Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -272,7 +272,7 @@ class BeamNode:
|
|
272 |
is_selected_sequence: bool
|
273 |
|
274 |
|
275 |
-
def generate_beams(start_sentence, scores, length_penalty, decoded_sequences, beam_indexes_source):
|
276 |
original_tree = BeamNode(
|
277 |
cumulative_score=0,
|
278 |
current_token_ix=None,
|
@@ -284,7 +284,6 @@ def generate_beams(start_sentence, scores, length_penalty, decoded_sequences, be
|
|
284 |
is_final=False,
|
285 |
is_selected_sequence=False,
|
286 |
)
|
287 |
-
n_beams = len(scores[0])
|
288 |
beam_trees = [original_tree] * n_beams
|
289 |
generation_length = len(scores)
|
290 |
|
@@ -429,7 +428,7 @@ def get_beam_search_html(
|
|
429 |
outputs = model.generate(
|
430 |
**inputs,
|
431 |
max_new_tokens=number_steps,
|
432 |
-
num_beams=number_beams,
|
433 |
num_return_sequences=num_return_sequences,
|
434 |
return_dict_in_generate=True,
|
435 |
length_penalty=length_penalty,
|
@@ -447,6 +446,7 @@ def get_beam_search_html(
|
|
447 |
markdown += f"\n- Score `{outputs.sequences_scores[i]:.2f}`: `{clean(sequence.replace('<s> ', ''))}`"
|
448 |
|
449 |
original_tree = generate_beams(
|
|
|
450 |
input_text,
|
451 |
outputs.scores[:],
|
452 |
length_penalty,
|
@@ -493,7 +493,7 @@ This parameter will not impact the beam search paths, but only influence the cho
|
|
493 |
label="Number of steps", minimum=1, maximum=12, step=1, value=5
|
494 |
)
|
495 |
n_beams = gr.Slider(
|
496 |
-
label="Number of beams", minimum=
|
497 |
)
|
498 |
length_penalty = gr.Slider(
|
499 |
label="Length penalty", minimum=-3, maximum=3, step=0.5, value=1
|
|
|
272 |
is_selected_sequence: bool
|
273 |
|
274 |
|
275 |
+
def generate_beams(n_beams, start_sentence, scores, length_penalty, decoded_sequences, beam_indexes_source):
|
276 |
original_tree = BeamNode(
|
277 |
cumulative_score=0,
|
278 |
current_token_ix=None,
|
|
|
284 |
is_final=False,
|
285 |
is_selected_sequence=False,
|
286 |
)
|
|
|
287 |
beam_trees = [original_tree] * n_beams
|
288 |
generation_length = len(scores)
|
289 |
|
|
|
428 |
outputs = model.generate(
|
429 |
**inputs,
|
430 |
max_new_tokens=number_steps,
|
431 |
+
num_beams=max(number_beams, 2),
|
432 |
num_return_sequences=num_return_sequences,
|
433 |
return_dict_in_generate=True,
|
434 |
length_penalty=length_penalty,
|
|
|
446 |
markdown += f"\n- Score `{outputs.sequences_scores[i]:.2f}`: `{clean(sequence.replace('<s> ', ''))}`"
|
447 |
|
448 |
original_tree = generate_beams(
|
449 |
+
number_beams,
|
450 |
input_text,
|
451 |
outputs.scores[:],
|
452 |
length_penalty,
|
|
|
493 |
label="Number of steps", minimum=1, maximum=12, step=1, value=5
|
494 |
)
|
495 |
n_beams = gr.Slider(
|
496 |
+
label="Number of beams", minimum=1, maximum=4, step=1, value=4
|
497 |
)
|
498 |
length_penalty = gr.Slider(
|
499 |
label="Length penalty", minimum=-3, maximum=3, step=0.5, value=1
|