Views
No views yet
google/medgemma-4b-it, republished under Convai Innovations for fine-tuning and experimentation in medical multimodal AI.1# pip install accelerate
2from transformers import AutoProcessor, AutoModelForImageTextToText
3from PIL import Image
4import requests
5import torch
6
7model_id = "convaiinnovations/gemma-finetuned-4b-it"
8
9model = AutoModelForImageTextToText.from_pretrained(
10 model_id,
11 torch_dtype=torch.bfloat16,
12 device_map="auto",
13)
14processor = AutoProcessor.from_pretrained(model_id)
15
16# Image attribution: Stillwaterising, CC0, via Wikimedia Commons
17image_url = "https://upload.wikimedia.org/wikipedia/commons/c/c8/Chest_Xray_PA_3-8-2010.png"
18image = Image.open(requests.get(image_url, headers={"User-Agent": "example"}, stream=True).raw)
19
20messages = [
21 {
22 "role": "system",
23 "content": [{"type": "text", "text": "You are an expert radiologist."}]
24 },
25 {
26 "role": "user",
27 "content": [
28 {"type": "text", "text": "Describe this X-ray"},
29 {"type": "image", "image": image}
30 ]
31 }
32]
33
34inputs = processor.apply_chat_template(
35 messages, add_generation_prompt=True, tokenize=True,
36 return_dict=True, return_tensors="pt"
37).to(model.device, dtype=torch.bfloat16)
38
39input_len = inputs["input_ids"].shape[-1]
40
41with torch.inference_mode():
42 generation = model.generate(**inputs, max_new_tokens=200, do_sample=False)
43 generation = generation[0][input_len:]
44
45decoded = processor.decode(generation, skip_special_tokens=True)
46print(decoded)