Rename gpt_model.py to qwen_model.py
Browse files- gpt_model.py +0 -30
- qwen_model.py +52 -0
gpt_model.py
DELETED
@@ -1,30 +0,0 @@
|
|
1 |
-
import openai
|
2 |
-
|
3 |
-
|
4 |
-
def generate_response(retrieved_texts, query, max_tokens=150):
|
5 |
-
"""
|
6 |
-
Generates a response based on the retrieved texts and query.
|
7 |
-
|
8 |
-
Args:
|
9 |
-
retrieved_texts (list): List of retrieved text strings.
|
10 |
-
query (str): Query string.
|
11 |
-
max_tokens (int): Maximum number of tokens for the response.
|
12 |
-
|
13 |
-
Returns:
|
14 |
-
str: Generated response.
|
15 |
-
"""
|
16 |
-
context = "\n".join(retrieved_texts)
|
17 |
-
prompt = f"This is the detail about the image: {context}\n\nQuestion: {query}\n\nAnswer:"
|
18 |
-
|
19 |
-
response = openai.ChatCompletion.create(
|
20 |
-
model="gpt-3.5-turbo",
|
21 |
-
messages=[
|
22 |
-
{"role": "system", "content": "You are a helpful assistant."},
|
23 |
-
{"role": "user", "content": prompt}
|
24 |
-
],
|
25 |
-
max_tokens=max_tokens,
|
26 |
-
n=1,
|
27 |
-
stop=None,
|
28 |
-
temperature=0.5,
|
29 |
-
)
|
30 |
-
return response.choices[0].message['content']
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
qwen_model.py
ADDED
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
|
2 |
+
|
3 |
+
# Replace with your target Qwen model on Hugging Face
|
4 |
+
MODEL_NAME = "Qwen/Qwen2.5-7B-Instruct-1M"
|
5 |
+
|
6 |
+
# Initialize tokenizer and model
|
7 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
|
8 |
+
model = AutoModelForCausalLM.from_pretrained(
|
9 |
+
MODEL_NAME,
|
10 |
+
device_map="auto", # or "cuda", etc. if you want to specify
|
11 |
+
trust_remote_code=True
|
12 |
+
)
|
13 |
+
|
14 |
+
# Create pipeline
|
15 |
+
qwen_pipeline = pipeline(
|
16 |
+
"text-generation",
|
17 |
+
model=model,
|
18 |
+
tokenizer=tokenizer
|
19 |
+
)
|
20 |
+
|
21 |
+
def generate_response(retrieved_texts, query, max_new_tokens=200):
|
22 |
+
"""
|
23 |
+
Generates a response based on the retrieved texts and query using Qwen.
|
24 |
+
Args:
|
25 |
+
retrieved_texts (list): List of retrieved text strings (e.g., from BLIP).
|
26 |
+
query (str): The user's question about the image.
|
27 |
+
max_new_tokens (int): Maximum tokens to generate for the answer.
|
28 |
+
Returns:
|
29 |
+
str: The generated answer.
|
30 |
+
"""
|
31 |
+
# Construct a prompt that includes the image details as context
|
32 |
+
context = "\n".join(retrieved_texts)
|
33 |
+
prompt = f"This is the detail about the image:\n{context}\n\nQuestion: {query}\nAnswer:"
|
34 |
+
|
35 |
+
# Generate the text
|
36 |
+
result = qwen_pipeline(
|
37 |
+
prompt,
|
38 |
+
max_new_tokens=max_new_tokens,
|
39 |
+
do_sample=True, # or False if you want deterministic output
|
40 |
+
temperature=0.7, # tweak as needed
|
41 |
+
)
|
42 |
+
|
43 |
+
# The pipeline returns a list of dicts with key "generated_text"
|
44 |
+
full_generation = result[0]["generated_text"]
|
45 |
+
|
46 |
+
# Optionally parse out the final answer if the model repeats the prompt
|
47 |
+
if "Answer:" in full_generation:
|
48 |
+
final_answer = full_generation.split("Answer:")[-1].strip()
|
49 |
+
else:
|
50 |
+
final_answer = full_generation
|
51 |
+
|
52 |
+
return final_answer
|