Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import GPT2Tokenizer, GPT2LMHeadModel
|
2 |
+
import torch
|
3 |
+
tokenizer = AutoTokenizer.from_pretrained("af1tang/personaGPT")
|
4 |
+
model = AutoModelForCausalLM.from_pretrained("af1tang/personaGPT")
|
5 |
+
if torch.cuda.is_available():
|
6 |
+
model = model.cuda()
|
7 |
+
## utility functions ##
|
8 |
+
flatten = lambda l: [item for sublist in l for item in sublist]
|
9 |
+
|
10 |
+
def to_data(x):
|
11 |
+
if torch.cuda.is_available():
|
12 |
+
x = x.cpu()
|
13 |
+
return x.data.numpy()
|
14 |
+
|
15 |
+
def to_var(x):
|
16 |
+
if not torch.is_tensor(x):
|
17 |
+
x = torch.Tensor(x)
|
18 |
+
if torch.cuda.is_available():
|
19 |
+
x = x.cuda()
|
20 |
+
return x
|
21 |
+
|
22 |
+
def display_dialog_history(dialog_hx):
|
23 |
+
for j, line in enumerate(dialog_hx):
|
24 |
+
msg = tokenizer.decode(line)
|
25 |
+
if j %2 == 0:
|
26 |
+
print(">> User: "+ msg)
|
27 |
+
else:
|
28 |
+
print("Bot: "+msg)
|
29 |
+
print()
|
30 |
+
|
31 |
+
def generate_next(bot_input_ids, do_sample=True, top_k=10, top_p=.92,
|
32 |
+
max_length=1000, pad_token=tokenizer.eos_token_id):
|
33 |
+
full_msg = model.generate(bot_input_ids, do_sample=True,
|
34 |
+
top_k=top_k, top_p=top_p,
|
35 |
+
max_length=max_length, pad_token_id=tokenizer.eos_token_id)
|
36 |
+
msg = to_data(full_msg.detach()[0])[bot_input_ids.shape[-1]:]
|
37 |
+
return msg
|
38 |
+
|
39 |
+
|
40 |
+
|
41 |
+
|
42 |
+
|
43 |
+
# get personality facts for conversation
|
44 |
+
personas = []
|
45 |
+
for i in range(3):
|
46 |
+
response = input(">> Fact %d: "%(i+1))+ tokenizer.eos_token
|
47 |
+
personas.append(response)
|
48 |
+
personas = tokenizer.encode(''.join(['<|p2|>'] + personas + ['<|sep|>'] + ['<|start|>']))
|
49 |
+
|
50 |
+
|
51 |
+
|