mesolitica/malaysian-tinyllama-1.1b-siglip-base-384-vision

Malaysian TinyLlama + siglip-base-patch16-384

WanDB https://wandb.ai/huseinzol05/vision-tinyllama?workspace=user-huseinzol05

how-to

from modeling_vision import MM_LLMs, MM_LLMs_Config
from transformers import AutoTokenizer, AutoProcessor
from PIL import Image
import requests

model = MM_LLMs.from_pretrained(
    'mesolitica/malaysian-tinyllama-1.1b-siglip-base-384-vision',
    flash_attention = True,
    dtype = torch.bfloat16,
    torch_dtype = torch.bfloat16
)
_ = model.cuda()

image_processor = AutoProcessor.from_pretrained('google/siglip-base-patch16-384')
tokenizer = AutoTokenizer.from_pretrained('mesolitica/malaysian-tinyllama-1.1b-siglip-base-384-vision')

def prepare_dataset(messages, images: List[str] = None):
    if images is not None:
        images = [Image.open(f).convert('RGB') for f in images]
        image_output = image_processor(images=images, return_tensors='pt')['pixel_values']
    else:
        image_output = None
    
    prompt = tokenizer.apply_chat_template(messages, tokenize = False)
    outputs = tokenizer(
                    prompt,
                    return_tensors='pt',
                    return_overflowing_tokens=False,
                    return_length=False)

    outputs['images'] = image_output
    outputs['image_index'] = torch.tensor([0] * len(outputs['images']))
    outputs['image_starts'] = torch.tensor([tokenizer.convert_tokens_to_ids('<image>')] * len(outputs['images']))
    return outputs

with open('Persian-cat-breed.jpg', 'wb') as fopen:
    fopen.write(requests.get('https://cdn.beautifulnara.net/wp-content/uploads/2017/12/10201620/Persian-cat-breed.jpg').content)

with open('nasi-goreng-1-23.jpg', 'wb') as fopen:
    fopen.write(requests.get('https://www.jocooks.com/wp-content/uploads/2023/09/nasi-goreng-1-23.jpg').content)

messages = [
    {'role': 'user', 'content': '<image> </image> ini gambar apa'},
]
outputs = prepare_dataset(messages, images = ['Persian-cat-breed.jpg'])
outputs['images'] = outputs['images'].type(model.dtype)
for k in outputs.keys():
    if outputs[k] is not None:
        outputs[k] = outputs[k].cuda()

with torch.no_grad():
    model_inputs = model.prepare_inputs_for_generation(**outputs)
r = model_inputs.pop('input_ids', None)

generate_kwargs = dict(
    model_inputs,
    max_new_tokens=300,
    top_p=0.95,
    top_k=50,
    temperature=0.1,
    do_sample=True,
    num_beams=1,
)

r = model.llm.generate(**generate_kwargs)
print(tokenizer.decode(r[0]))

<s>Imej itu menunjukkan seekor kucing putih yang comel duduk di atas sofa hitam.</s>

messages = [
    {'role': 'user', 'content': '<image> </image> <image> </image> apa kaitan 2 gambar ni'},
]
outputs = prepare_dataset(messages, images = ['Persian-cat-breed.jpg', 'nasi-goreng-1-23.jpg'])
outputs['images'] = outputs['images'].type(model.dtype)
for k in outputs.keys():
    if outputs[k] is not None:
        outputs[k] = outputs[k].cuda()

with torch.no_grad():
    model_inputs = model.prepare_inputs_for_generation(**outputs)
r = model_inputs.pop('input_ids', None)

generate_kwargs = dict(
    model_inputs,
    max_new_tokens=300,
    top_p=0.95,
    top_k=50,
    temperature=0.1,
    do_sample=True,
    num_beams=1,
)

r = model.llm.generate(**generate_kwargs)
print(tokenizer.decode(r[0]))

<s>Tiada hubungan yang jelas antara gambar 1 (anak kucing putih duduk di atas sofa) dan gambar 2 (foto penutup mangkuk mi telur dengan nasi dan cili). Gambar pertama ialah imej haiwan, manakala gambar kedua ialah imej makanan. Mereka tergolong dalam kategori yang berbeza dan tidak mempunyai hubungan antara satu sama lain.</s>