Views
No views yet
1from modeling_combine import MM_LLMs, MM_LLMs_Config
2from transformers import AutoTokenizer, AutoProcessor
3from PIL import Image
4import librosa
5import requests
6
7model = MM_LLMs.from_pretrained(
8 'mesolitica/malaysian-mistral-mmmmodal',
9 flash_attention = True,
10 dtype = torch.bfloat16,
11 torch_dtype = torch.bfloat16
12)
13_ = model.cuda()
14
15image_processor = AutoProcessor.from_pretrained('google/siglip-base-patch16-384')
16audio_processor = AutoProcessor.from_pretrained('mesolitica/malaysian-whisper-small')
17tokenizer = AutoTokenizer.from_pretrained('mesolitica/malaysian-mistral-mmmmodal')
18
19def prepare_dataset(messages, images: List[str] = None, audio: List[str] = None, sr = 16000):
20
21 if images is not None:
22 images = [Image.open(f).convert('RGB') for f in images]
23 image_output = image_processor(images=images, return_tensors='pt')['pixel_values']
24 else:
25 image_output = None
26
27 if audio is not None:
28 audio = [librosa.load(f, sr=sr)[0] for f in audio]
29 audio_features = audio_processor(audio, sampling_rate=sr, return_tensors='pt',)['input_features']
30 else:
31 audio_features = None
32
33 prompt = tokenizer.apply_chat_template(messages, tokenize = False)
34 outputs = tokenizer(
35 prompt,
36 return_tensors='pt',
37 return_overflowing_tokens=False,
38 return_length=False
39 )
40
41 outputs['images'] = image_output
42 outputs['audios'] = audio_features
43
44 image_token = tokenizer.convert_tokens_to_ids('<image>')
45 audio_token = tokenizer.convert_tokens_to_ids('<audio>')
46
47 if image_output is not None:
48 len_image = len(image_output)
49 else:
50 len_image = 0
51
52 if audio_features is not None:
53 len_audio = len(audio_features)
54 else:
55 len_audio = 0
56
57 outputs['image_index'] = torch.tensor([0] * len_image)
58 outputs['image_starts'] = torch.tensor([image_token] * (len_image + 1))
59 outputs['audio_index'] = torch.tensor([0] * len_audio)
60 outputs['audio_starts'] = torch.tensor([audio_token] * (len_audio + 1))
61
62 where_is = torch.where((outputs['input_ids'] == image_token) | (outputs['input_ids'] == audio_token))
63 ls = []
64 for i in range(len(where_is[0])):
65 b, k = where_is[0][i], where_is[1][i]
66 l = int(outputs['input_ids'][b, k])
67 ls.append(l)
68
69 ls = torch.tensor(ls)
70 outputs['where_is_b'] = where_is[0]
71 outputs['where_is_k'] = where_is[1]
72 outputs['ls'] = ls
73
74 return outputs
75
76with open('Persian-cat-breed.jpg', 'wb') as fopen:
77 fopen.write(requests.get('https://cdn.beautifulnara.net/wp-content/uploads/2017/12/10201620/Persian-cat-breed.jpg').content)
78
79with open('nasi-goreng-1-23.jpg', 'wb') as fopen:
80 fopen.write(requests.get('https://www.jocooks.com/wp-content/uploads/2023/09/nasi-goreng-1-23.jpg').content)
81
82with open('test.mp3', 'wb') as fopen:
83 fopen.write(requests.get('https://github.com/mesolitica/multimodal-LLM/raw/master/data/test.mp3').content)
84
85messages = [
86 {'role': 'user', 'content': '<image> </image> ini gambar apa'},
87]
88outputs = prepare_dataset(messages, images = ['Persian-cat-breed.jpg'])
89if outputs['images'] is not None:
90 outputs['images'] = outputs['images'].type(model.dtype)
91if outputs['audios'] is not None:
92 outputs['audios'] = outputs['audios'].type(model.dtype)
93for k in outputs.keys():
94 if outputs[k] is not None:
95 outputs[k] = outputs[k].cuda()
96
97with torch.no_grad():
98 model_inputs = model.prepare_inputs_for_generation(**outputs, inference = True)
99r = model_inputs.pop('input_ids', None)
100
101generate_kwargs = dict(
102 model_inputs,
103 max_new_tokens=300,
104 top_p=0.95,
105 top_k=50,
106 temperature=0.1,
107 do_sample=True,
108 num_beams=1,
109)
110
111r = model.llm.generate(**generate_kwargs)
112print(tokenizer.decode(r[0]))<s>Imej itu menunjukkan seekor kucing putih yang comel duduk di atas sofa hitam.</s>1messages = [
2 {'role': 'user', 'content': '<image> </image> <image> </image> apa kaitan 2 gambar ni'},
3]
4outputs = prepare_dataset(messages, images = ['Persian-cat-breed.jpg', 'nasi-goreng-1-23.jpg'])
5if outputs['images'] is not None:
6 outputs['images'] = outputs['images'].type(model.dtype)
7if outputs['audios'] is not None:
8 outputs['audios'] = outputs['audios'].type(model.dtype)
9for k in outputs.keys():
10 if outputs[k] is not None:
11 outputs[k] = outputs[k].cuda()
12
13with torch.no_grad():
14 model_inputs = model.prepare_inputs_for_generation(**outputs, inference = True)
15r = model_inputs.pop('input_ids', None)
16
17generate_kwargs = dict(
18 model_inputs,
19 max_new_tokens=300,
20 top_p=0.95,
21 top_k=50,
22 temperature=0.1,
23 do_sample=True,
24 num_beams=1,
25)
26
27r = model.llm.generate(**generate_kwargs)
28print(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>1messages = [
2 {'role': 'user', 'content': '<audio> </audio> apa isu audio ni'},
3]
4outputs = prepare_dataset(messages, audio = [audio])
5if outputs['images'] is not None:
6 outputs['images'] = outputs['images'].type(model.dtype)
7if outputs['audios'] is not None:
8 outputs['audios'] = outputs['audios'].type(model.dtype)
9for k in outputs.keys():
10 if outputs[k] is not None:
11 outputs[k] = outputs[k].cuda()
12
13with torch.no_grad():
14 model_inputs = model.prepare_inputs_for_generation(**outputs, inference = True)
15
16r = model_inputs.pop('input_ids', None)
17generate_kwargs = dict(
18 model_inputs,
19 max_new_tokens=300,
20 top_p=0.95,
21 top_k=50,
22 temperature=0.9,
23 do_sample=True,
24 num_beams=1,
25)
26
27r = model.llm.generate(**generate_kwargs)
28print(tokenizer.decode(r[0]))<s>Isu audio ini berkisar tentang persepsi salah faham dan sikap bakhil berkenaan wang dalam konteks menggalakkan penggunaan e-dompet. Penceramah mencadangkan bahawa orang mungkin keberatan untuk menerima wang kerana tidak melihat manfaat atau nilai menggunakan e-dompet, dan kebimbangan tentang tidak dapat mengakses wang itu jika mereka memerlukannya segera. Penceramah juga menyebut isu ekonomi sistem dan kekurangan sistem yang berkesan di Malaysia. Secara keseluruhannya, isu ini menekankan keperluan untuk pemahaman dan kesedaran yang lebih baik tentang faedah menggunakan e-dompet, serta keperluan untuk pembaharuan sistemik untuk memastikan akses yang saksama kepada wang dan sumber lain.</s>1messages = [
2 {'role': 'user', 'content': '<image> </image> <audio> </audio> apa kaitan gambar dan audio ni'},
3]
4outputs = prepare_dataset(messages, images = [test_image], audio = [audio])
5if outputs['images'] is not None:
6 outputs['images'] = outputs['images'].type(model.dtype)
7if outputs['audios'] is not None:
8 outputs['audios'] = outputs['audios'].type(model.dtype)
9for k in outputs.keys():
10 if outputs[k] is not None:
11 outputs[k] = outputs[k].cuda()
12
13with torch.no_grad():
14 model_inputs = model.prepare_inputs_for_generation(**outputs, inference = True)
15
16r = model_inputs.pop('input_ids', None)
17generate_kwargs = dict(
18 model_inputs,
19 max_new_tokens=300,
20 top_p=0.95,
21 top_k=50,
22 temperature=0.9,
23 do_sample=True,
24 num_beams=1,
25)
26
27r = model.llm.generate(**generate_kwargs)
28print(tokenizer.decode(r[0]))<s>Tidak jelas bagaimana gambar dan audio berkaitan antara satu sama lain. Gambar itu menunjukkan bas pelancongan dengan iklan yang menggalakkan orang ramai menggunakan e-dompet mereka, tetapi ia tidak menyatakan tujuan iklan itu. Audio itu membincangkan idea pembaziran dana sebanyak RM5 juta (kira-kira 1.2 juta USD) ke atas sesuatu projek, tetapi ia tidak menyebut secara langsung bas pelancongan atau e-dompet. Ada kemungkinan bahawa kedua-dua gambar dan audio sedang membincangkan topik yang sama, tetapi lebih banyak konteks diperlukan untuk membuat perkaitan yang pasti.</s>