Update app.py
Browse files
app.py
CHANGED
@@ -77,7 +77,7 @@ def proc_submission(
|
|
77 |
|
78 |
if processed["was_truncated"]:
|
79 |
tr_in = processed["truncated_text"]
|
80 |
-
msg = f"Input text was truncated to {max_input_length} words to fit within the computational constraints"
|
81 |
logging.warning(msg)
|
82 |
history["WARNING"] = msg
|
83 |
else:
|
@@ -92,18 +92,12 @@ def proc_submission(
|
|
92 |
**settings,
|
93 |
)
|
94 |
sum_text = [f"Section {i}: " + s["summary"][0] for i, s in enumerate(_summaries)]
|
95 |
-
sum_scores = [
|
96 |
-
f" - Section {i}: {round(s['summary_score'],4)}"
|
97 |
-
for i, s in enumerate(_summaries)
|
98 |
-
]
|
99 |
rates = [
|
100 |
f" - Section {i}: {round(s['compression_rate'],3)}"
|
101 |
for i, s in enumerate(_summaries)
|
102 |
]
|
103 |
|
104 |
sum_text_out = "\n".join(sum_text)
|
105 |
-
history["Summary Scores"] = "<br><br>"
|
106 |
-
scores_out = "\n".join(sum_scores)
|
107 |
history["Compression Rates"] = "<br><br>"
|
108 |
rates_out = "\n".join(rates)
|
109 |
rt = round((time.perf_counter() - st) / 60, 2)
|
@@ -188,7 +182,7 @@ if __name__ == "__main__":
|
|
188 |
)
|
189 |
num_beams = gr.Radio(
|
190 |
choices=[2, 3, 4],
|
191 |
-
label="Beam Search:
|
192 |
value=2,
|
193 |
)
|
194 |
gr.Markdown(
|
@@ -249,12 +243,6 @@ if __name__ == "__main__":
|
|
249 |
gr.Markdown(
|
250 |
"The summary scores can be thought of as representing the quality of the summary. less-negative numbers (closer to 0) are better:"
|
251 |
)
|
252 |
-
summary_scores = gr.Textbox(
|
253 |
-
label="Summary Scores ", placeholder="Summary scores will appear here"
|
254 |
-
)
|
255 |
-
gr.Markdown(
|
256 |
-
"The compression rate indicates the ratio between the machine-generated summary length and the input text (from 0% to 100%). The higher the compression rate the more extreme the summary is."
|
257 |
-
)
|
258 |
compression_rate = gr.Textbox(
|
259 |
label="Compression rate π", placeholder="The π will appear here"
|
260 |
)
|
@@ -266,7 +254,7 @@ if __name__ == "__main__":
|
|
266 |
"- [Blaise-g/longt5_tglobal_large_sumpubmed](https://huggingface.co/Blaise-g/longt5_tglobal_large_sumpubmed) is a fine-tuned checkpoint of [Stancld/longt5-tglobal-large-16384-pubmed-3k_steps](https://huggingface.co/Stancld/longt5-tglobal-large-16384-pubmed-3k_steps) on the [SumPubMed dataset](https://aclanthology.org/2021.acl-srw.30/). [Blaise-g/longt5_tglobal_large_scitldr](https://huggingface.co/Blaise-g/longt5_tglobal_large_scitldr) is a fine-tuned checkpoint of [Blaise-g/longt5_tglobal_large_sumpubmed](https://huggingface.co/Blaise-g/longt5_tglobal_large_sumpubmed) on the [Scitldr dataset](https://arxiv.org/abs/2004.15011). The goal was to create two models capable of handling the complex information contained in long biomedical documents and subsequently producing scientific summaries according to one of the two possible levels of conciseness: 1) A long explanatory synopsis that retains the majority of domain-specific language used in the original source text. 2)A one sentence long, TLDR style summary."
|
267 |
)
|
268 |
gr.Markdown(
|
269 |
-
"- The two most important text generation parameters are the `num_beams` and
|
270 |
)
|
271 |
gr.Markdown("---")
|
272 |
|
@@ -287,7 +275,7 @@ if __name__ == "__main__":
|
|
287 |
token_batch_length,
|
288 |
length_penalty,
|
289 |
],
|
290 |
-
outputs=[output_text, summary_text,
|
291 |
)
|
292 |
|
293 |
demo.launch(enable_queue=True, share=False)
|
|
|
77 |
|
78 |
if processed["was_truncated"]:
|
79 |
tr_in = processed["truncated_text"]
|
80 |
+
msg = f"Input text was truncated to {max_input_length} words to fit within the computational constraints of the inference API"
|
81 |
logging.warning(msg)
|
82 |
history["WARNING"] = msg
|
83 |
else:
|
|
|
92 |
**settings,
|
93 |
)
|
94 |
sum_text = [f"Section {i}: " + s["summary"][0] for i, s in enumerate(_summaries)]
|
|
|
|
|
|
|
|
|
95 |
rates = [
|
96 |
f" - Section {i}: {round(s['compression_rate'],3)}"
|
97 |
for i, s in enumerate(_summaries)
|
98 |
]
|
99 |
|
100 |
sum_text_out = "\n".join(sum_text)
|
|
|
|
|
101 |
history["Compression Rates"] = "<br><br>"
|
102 |
rates_out = "\n".join(rates)
|
103 |
rt = round((time.perf_counter() - st) / 60, 2)
|
|
|
182 |
)
|
183 |
num_beams = gr.Radio(
|
184 |
choices=[2, 3, 4],
|
185 |
+
label="Beam Search: Number of Beams",
|
186 |
value=2,
|
187 |
)
|
188 |
gr.Markdown(
|
|
|
243 |
gr.Markdown(
|
244 |
"The summary scores can be thought of as representing the quality of the summary. less-negative numbers (closer to 0) are better:"
|
245 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
246 |
compression_rate = gr.Textbox(
|
247 |
label="Compression rate π", placeholder="The π will appear here"
|
248 |
)
|
|
|
254 |
"- [Blaise-g/longt5_tglobal_large_sumpubmed](https://huggingface.co/Blaise-g/longt5_tglobal_large_sumpubmed) is a fine-tuned checkpoint of [Stancld/longt5-tglobal-large-16384-pubmed-3k_steps](https://huggingface.co/Stancld/longt5-tglobal-large-16384-pubmed-3k_steps) on the [SumPubMed dataset](https://aclanthology.org/2021.acl-srw.30/). [Blaise-g/longt5_tglobal_large_scitldr](https://huggingface.co/Blaise-g/longt5_tglobal_large_scitldr) is a fine-tuned checkpoint of [Blaise-g/longt5_tglobal_large_sumpubmed](https://huggingface.co/Blaise-g/longt5_tglobal_large_sumpubmed) on the [Scitldr dataset](https://arxiv.org/abs/2004.15011). The goal was to create two models capable of handling the complex information contained in long biomedical documents and subsequently producing scientific summaries according to one of the two possible levels of conciseness: 1) A long explanatory synopsis that retains the majority of domain-specific language used in the original source text. 2)A one sentence long, TLDR style summary."
|
255 |
)
|
256 |
gr.Markdown(
|
257 |
+
"- The two most important text generation parameters are the `num_beams` and 'length_penalty': 1) Choosing a higher number of beams for the beam search algorithm results in generating a summary with higher probability (hence theoretically higher quality) at the cost of increasing computation times and memory usage. 2) The length penalty encourages the model to generate longer or shorter summary sequences by placing an exponential penalty on the beam score according to the current sequence length."
|
258 |
)
|
259 |
gr.Markdown("---")
|
260 |
|
|
|
275 |
token_batch_length,
|
276 |
length_penalty,
|
277 |
],
|
278 |
+
outputs=[output_text, summary_text, compression_rate],
|
279 |
)
|
280 |
|
281 |
demo.launch(enable_queue=True, share=False)
|