
button.click(
|
| 452 |
+
get_beam_search_html,
|
| 453 |
+
inputs=[text, steps, beams, length_penalty],
|
| 454 |
+
outputs=[out_html, out_markdown],
|
| 455 |
+
)
|
| 456 |
|
| 457 |
demo.launch()
|