Spaces:
Sleeping
Sleeping
Commit
β’
944d9b4
1
Parent(s):
e59f863
Update app.py
Browse files
app.py
CHANGED
@@ -32,7 +32,45 @@ def infer(input_ids, max_length, temperature, top_k, top_p):
|
|
32 |
|
33 |
return output_sequences
|
34 |
|
35 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
36 |
|
37 |
'''
|
38 |
'''
|
@@ -41,40 +79,47 @@ default_value = test_book[0]['book']
|
|
41 |
st.title("Book Summarization π")
|
42 |
st.write("The almighty king of text generation, GPT-2 comes in four available sizes, only three of which have been publicly made available. Feared for its fake news generation capabilities, it currently stands as the most syntactically coherent model. A direct successor to the original GPT, it reinforces the already established pre-training/fine-tuning killer duo. From the paper: Language Models are Unsupervised Multitask Learners by Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei and Ilya Sutskever.")
|
43 |
|
44 |
-
|
|
|
|
|
|
|
|
|
|
|
45 |
max_length = st.sidebar.slider("Max Length", value = 512,min_value = 10, max_value=1024)
|
46 |
temperature = st.sidebar.slider("Temperature", value = 1.0, min_value = 0.0, max_value=1.0, step=0.05)
|
47 |
top_k = st.sidebar.slider("Top-k", min_value = 0, max_value=5, value = 0)
|
48 |
top_p = st.sidebar.slider("Top-p", min_value = 0.0, max_value=1.0, step = 0.05, value = 0.92)
|
49 |
|
50 |
-
encoded_prompt = tokenizer.encode(sent, add_special_tokens=False, return_tensors="pt")
|
51 |
-
if encoded_prompt.size()[-1] == 0:
|
52 |
-
input_ids = None
|
53 |
-
else:
|
54 |
-
input_ids = encoded_prompt
|
55 |
-
|
56 |
-
|
57 |
-
output_sequences = infer(input_ids, max_length, temperature, top_k, top_p)
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
for generated_sequence_idx, generated_sequence in enumerate(output_sequences):
|
62 |
-
print(f"=== GENERATED SEQUENCE {generated_sequence_idx + 1} ===")
|
63 |
-
generated_sequences = generated_sequence.tolist()
|
64 |
-
|
65 |
-
# Decode text
|
66 |
-
text = tokenizer.decode(generated_sequence, clean_up_tokenization_spaces=True)
|
67 |
-
|
68 |
-
# Remove all text after the stop token
|
69 |
-
#text = text[: text.find(args.stop_token) if args.stop_token else None]
|
70 |
-
|
71 |
-
# Add the prompt at the beginning of the sequence. Remove the excess text that was used for pre-processing
|
72 |
-
total_sequence = (
|
73 |
-
sent + text[len(tokenizer.decode(encoded_prompt[0], clean_up_tokenization_spaces=True)) :]
|
74 |
-
)
|
75 |
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
32 |
|
33 |
return output_sequences
|
34 |
|
35 |
+
|
36 |
+
def chunking(book_text):
|
37 |
+
segments = []
|
38 |
+
#sentences, token_lens
|
39 |
+
current_segment = ""
|
40 |
+
total_token_lens = 0
|
41 |
+
for i in range(len(sentences)):
|
42 |
+
|
43 |
+
if total_token_lens < 512:
|
44 |
+
total_token_lens += token_lens[i]
|
45 |
+
current_segment += (sentences[i] + " ")
|
46 |
+
|
47 |
+
elif total_token_lens > 768:
|
48 |
+
segments.append(current_segment)
|
49 |
+
current_segment = sentences[i]
|
50 |
+
total_token_lens = token_lens[i]
|
51 |
+
|
52 |
+
else:
|
53 |
+
#make next_pseudo_segment
|
54 |
+
next_pseudo_segment = ""
|
55 |
+
next_token_len = 0
|
56 |
+
for t in range(30):
|
57 |
+
if (i+t < len(sentences)) and (next_token_len + token_lens[i+t] < 512):
|
58 |
+
next_token_len += token_lens[i+t]
|
59 |
+
next_pseudo_segment += sentences[i+t]
|
60 |
+
|
61 |
+
embs = model.encode([current_segment, next_pseudo_segment, sentences[i]]) # current, next, sent
|
62 |
+
if cos_similarity(embs[1],embs[2]) > cos_similarity(embs[0],embs[2]):
|
63 |
+
segments.append(current_segment)
|
64 |
+
current_segment = sentences[i]
|
65 |
+
total_token_lens = token_lens[i]
|
66 |
+
else:
|
67 |
+
total_token_lens += token_lens[i]
|
68 |
+
current_segment += (sentences[i] + " ")
|
69 |
+
|
70 |
+
return segments
|
71 |
+
|
72 |
+
|
73 |
+
chunked_segments = chunking(test_book[0]['book'])
|
74 |
|
75 |
'''
|
76 |
'''
|
|
|
79 |
st.title("Book Summarization π")
|
80 |
st.write("The almighty king of text generation, GPT-2 comes in four available sizes, only three of which have been publicly made available. Feared for its fake news generation capabilities, it currently stands as the most syntactically coherent model. A direct successor to the original GPT, it reinforces the already established pre-training/fine-tuning killer duo. From the paper: Language Models are Unsupervised Multitask Learners by Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei and Ilya Sutskever.")
|
81 |
|
82 |
+
book_index = st.sidebar.slider("Select Book Example", value = 0,min_value = 0, max_value=4)
|
83 |
+
|
84 |
+
_book = test_book[book_index]['book']
|
85 |
+
chunked_segments = chunking(_book)
|
86 |
+
|
87 |
+
sent = st.text_area("Text", _book, height = 550)
|
88 |
max_length = st.sidebar.slider("Max Length", value = 512,min_value = 10, max_value=1024)
|
89 |
temperature = st.sidebar.slider("Temperature", value = 1.0, min_value = 0.0, max_value=1.0, step=0.05)
|
90 |
top_k = st.sidebar.slider("Top-k", min_value = 0, max_value=5, value = 0)
|
91 |
top_p = st.sidebar.slider("Top-p", min_value = 0.0, max_value=1.0, step = 0.05, value = 0.92)
|
92 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
93 |
|
94 |
+
for segment in range(len(chunked_segments)):
|
95 |
+
|
96 |
+
encoded_prompt = tokenizer.encode(segment, add_special_tokens=False, return_tensors="pt")
|
97 |
+
if encoded_prompt.size()[-1] == 0:
|
98 |
+
input_ids = None
|
99 |
+
else:
|
100 |
+
input_ids = encoded_prompt
|
101 |
+
|
102 |
+
|
103 |
+
output_sequences = infer(input_ids, max_length, temperature, top_k, top_p)
|
104 |
+
|
105 |
+
|
106 |
+
for generated_sequence_idx, generated_sequence in enumerate(output_sequences):
|
107 |
+
print(f"=== GENERATED SEQUENCE {generated_sequence_idx + 1} ===")
|
108 |
+
generated_sequences = generated_sequence.tolist()
|
109 |
+
|
110 |
+
# Decode text
|
111 |
+
text = tokenizer.decode(generated_sequence, clean_up_tokenization_spaces=True)
|
112 |
+
|
113 |
+
# Remove all text after the stop token
|
114 |
+
#text = text[: text.find(args.stop_token) if args.stop_token else None]
|
115 |
+
|
116 |
+
# Add the prompt at the beginning of the sequence. Remove the excess text that was used for pre-processing
|
117 |
+
total_sequence = (
|
118 |
+
sent + text[len(tokenizer.decode(encoded_prompt[0], clean_up_tokenization_spaces=True)) :]
|
119 |
+
)
|
120 |
+
|
121 |
+
generated_sequences.append(total_sequence)
|
122 |
+
print(total_sequence)
|
123 |
+
|
124 |
+
|
125 |
+
st.write(generated_sequences[-1])
|