File size: 8,390 Bytes
8235038
 
 
 
 
d3631b5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5e15b3f
d3631b5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5e15b3f
d3631b5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
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
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
---
library_name: transformers
tags: []
---

# Multi-images Multi-audio Multi-turn Malaysian 7B Mistral

WanDB https://wandb.ai/huseinzol05/multimodal-mistral?workspace=user-huseinzol05

## how-to

```python
from modeling_combine import MM_LLMs, MM_LLMs_Config
from transformers import AutoTokenizer, AutoProcessor
from PIL import Image
import librosa
import requests

model = MM_LLMs.from_pretrained(
    'mesolitica/malaysian-mistral-mmmmodal',
    flash_attention = True,
    dtype = torch.bfloat16,
    torch_dtype = torch.bfloat16
)
_ = model.cuda()

image_processor = AutoProcessor.from_pretrained('google/siglip-base-patch16-384')
audio_processor = AutoProcessor.from_pretrained('mesolitica/malaysian-whisper-small')
tokenizer = AutoTokenizer.from_pretrained('mesolitica/malaysian-mistral-mmmmodal')

def prepare_dataset(messages, images: List[str] = None, audio: List[str] = None, sr = 16000):
    
    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
        
    if audio is not None:
        audio = [librosa.load(f, sr=sr)[0] for f in audio]
        audio_features = audio_processor(audio, sampling_rate=sr, return_tensors='pt',)['input_features']
    else:
        audio_features = 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['audios'] = audio_features
    
    image_token = tokenizer.convert_tokens_to_ids('<image>')
    audio_token = tokenizer.convert_tokens_to_ids('<audio>')
    
    if image_output is not None:
        len_image = len(image_output)
    else:
        len_image = 0
        
    if audio_features is not None:
        len_audio = len(audio_features)
    else:
        len_audio = 0
        
    outputs['image_index'] = torch.tensor([0] * len_image)
    outputs['image_starts'] = torch.tensor([image_token] * (len_image + 1))
    outputs['audio_index'] = torch.tensor([0] * len_audio)
    outputs['audio_starts'] = torch.tensor([audio_token] * (len_audio + 1))
        
    where_is = torch.where((outputs['input_ids'] == image_token) | (outputs['input_ids'] == audio_token))
    ls = []
    for i in range(len(where_is[0])):
        b, k = where_is[0][i], where_is[1][i]
        l = int(outputs['input_ids'][b, k])
        ls.append(l)

    ls = torch.tensor(ls)
    outputs['where_is_b'] = where_is[0]
    outputs['where_is_k'] = where_is[1]
    outputs['ls'] = ls
        
    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)

with open('test.mp3', 'wb') as fopen:
    fopen.write(requests.get('https://github.com/mesolitica/multimodal-LLM/raw/master/data/test.mp3').content)

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

```python
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'])
if outputs['images'] is not None:
    outputs['images'] = outputs['images'].type(model.dtype)
if outputs['audios'] is not None:
    outputs['audios'] = outputs['audios'].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, inference = True)
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>
```

```python
messages = [
    {'role': 'user', 'content': '<audio> </audio> apa isu audio ni'},
]
outputs = prepare_dataset(messages, audio = [audio])
if outputs['images'] is not None:
    outputs['images'] = outputs['images'].type(model.dtype)
if outputs['audios'] is not None:
    outputs['audios'] = outputs['audios'].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, inference = True)
    
r = model_inputs.pop('input_ids', None)
generate_kwargs = dict(
    model_inputs,
    max_new_tokens=300,
    top_p=0.95,
    top_k=50,
    temperature=0.9,
    do_sample=True,
    num_beams=1,
)

r = model.llm.generate(**generate_kwargs)
print(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>
```

```python
messages = [
    {'role': 'user', 'content': '<image> </image> <audio> </audio> apa kaitan gambar dan audio ni'},
]
outputs = prepare_dataset(messages, images = [test_image], audio = [audio])
if outputs['images'] is not None:
    outputs['images'] = outputs['images'].type(model.dtype)
if outputs['audios'] is not None:
    outputs['audios'] = outputs['audios'].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, inference = True)
    
r = model_inputs.pop('input_ids', None)
generate_kwargs = dict(
    model_inputs,
    max_new_tokens=300,
    top_p=0.95,
    top_k=50,
    temperature=0.9,
    do_sample=True,
    num_beams=1,
)

r = model.llm.generate(**generate_kwargs)
print(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>
```