Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,54 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
|
3 |
+
import gradio as gr
|
4 |
+
import torch.nn.functional as F
|
5 |
+
|
6 |
+
from transformers import BertTokenizer, GPT2LMHeadModel
|
7 |
+
|
8 |
+
tokenizer = BertTokenizer.from_pretrained("supermy/couplet-gpt2")
|
9 |
+
model = GPT2LMHeadModel.from_pretrained("supermy/couplet-gpt2")
|
10 |
+
model.eval()
|
11 |
+
|
12 |
+
def top_k_top_p_filtering( logits, top_k=0, top_p=0.0, filter_value=-float('Inf') ):
|
13 |
+
assert logits.dim() == 1
|
14 |
+
top_k = min( top_k, logits.size(-1) )
|
15 |
+
if top_k > 0:
|
16 |
+
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
|
17 |
+
logits[indices_to_remove] = filter_value
|
18 |
+
if top_p > 0.0:
|
19 |
+
sorted_logits, sorted_indices = torch.sort(logits, descending=True)
|
20 |
+
cumulative_probs = torch.cumsum( F.softmax(sorted_logits, dim=-1), dim=-1 )
|
21 |
+
sorted_indices_to_remove = cumulative_probs > top_p
|
22 |
+
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
|
23 |
+
sorted_indices_to_remove[..., 0] = 0
|
24 |
+
indices_to_remove = sorted_indices[sorted_indices_to_remove]
|
25 |
+
logits[indices_to_remove] = filter_value
|
26 |
+
return logits
|
27 |
+
|
28 |
+
def generate(input_text):
|
29 |
+
input_ids = [tokenizer.cls_token_id]
|
30 |
+
input_ids.extend( tokenizer.encode(input_text + "-", add_special_tokens=False) )
|
31 |
+
input_ids = torch.tensor( [input_ids] )
|
32 |
+
|
33 |
+
generated = []
|
34 |
+
for _ in range(100):
|
35 |
+
output = model(input_ids)
|
36 |
+
|
37 |
+
next_token_logits = output.logits[0, -1, :]
|
38 |
+
next_token_logits[ tokenizer.convert_tokens_to_ids('[UNK]') ] = -float('Inf')
|
39 |
+
filtered_logits = top_k_top_p_filtering(next_token_logits, top_k=8, top_p=1)
|
40 |
+
next_token = torch.multinomial( F.softmax(filtered_logits, dim=-1), num_samples=1 )
|
41 |
+
if next_token == tokenizer.sep_token_id:
|
42 |
+
break
|
43 |
+
generated.append( next_token.item() )
|
44 |
+
input_ids = torch.cat( (input_ids, next_token.unsqueeze(0)), dim=1 )
|
45 |
+
|
46 |
+
return "".join( tokenizer.convert_ids_to_tokens(generated) )
|
47 |
+
|
48 |
+
if __name__ == "__main__":
|
49 |
+
|
50 |
+
gr.Interface(
|
51 |
+
fn=generate,
|
52 |
+
inputs="text",
|
53 |
+
outputs="text"
|
54 |
+
).launch()
|