Update README.md
Browse files
README.md
CHANGED
@@ -1,199 +1,538 @@
|
|
1 |
---
|
2 |
library_name: transformers
|
3 |
-
|
|
|
|
|
|
|
4 |
---
|
5 |
|
6 |
-
#
|
7 |
|
8 |
-
|
9 |
|
10 |
|
|
|
|
|
11 |
|
12 |
-
|
|
|
13 |
|
14 |
-
###
|
|
|
15 |
|
16 |
-
|
17 |
-
|
18 |
-
This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
|
19 |
-
|
20 |
-
- **Developed by:** [More Information Needed]
|
21 |
-
- **Funded by [optional]:** [More Information Needed]
|
22 |
-
- **Shared by [optional]:** [More Information Needed]
|
23 |
-
- **Model type:** [More Information Needed]
|
24 |
-
- **Language(s) (NLP):** [More Information Needed]
|
25 |
-
- **License:** [More Information Needed]
|
26 |
-
- **Finetuned from model [optional]:** [More Information Needed]
|
27 |
-
|
28 |
-
### Model Sources [optional]
|
29 |
-
|
30 |
-
<!-- Provide the basic links for the model. -->
|
31 |
-
|
32 |
-
- **Repository:** [More Information Needed]
|
33 |
-
- **Paper [optional]:** [More Information Needed]
|
34 |
-
- **Demo [optional]:** [More Information Needed]
|
35 |
-
|
36 |
-
## Uses
|
37 |
-
|
38 |
-
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
|
39 |
-
|
40 |
-
### Direct Use
|
41 |
-
|
42 |
-
<!-- This section is for the model use without fine-tuning or plugging into a larger ecosystem/app. -->
|
43 |
-
|
44 |
-
[More Information Needed]
|
45 |
-
|
46 |
-
### Downstream Use [optional]
|
47 |
-
|
48 |
-
<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->
|
49 |
-
|
50 |
-
[More Information Needed]
|
51 |
-
|
52 |
-
### Out-of-Scope Use
|
53 |
-
|
54 |
-
<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
|
55 |
-
|
56 |
-
[More Information Needed]
|
57 |
-
|
58 |
-
## Bias, Risks, and Limitations
|
59 |
-
|
60 |
-
<!-- This section is meant to convey both technical and sociotechnical limitations. -->
|
61 |
-
|
62 |
-
[More Information Needed]
|
63 |
-
|
64 |
-
### Recommendations
|
65 |
-
|
66 |
-
<!-- This section is meant to convey recommendations with respect to the bias, risk, and technical limitations. -->
|
67 |
-
|
68 |
-
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model. More information needed for further recommendations.
|
69 |
-
|
70 |
-
## How to Get Started with the Model
|
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 |
-
- **Training regime:** [More Information Needed] <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
|
96 |
-
|
97 |
-
#### Speeds, Sizes, Times [optional]
|
98 |
-
|
99 |
-
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
|
100 |
-
|
101 |
-
[More Information Needed]
|
102 |
-
|
103 |
-
## Evaluation
|
104 |
-
|
105 |
-
<!-- This section describes the evaluation protocols and provides the results. -->
|
106 |
-
|
107 |
-
### Testing Data, Factors & Metrics
|
108 |
-
|
109 |
-
#### Testing Data
|
110 |
-
|
111 |
-
<!-- This should link to a Dataset Card if possible. -->
|
112 |
-
|
113 |
-
[More Information Needed]
|
114 |
-
|
115 |
-
#### Factors
|
116 |
-
|
117 |
-
<!-- These are the things the evaluation is disaggregating by, e.g., subpopulations or domains. -->
|
118 |
-
|
119 |
-
[More Information Needed]
|
120 |
-
|
121 |
-
#### Metrics
|
122 |
-
|
123 |
-
<!-- These are the evaluation metrics being used, ideally with a description of why. -->
|
124 |
-
|
125 |
-
[More Information Needed]
|
126 |
-
|
127 |
-
### Results
|
128 |
-
|
129 |
-
[More Information Needed]
|
130 |
-
|
131 |
-
#### Summary
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
## Model Examination [optional]
|
136 |
-
|
137 |
-
<!-- Relevant interpretability work for the model goes here -->
|
138 |
|
139 |
-
|
140 |
|
141 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
142 |
|
143 |
-
|
|
|
|
|
|
|
144 |
|
145 |
-
|
|
|
|
|
|
|
|
|
146 |
|
147 |
-
|
148 |
-
- **Hours used:** [More Information Needed]
|
149 |
-
- **Cloud Provider:** [More Information Needed]
|
150 |
-
- **Compute Region:** [More Information Needed]
|
151 |
-
- **Carbon Emitted:** [More Information Needed]
|
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 |
-
[More Information Needed]
|
178 |
|
179 |
-
|
|
|
180 |
|
181 |
-
|
182 |
|
183 |
-
## Glossary [optional]
|
184 |
|
185 |
-
|
|
|
186 |
|
187 |
-
|
188 |
|
189 |
-
## More Information [optional]
|
190 |
|
191 |
-
[
|
|
|
|
|
|
|
|
|
|
|
192 |
|
193 |
-
## Model Card Authors [optional]
|
194 |
|
195 |
-
[
|
|
|
196 |
|
197 |
-
|
|
|
198 |
|
199 |
-
[More Information Needed]
|
|
|
1 |
---
|
2 |
library_name: transformers
|
3 |
+
license: mit
|
4 |
+
language:
|
5 |
+
- ja
|
6 |
+
pipeline_tag: image-text-to-text
|
7 |
---
|
8 |
|
9 |
+
# [EZO model card]
|
10 |
|
11 |
+

|
12 |
|
13 |
|
14 |
+
## [Model Information]
|
15 |
+
Based on the InternVL2-26B, this model has improved performance, especially in image recognition and Japanese language performance, by employing multiple tuning methods. It excels in Japanese language tasks, it's designed to meet diverse needs globally.
|
16 |
|
17 |
+
InternVL2-26Bをベースとして、複数のチューニング手法を採用のうえ、画像認識と日本語性能を中心に、性能を向上させたモデルです。
|
18 |
+
オープンソースモデルでは、2024. 08. 19時点で最高峰の日本語能力を備えており、同様の手法でトレーニングを行うことで、世界中の多様なニーズに応えることができる設計となっています。
|
19 |
|
20 |
+
### [Benchmark Results]
|
21 |
+
TODO:
|
22 |
|
23 |
+
**Terms of Use**:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
+
This model is based on InternVL2-26B and is Mit.
|
26 |
+
このモデルはInternVL2-26Bをベースにしており、Mitとしています。
|
27 |
|
28 |
+
### 推奨される使用ガイドライン / Recommended Usage Guidelines
|
29 |
|
30 |
+
1. **商用利用**: 本モデルを商用目的で使用する場合、info@axcxept.com へのメール連絡を強く推奨します。これにより、モデルの応用や改善についての協力の機会が生まれる可能性があります。
|
31 |
|
32 |
+
2. **クレジット表記**: 本モデルを使用または改変する際は、以下のようなクレジット表記を行うことを推奨します:
|
33 |
+
"This project utilizes HODACHI/EZO-InternVL2-26B, a model based on OpenGVLab/InternVL2-26B and fine-tuned by Axcxept co., ltd."
|
34 |
|
35 |
+
3. **フィードバック**: モデルの使用経験に関するフィードバックを歓迎します。info@axcxept.com までご連絡ください。
|
36 |
|
37 |
+
これらは推奨事項であり、法的要件ではありません。本モデルの使用は主に Mit に準拠します。
|
38 |
|
39 |
+
1. **Commercial Use**: If you intend to use this model for commercial purposes, we strongly encourage you to contact us by email at info@axcxept.com. This may create opportunities for collaboration on applications and improvements of the model. 2.
|
40 |
|
41 |
+
2.**Credit notation**: When using or modifying the model, it is recommended that the following credit be given: “This project utilizes HODAC:
|
42 |
+
���This project utilizes HODACHI/EZO-InternVL2-26B, a model based on OpenGVLab/InternVL2-26B and fine-tuned by Axcxept co. ltd.”
|
43 |
|
44 |
+
3.**Feedback**: We welcome feedback on your experience with the model, please contact us at info@axcxept.com.
|
45 |
|
46 |
+
These are recommendations, not legal requirements. Use of this model is primarily governed by Mit.
|
47 |
|
48 |
|
49 |
+
<!-- Provide a longer summary of what this model is. -->
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
50 |
|
51 |
+
This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.
|
52 |
|
53 |
+
- **Developed by:** Axcxept co., ltd.
|
54 |
+
- **Language(s) (NLP): English, Japanese
|
55 |
+
|
56 |
+
## [Usage]
|
57 |
+
|
58 |
+
### INSTALL
|
59 |
+
|
60 |
+
`
|
61 |
+
pip install -U transformers==4.37.2 sentencepiece
|
62 |
+
`
|
63 |
+
|
64 |
+
|
65 |
+
### Model Loading
|
66 |
+
|
67 |
+
#### 16-bit (bf16 / fp16)
|
68 |
+
|
69 |
+
```python
|
70 |
+
import torch
|
71 |
+
from transformers import AutoTokenizer, AutoModel
|
72 |
+
path = "HODACHI/EZO-InternVL2-26B"
|
73 |
+
model = AutoModel.from_pretrained(
|
74 |
+
path,
|
75 |
+
torch_dtype=torch.bfloat16,
|
76 |
+
low_cpu_mem_usage=True,
|
77 |
+
trust_remote_code=True).eval().cuda()
|
78 |
+
```
|
79 |
+
|
80 |
+
#### BNB 8-bit Quantization
|
81 |
+
|
82 |
+
```python
|
83 |
+
import torch
|
84 |
+
from transformers import AutoTokenizer, AutoModel
|
85 |
+
path = "HODACHI/EZO-InternVL2-26B"
|
86 |
+
model = AutoModel.from_pretrained(
|
87 |
+
path,
|
88 |
+
torch_dtype=torch.bfloat16,
|
89 |
+
load_in_8bit=True,
|
90 |
+
low_cpu_mem_usage=True,
|
91 |
+
trust_remote_code=True).eval()
|
92 |
+
```
|
93 |
+
|
94 |
+
#### BNB 4-bit Quantization
|
95 |
+
|
96 |
+
> **⚠️ Warning:** Due to significant quantization errors with BNB 4-bit quantization on InternViT-6B, the model may produce nonsensical outputs and fail to understand images. Therefore, please avoid using BNB 4-bit quantization.
|
97 |
+
|
98 |
+
#### Multiple GPUs
|
99 |
+
|
100 |
+
The reason for writing the code this way is to avoid errors that occur during multi-GPU inference due to tensors not being on the same device. By ensuring that the first and last layers of the large language model (LLM) are on the same device, we prevent such errors.
|
101 |
+
|
102 |
+
```python
|
103 |
+
import math
|
104 |
+
import torch
|
105 |
+
from transformers import AutoTokenizer, AutoModel
|
106 |
+
|
107 |
+
def split_model(model_name):
|
108 |
+
device_map = {}
|
109 |
+
world_size = torch.cuda.device_count()
|
110 |
+
num_layers = {
|
111 |
+
'InternVL2-1B': 24, 'InternVL2-2B': 24, 'InternVL2-4B': 32, 'InternVL2-8B': 32,
|
112 |
+
'InternVL2-26B': 48, 'InternVL2-40B': 60, 'InternVL2-Llama3-76B': 80}[model_name]
|
113 |
+
# Since the first GPU will be used for ViT, treat it as half a GPU.
|
114 |
+
num_layers_per_gpu = math.ceil(num_layers / (world_size - 0.5))
|
115 |
+
num_layers_per_gpu = [num_layers_per_gpu] * world_size
|
116 |
+
num_layers_per_gpu[0] = math.ceil(num_layers_per_gpu[0] * 0.5)
|
117 |
+
layer_cnt = 0
|
118 |
+
for i, num_layer in enumerate(num_layers_per_gpu):
|
119 |
+
for j in range(num_layer):
|
120 |
+
device_map[f'language_model.model.layers.{layer_cnt}'] = i
|
121 |
+
layer_cnt += 1
|
122 |
+
device_map['vision_model'] = 0
|
123 |
+
device_map['mlp1'] = 0
|
124 |
+
device_map['language_model.model.tok_embeddings'] = 0
|
125 |
+
device_map['language_model.model.embed_tokens'] = 0
|
126 |
+
device_map['language_model.output'] = 0
|
127 |
+
device_map['language_model.model.norm'] = 0
|
128 |
+
device_map['language_model.lm_head'] = 0
|
129 |
+
device_map[f'language_model.model.layers.{num_layers - 1}'] = 0
|
130 |
+
|
131 |
+
return device_map
|
132 |
+
|
133 |
+
path = "HODACHI/EZO-InternVL2-26B"
|
134 |
+
device_map = split_model('InternVL2-26B')
|
135 |
+
model = AutoModel.from_pretrained(
|
136 |
+
path,
|
137 |
+
torch_dtype=torch.bfloat16,
|
138 |
+
low_cpu_mem_usage=True,
|
139 |
+
trust_remote_code=True,
|
140 |
+
device_map=device_map).eval()
|
141 |
+
```
|
142 |
+
|
143 |
+
### Inference with Transformers
|
144 |
+
|
145 |
+
```python
|
146 |
+
import numpy as np
|
147 |
+
import torch
|
148 |
+
import torchvision.transforms as T
|
149 |
+
from decord import VideoReader, cpu
|
150 |
+
from PIL import Image
|
151 |
+
from torchvision.transforms.functional import InterpolationMode
|
152 |
+
from transformers import AutoModel, AutoTokenizer
|
153 |
+
|
154 |
+
IMAGENET_MEAN = (0.485, 0.456, 0.406)
|
155 |
+
IMAGENET_STD = (0.229, 0.224, 0.225)
|
156 |
+
|
157 |
+
def build_transform(input_size):
|
158 |
+
MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
|
159 |
+
transform = T.Compose([
|
160 |
+
T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
|
161 |
+
T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
|
162 |
+
T.ToTensor(),
|
163 |
+
T.Normalize(mean=MEAN, std=STD)
|
164 |
+
])
|
165 |
+
return transform
|
166 |
+
|
167 |
+
def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
|
168 |
+
best_ratio_diff = float('inf')
|
169 |
+
best_ratio = (1, 1)
|
170 |
+
area = width * height
|
171 |
+
for ratio in target_ratios:
|
172 |
+
target_aspect_ratio = ratio[0] / ratio[1]
|
173 |
+
ratio_diff = abs(aspect_ratio - target_aspect_ratio)
|
174 |
+
if ratio_diff < best_ratio_diff:
|
175 |
+
best_ratio_diff = ratio_diff
|
176 |
+
best_ratio = ratio
|
177 |
+
elif ratio_diff == best_ratio_diff:
|
178 |
+
if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
|
179 |
+
best_ratio = ratio
|
180 |
+
return best_ratio
|
181 |
+
|
182 |
+
def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
|
183 |
+
orig_width, orig_height = image.size
|
184 |
+
aspect_ratio = orig_width / orig_height
|
185 |
+
|
186 |
+
# calculate the existing image aspect ratio
|
187 |
+
target_ratios = set(
|
188 |
+
(i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
|
189 |
+
i * j <= max_num and i * j >= min_num)
|
190 |
+
target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])
|
191 |
+
|
192 |
+
# find the closest aspect ratio to the target
|
193 |
+
target_aspect_ratio = find_closest_aspect_ratio(
|
194 |
+
aspect_ratio, target_ratios, orig_width, orig_height, image_size)
|
195 |
+
|
196 |
+
# calculate the target width and height
|
197 |
+
target_width = image_size * target_aspect_ratio[0]
|
198 |
+
target_height = image_size * target_aspect_ratio[1]
|
199 |
+
blocks = target_aspect_ratio[0] * target_aspect_ratio[1]
|
200 |
+
|
201 |
+
# resize the image
|
202 |
+
resized_img = image.resize((target_width, target_height))
|
203 |
+
processed_images = []
|
204 |
+
for i in range(blocks):
|
205 |
+
box = (
|
206 |
+
(i % (target_width // image_size)) * image_size,
|
207 |
+
(i // (target_width // image_size)) * image_size,
|
208 |
+
((i % (target_width // image_size)) + 1) * image_size,
|
209 |
+
((i // (target_width // image_size)) + 1) * image_size
|
210 |
+
)
|
211 |
+
# split the image
|
212 |
+
split_img = resized_img.crop(box)
|
213 |
+
processed_images.append(split_img)
|
214 |
+
assert len(processed_images) == blocks
|
215 |
+
if use_thumbnail and len(processed_images) != 1:
|
216 |
+
thumbnail_img = image.resize((image_size, image_size))
|
217 |
+
processed_images.append(thumbnail_img)
|
218 |
+
return processed_images
|
219 |
+
|
220 |
+
def load_image(image_file, input_size=448, max_num=12):
|
221 |
+
image = Image.open(image_file).convert('RGB')
|
222 |
+
transform = build_transform(input_size=input_size)
|
223 |
+
images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
|
224 |
+
pixel_values = [transform(image) for image in images]
|
225 |
+
pixel_values = torch.stack(pixel_values)
|
226 |
+
return pixel_values
|
227 |
+
|
228 |
+
# If you have an 80G A100 GPU, you can put the entire model on a single GPU.
|
229 |
+
# Otherwise, you need to load a model using multiple GPUs, please refer to the `Multiple GPUs` section.
|
230 |
+
path = 'HODACHI/EZO-InternVL2-26B'
|
231 |
+
model = AutoModel.from_pretrained(
|
232 |
+
path,
|
233 |
+
torch_dtype=torch.bfloat16,
|
234 |
+
low_cpu_mem_usage=True,
|
235 |
+
trust_remote_code=True).eval().cuda()
|
236 |
+
tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
|
237 |
+
|
238 |
+
# set the max number of tiles in `max_num`
|
239 |
+
pixel_values = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
|
240 |
+
generation_config = dict(max_new_tokens=1024, do_sample=False)
|
241 |
+
|
242 |
+
# pure-text conversation (纯文本对话)
|
243 |
+
question = 'Hello, who are you?'
|
244 |
+
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
|
245 |
+
print(f'User: {question}\nAssistant: {response}')
|
246 |
+
|
247 |
+
question = 'Can you tell me a story?'
|
248 |
+
response, history = model.chat(tokenizer, None, question, generation_config, history=history, return_history=True)
|
249 |
+
print(f'User: {question}\nAssistant: {response}')
|
250 |
+
|
251 |
+
# single-image single-round conversation (单图单轮对话)
|
252 |
+
question = '<image>\nPlease describe the image shortly.'
|
253 |
+
response = model.chat(tokenizer, pixel_values, question, generation_config)
|
254 |
+
print(f'User: {question}\nAssistant: {response}')
|
255 |
+
|
256 |
+
# single-image multi-round conversation (单图多轮对话)
|
257 |
+
question = '<image>\nPlease describe the image in detail.'
|
258 |
+
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=None, return_history=True)
|
259 |
+
print(f'User: {question}\nAssistant: {response}')
|
260 |
+
|
261 |
+
question = 'Please write a poem according to the image.'
|
262 |
+
response, history = model.chat(tokenizer, pixel_values, question, generation_config, history=history, return_history=True)
|
263 |
+
print(f'User: {question}\nAssistant: {response}')
|
264 |
+
|
265 |
+
# multi-image multi-round conversation, combined images (多图多轮对话,拼接图像)
|
266 |
+
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
|
267 |
+
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
|
268 |
+
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
|
269 |
+
|
270 |
+
question = '<image>\nDescribe the two images in detail.'
|
271 |
+
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
|
272 |
+
history=None, return_history=True)
|
273 |
+
print(f'User: {question}\nAssistant: {response}')
|
274 |
+
|
275 |
+
question = 'What are the similarities and differences between these two images.'
|
276 |
+
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
|
277 |
+
history=history, return_history=True)
|
278 |
+
print(f'User: {question}\nAssistant: {response}')
|
279 |
+
|
280 |
+
# multi-image multi-round conversation, separate images (多图多轮对话,独立图像)
|
281 |
+
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
|
282 |
+
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
|
283 |
+
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
|
284 |
+
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
|
285 |
+
|
286 |
+
question = 'Image-1: <image>\nImage-2: <image>\nDescribe the two images in detail.'
|
287 |
+
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
|
288 |
+
num_patches_list=num_patches_list,
|
289 |
+
history=None, return_history=True)
|
290 |
+
print(f'User: {question}\nAssistant: {response}')
|
291 |
+
|
292 |
+
question = 'What are the similarities and differences between these two images.'
|
293 |
+
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
|
294 |
+
num_patches_list=num_patches_list,
|
295 |
+
history=history, return_history=True)
|
296 |
+
print(f'User: {question}\nAssistant: {response}')
|
297 |
+
|
298 |
+
# batch inference, single image per sample (单图批处理)
|
299 |
+
pixel_values1 = load_image('./examples/image1.jpg', max_num=12).to(torch.bfloat16).cuda()
|
300 |
+
pixel_values2 = load_image('./examples/image2.jpg', max_num=12).to(torch.bfloat16).cuda()
|
301 |
+
num_patches_list = [pixel_values1.size(0), pixel_values2.size(0)]
|
302 |
+
pixel_values = torch.cat((pixel_values1, pixel_values2), dim=0)
|
303 |
+
|
304 |
+
questions = ['<image>\nDescribe the image in detail.'] * len(num_patches_list)
|
305 |
+
responses = model.batch_chat(tokenizer, pixel_values,
|
306 |
+
num_patches_list=num_patches_list,
|
307 |
+
questions=questions,
|
308 |
+
generation_config=generation_config)
|
309 |
+
for question, response in zip(questions, responses):
|
310 |
+
print(f'User: {question}\nAssistant: {response}')
|
311 |
+
|
312 |
+
# video multi-round conversation (视频多轮对话)
|
313 |
+
def get_index(bound, fps, max_frame, first_idx=0, num_segments=32):
|
314 |
+
if bound:
|
315 |
+
start, end = bound[0], bound[1]
|
316 |
+
else:
|
317 |
+
start, end = -100000, 100000
|
318 |
+
start_idx = max(first_idx, round(start * fps))
|
319 |
+
end_idx = min(round(end * fps), max_frame)
|
320 |
+
seg_size = float(end_idx - start_idx) / num_segments
|
321 |
+
frame_indices = np.array([
|
322 |
+
int(start_idx + (seg_size / 2) + np.round(seg_size * idx))
|
323 |
+
for idx in range(num_segments)
|
324 |
+
])
|
325 |
+
return frame_indices
|
326 |
+
|
327 |
+
def load_video(video_path, bound=None, input_size=448, max_num=1, num_segments=32):
|
328 |
+
vr = VideoReader(video_path, ctx=cpu(0), num_threads=1)
|
329 |
+
max_frame = len(vr) - 1
|
330 |
+
fps = float(vr.get_avg_fps())
|
331 |
+
|
332 |
+
pixel_values_list, num_patches_list = [], []
|
333 |
+
transform = build_transform(input_size=input_size)
|
334 |
+
frame_indices = get_index(bound, fps, max_frame, first_idx=0, num_segments=num_segments)
|
335 |
+
for frame_index in frame_indices:
|
336 |
+
img = Image.fromarray(vr[frame_index].asnumpy()).convert('RGB')
|
337 |
+
img = dynamic_preprocess(img, image_size=input_size, use_thumbnail=True, max_num=max_num)
|
338 |
+
pixel_values = [transform(tile) for tile in img]
|
339 |
+
pixel_values = torch.stack(pixel_values)
|
340 |
+
num_patches_list.append(pixel_values.shape[0])
|
341 |
+
pixel_values_list.append(pixel_values)
|
342 |
+
pixel_values = torch.cat(pixel_values_list)
|
343 |
+
return pixel_values, num_patches_list
|
344 |
+
|
345 |
+
video_path = './examples/red-panda.mp4'
|
346 |
+
pixel_values, num_patches_list = load_video(video_path, num_segments=8, max_num=1)
|
347 |
+
pixel_values = pixel_values.to(torch.bfloat16).cuda()
|
348 |
+
video_prefix = ''.join([f'Frame{i+1}: <image>\n' for i in range(len(num_patches_list))])
|
349 |
+
question = video_prefix + 'What is the red panda doing?'
|
350 |
+
# Frame1: <image>\nFrame2: <image>\n...\nFrame8: <image>\n{question}
|
351 |
+
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
|
352 |
+
num_patches_list=num_patches_list, history=None, return_history=True)
|
353 |
+
print(f'User: {question}\nAssistant: {response}')
|
354 |
+
|
355 |
+
question = 'Describe this video in detail. Don\'t repeat.'
|
356 |
+
response, history = model.chat(tokenizer, pixel_values, question, generation_config,
|
357 |
+
num_patches_list=num_patches_list, history=history, return_history=True)
|
358 |
+
print(f'User: {question}\nAssistant: {response}')
|
359 |
+
```
|
360 |
+
|
361 |
+
#### Streaming output
|
362 |
+
|
363 |
+
Besides this method, you can also use the following code to get streamed output.
|
364 |
+
|
365 |
+
```python
|
366 |
+
from transformers import TextIteratorStreamer
|
367 |
+
from threading import Thread
|
368 |
+
|
369 |
+
# Initialize the streamer
|
370 |
+
streamer = TextIteratorStreamer(tokenizer, skip_prompt=True, skip_special_tokens=True, timeout=10)
|
371 |
+
# Define the generation configuration
|
372 |
+
generation_config = dict(max_new_tokens=1024, do_sample=False, streamer=streamer)
|
373 |
+
# Start the model chat in a separate thread
|
374 |
+
thread = Thread(target=model.chat, kwargs=dict(
|
375 |
+
tokenizer=tokenizer, pixel_values=pixel_values, question=question,
|
376 |
+
history=None, return_history=False, generation_config=generation_config,
|
377 |
+
))
|
378 |
+
thread.start()
|
379 |
+
|
380 |
+
# Initialize an empty string to store the generated text
|
381 |
+
generated_text = ''
|
382 |
+
# Loop through the streamer to get the new text as it is generated
|
383 |
+
for new_text in streamer:
|
384 |
+
if new_text == model.conv_template.sep:
|
385 |
+
break
|
386 |
+
generated_text += new_text
|
387 |
+
print(new_text, end='', flush=True) # Print each new chunk of generated text on the same line
|
388 |
+
```
|
389 |
+
|
390 |
+
## Finetune
|
391 |
+
|
392 |
+
SWIFT from ModelScope community has supported the fine-tuning (Image/Video) of InternVL, please check [this link](https://github.com/modelscope/swift/blob/main/docs/source_en/Multi-Modal/internvl-best-practice.md) for more details.
|
393 |
+
|
394 |
+
## Deployment
|
395 |
+
|
396 |
+
### LMDeploy
|
397 |
+
|
398 |
+
LMDeploy is a toolkit for compressing, deploying, and serving LLM, developed by the MMRazor and MMDeploy teams.
|
399 |
+
|
400 |
+
```sh
|
401 |
+
pip install lmdeploy
|
402 |
+
```
|
403 |
+
|
404 |
+
LMDeploy abstracts the complex inference process of multi-modal Vision-Language Models (VLM) into an easy-to-use pipeline, similar to the Large Language Model (LLM) inference pipeline.
|
405 |
+
|
406 |
+
#### A 'Hello, world' example
|
407 |
+
|
408 |
+
```python
|
409 |
+
from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig
|
410 |
+
from lmdeploy.vl import load_image
|
411 |
+
|
412 |
+
model = 'HODACHI/EZO-InternVL2-26B'
|
413 |
+
system_prompt = 'あなたは優秀な、マルチモーダル大規模言語モデルである。'
|
414 |
+
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
|
415 |
+
chat_template_config = ChatTemplateConfig('internvl-internlm2')
|
416 |
+
chat_template_config.meta_instruction = system_prompt
|
417 |
+
pipe = pipeline(model, chat_template_config=chat_template_config,
|
418 |
+
backend_config=TurbomindEngineConfig(session_len=8192))
|
419 |
+
response = pipe(('describe this image', image))
|
420 |
+
print(response.text)
|
421 |
+
```
|
422 |
+
|
423 |
+
If `ImportError` occurs while executing this case, please install the required dependency packages as prompted.
|
424 |
+
|
425 |
+
#### Multi-images inference
|
426 |
+
|
427 |
+
When dealing with multiple images, you can put them all in one list. Keep in mind that multiple images will lead to a higher number of input tokens, and as a result, the size of the context window typically needs to be increased.
|
428 |
+
|
429 |
+
> Warning: Due to the scarcity of multi-image conversation data, the performance on multi-image tasks may be unstable, and it may require multiple attempts to achieve satisfactory results.
|
430 |
+
|
431 |
+
```python
|
432 |
+
from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig
|
433 |
+
from lmdeploy.vl import load_image
|
434 |
+
from lmdeploy.vl.constants import IMAGE_TOKEN
|
435 |
+
|
436 |
+
model = 'HODACHI/EZO-InternVL2-26B'
|
437 |
+
system_prompt = 'あなたは優秀な、マルチモーダル大規模言語モデルである。'
|
438 |
+
chat_template_config = ChatTemplateConfig('internvl-internlm2')
|
439 |
+
chat_template_config.meta_instruction = system_prompt
|
440 |
+
pipe = pipeline(model, chat_template_config=chat_template_config,
|
441 |
+
backend_config=TurbomindEngineConfig(session_len=8192))
|
442 |
|
443 |
+
image_urls=[
|
444 |
+
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg',
|
445 |
+
'https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg'
|
446 |
+
]
|
447 |
|
448 |
+
images = [load_image(img_url) for img_url in image_urls]
|
449 |
+
# Numbering images improves multi-image conversations
|
450 |
+
response = pipe((f'Image-1: {IMAGE_TOKEN}\nImage-2: {IMAGE_TOKEN}\ndescribe these two images', images))
|
451 |
+
print(response.text)
|
452 |
+
```
|
453 |
|
454 |
+
#### Batch prompts inference
|
|
|
|
|
|
|
|
|
455 |
|
456 |
+
Conducting inference with batch prompts is quite straightforward; just place them within a list structure:
|
457 |
|
458 |
+
```python
|
459 |
+
from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig
|
460 |
+
from lmdeploy.vl import load_image
|
461 |
|
462 |
+
model = 'HODACHI/EZO-InternVL2-26B'
|
463 |
+
system_prompt = 'あなたは優秀な、マルチモーダル大規模言語モデルである。'
|
464 |
+
chat_template_config = ChatTemplateConfig('internvl-internlm2')
|
465 |
+
chat_template_config.meta_instruction = system_prompt
|
466 |
+
pipe = pipeline(model, chat_template_config=chat_template_config,
|
467 |
+
backend_config=TurbomindEngineConfig(session_len=8192))
|
468 |
|
469 |
+
image_urls=[
|
470 |
+
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg",
|
471 |
+
"https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/det.jpg"
|
472 |
+
]
|
473 |
+
prompts = [('describe this image', load_image(img_url)) for img_url in image_urls]
|
474 |
+
response = pipe(prompts)
|
475 |
+
print(response)
|
476 |
+
```
|
477 |
|
478 |
+
#### Multi-turn conversation
|
479 |
|
480 |
+
There are two ways to do the multi-turn conversations with the pipeline. One is to construct messages according to the format of OpenAI and use above introduced method, the other is to use the `pipeline.chat` interface.
|
481 |
|
482 |
+
```python
|
483 |
+
from lmdeploy import pipeline, TurbomindEngineConfig, ChatTemplateConfig, GenerationConfig
|
484 |
+
from lmdeploy.vl import load_image
|
485 |
|
486 |
+
model = 'HODACHI/EZO-InternVL2-26B'
|
487 |
+
system_prompt = 'あなたは優秀な、マルチモーダル大規模言語モデルです'
|
488 |
+
chat_template_config = ChatTemplateConfig('internvl-internlm2')
|
489 |
+
chat_template_config.meta_instruction = system_prompt
|
490 |
+
pipe = pipeline(model, chat_template_config=chat_template_config,
|
491 |
+
backend_config=TurbomindEngineConfig(session_len=8192))
|
492 |
|
493 |
+
image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/demo/resources/human-pose.jpg')
|
494 |
+
gen_config = GenerationConfig(top_k=40, top_p=0.8, temperature=0.8)
|
495 |
+
sess = pipe.chat(('describe this image', image), gen_config=gen_config)
|
496 |
+
print(sess.response.text)
|
497 |
+
sess = pipe.chat('What is the woman doing?', session=sess, gen_config=gen_config)
|
498 |
+
print(sess.response.text)
|
499 |
+
```
|
500 |
|
501 |
+
### [Model Data]
|
502 |
+
#### Training Dataset]
|
503 |
+
We extracted high-quality data from Japanese Wikipedia and FineWeb to create instruction data. Our innovative training approach allows for performance improvements across various languages and domains, making the model suitable for global use despite its focus on Japanese data.
|
504 |
|
505 |
+
日本語のWikiデータおよび、FineWebから良質なデータ、画像分類を抽出し、Instructionデータを作成しました。
|
506 |
+
このモデルでは日本語に特化させていますが、世界中のどんなユースケースでも利用可能なアプローチです。
|
507 |
|
508 |
+
https://huggingface.co/datasets/legacy-datasets/wikipedia
|
509 |
+
https://huggingface.co/datasets/HuggingFaceFW/fineweb
|
510 |
+
https://huggingface.co/datasets/STIC-LVLM/stic-coco-preference-6k/blob/main/images.zip
|
511 |
|
|
|
512 |
|
513 |
+
#### Data Preprocessing
|
514 |
+
We used a plain instruction tuning method to train the model on exemplary responses. This approach enhances the model's ability to understand and generate high-quality responses across various languages and contexts.
|
515 |
|
516 |
+
プレインストラクトチューニング手法を用いて、模範的回答を学習させました。この手法により、モデルは様々な言語やコンテキストにおいて高品質な応答を理解し生成する能力が向上しています。
|
517 |
|
|
|
518 |
|
519 |
+
#### Implementation Information
|
520 |
+
[Pre-Instruction Training]
|
521 |
|
522 |
+
https://huggingface.co/instruction-pretrain/instruction-synthesizer
|
523 |
|
|
|
524 |
|
525 |
+
### [Disclaimer]
|
526 |
+
このモデルは研究開発のみを目的として提供されるものであり、実験的なプロトタイプとみなされるべきモデルです。
|
527 |
+
商業的な使用やミッションクリティカルな環境への配備を意図したものではありません。
|
528 |
+
本モデルの使用は、使用者の責任において行われるものとし、その性能および結果は保証されません。
|
529 |
+
Axcxept株式会社は、直接的、間接的、特別、偶発的、結果的な損害、または本モデルの使用から生じるいかなる損失に対しても、得られた結果にかかわらず、一切の責任を負いません。
|
530 |
+
利用者は、本モデルの使用に伴うリスクを十分に理解し、自己の判断で使用するものとします。
|
531 |
|
|
|
532 |
|
533 |
+
### [Hardware]
|
534 |
+
H100 × 8(Running in 8h)
|
535 |
|
536 |
+
### [We are.]
|
537 |
+
[](https://axcxept.com)
|
538 |
|
|