Spaces:
Sleeping
Sleeping
Update inference.py
Browse files- inference.py +77 -35
inference.py
CHANGED
@@ -1,36 +1,78 @@
|
|
1 |
-
import os
|
2 |
-
import faiss
|
3 |
import torch
|
4 |
-
from
|
5 |
-
from
|
6 |
-
from
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
|
22 |
-
|
23 |
-
|
24 |
-
|
25 |
-
|
26 |
-
|
27 |
-
|
28 |
-
|
29 |
-
|
30 |
-
|
31 |
-
|
32 |
-
|
33 |
-
|
34 |
-
|
35 |
-
|
36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import torch
|
2 |
+
from evo_model import EvoTransformer
|
3 |
+
from transformers import AutoTokenizer, pipeline
|
4 |
+
from rag_utils import RAGRetriever, extract_text_from_file
|
5 |
+
import os
|
6 |
+
|
7 |
+
# Load Evo model
|
8 |
+
def load_evo_model(model_path="evo_hellaswag.pt", device=None):
|
9 |
+
if device is None:
|
10 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
11 |
+
|
12 |
+
model = EvoTransformer()
|
13 |
+
model.load_state_dict(torch.load(model_path, map_location=device))
|
14 |
+
model.to(device)
|
15 |
+
model.eval()
|
16 |
+
return model, device
|
17 |
+
|
18 |
+
evo_model, device = load_evo_model()
|
19 |
+
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
|
20 |
+
|
21 |
+
# Load GPT-3.5 (via OpenAI API)
|
22 |
+
import openai
|
23 |
+
openai.api_key = os.getenv("OPENAI_API_KEY")
|
24 |
+
|
25 |
+
# RAG Retriever
|
26 |
+
retriever = RAGRetriever()
|
27 |
+
|
28 |
+
def get_context_from_file(file_obj):
|
29 |
+
file_path = file_obj.name
|
30 |
+
text = extract_text_from_file(file_path)
|
31 |
+
retriever.add_document(text)
|
32 |
+
return text
|
33 |
+
|
34 |
+
# Evo prediction
|
35 |
+
def get_evo_response(prompt, file=None):
|
36 |
+
# Step 1: augment context if document is uploaded
|
37 |
+
context = ""
|
38 |
+
if file is not None:
|
39 |
+
context_list = retriever.retrieve(prompt)
|
40 |
+
context = "\n".join(context_list)
|
41 |
+
|
42 |
+
full_prompt = f"{prompt}\n{context}"
|
43 |
+
|
44 |
+
# Step 2: use Evo to predict
|
45 |
+
options = ["Yes, proceed with the action.", "No, maintain current strategy."]
|
46 |
+
inputs = [f"{full_prompt} {opt}" for opt in options]
|
47 |
+
|
48 |
+
encoded = tokenizer(inputs, padding=True, truncation=True, return_tensors="pt").to(device)
|
49 |
+
|
50 |
+
with torch.no_grad():
|
51 |
+
logits = evo_model(encoded["input_ids"]).squeeze(-1)
|
52 |
+
probs = torch.softmax(logits, dim=0)
|
53 |
+
best = torch.argmax(probs).item()
|
54 |
+
|
55 |
+
return f"Evo suggests: {options[best]} (Confidence: {probs[best]:.2f})"
|
56 |
+
|
57 |
+
# GPT-3.5 response
|
58 |
+
def get_gpt_response(prompt, file=None):
|
59 |
+
context = ""
|
60 |
+
if file is not None:
|
61 |
+
context_list = retriever.retrieve(prompt)
|
62 |
+
context = "\n".join(context_list)
|
63 |
+
|
64 |
+
full_prompt = (
|
65 |
+
f"Question: {prompt}\n"
|
66 |
+
f"Relevant Context:\n{context}\n"
|
67 |
+
f"Answer like a financial advisor."
|
68 |
+
)
|
69 |
+
|
70 |
+
response = openai.ChatCompletion.create(
|
71 |
+
model="gpt-3.5-turbo",
|
72 |
+
messages=[
|
73 |
+
{"role": "user", "content": full_prompt}
|
74 |
+
],
|
75 |
+
temperature=0.4,
|
76 |
+
)
|
77 |
+
|
78 |
+
return response.choices[0].message.content.strip()
|