Update README.md
Browse files
README.md
CHANGED
@@ -1,201 +1,238 @@
|
|
1 |
---
|
2 |
library_name: transformers
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
---
|
5 |
|
6 |
-
#
|
7 |
|
8 |
-
|
9 |
|
|
|
10 |
|
|
|
11 |
|
12 |
-
##
|
13 |
-
|
14 |
-
### Model Description
|
15 |
-
|
16 |
-
<!-- Provide a longer summary of what this model is. -->
|
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 |
-
|
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 |
-
|
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 |
-
<!-- This section describes the evaluation protocols and provides the results. -->
|
106 |
-
|
107 |
-
### Testing Data, Factors & Metrics
|
108 |
-
|
109 |
-
#### Testing Data
|
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 |
-
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
|
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 |
-
## Technical Specifications
|
154 |
|
155 |
### Model Architecture and Objective
|
156 |
|
157 |
-
[
|
158 |
|
159 |
### Compute Infrastructure
|
160 |
|
161 |
-
|
162 |
|
163 |
#### Hardware
|
164 |
|
165 |
-
|
166 |
|
167 |
#### Software
|
168 |
|
169 |
-
[
|
170 |
-
|
171 |
-
|
172 |
-
|
173 |
-
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
|
174 |
-
|
175 |
-
**BibTeX:**
|
176 |
-
|
177 |
-
[More Information Needed]
|
178 |
-
|
179 |
-
**APA:**
|
180 |
-
|
181 |
-
[More Information Needed]
|
182 |
-
|
183 |
-
## Glossary [optional]
|
184 |
-
|
185 |
-
<!-- If relevant, include terms and calculations in this section that can help readers understand the model or model card. -->
|
186 |
-
|
187 |
-
[More Information Needed]
|
188 |
|
189 |
## More Information [optional]
|
190 |
|
191 |
-
[More Information Needed]
|
192 |
-
|
193 |
-
## Model Card Authors [optional]
|
194 |
-
|
195 |
-
[More Information Needed]
|
196 |
-
|
197 |
-
## Model Card Contact
|
198 |
-
|
199 |
-
[More Information Needed]
|
200 |
-
|
201 |
-
|
|
|
1 |
---
|
2 |
library_name: transformers
|
3 |
+
license: apache-2.0
|
4 |
+
datasets:
|
5 |
+
- isek-ai/danbooru-tags-2023
|
6 |
+
base_model: p1atdev/dart-v1-base
|
7 |
+
tags:
|
8 |
+
- trl
|
9 |
+
- sft
|
10 |
+
- danbooru
|
11 |
---
|
12 |
|
13 |
+
# Dart (Danbooru Tags Transformer) v1
|
14 |
|
15 |
+
This model is a fine-tuned Dart (**Da**nboo**r**u **T**ags Transformer) model that generates danbooru tags.
|
16 |
|
17 |
+
Demo: [🤗 Space](https://huggingface.co/spaces/p1atdev/danbooru-tags-transformer)
|
18 |
|
19 |
+
If you are a developer and want to finetune, it's recommended using the base version, [p1atdev/dart-v1-base](https://huggingface.co/p1atdev/dart-v1-base), instead
|
20 |
|
21 |
+
## Usage
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
|
23 |
+
### Using AutoModel
|
|
|
|
|
|
|
|
|
|
|
|
|
24 |
|
25 |
+
🤗 Transformers library is required.
|
26 |
|
27 |
+
```bash
|
28 |
+
pip install -U transformers
|
29 |
+
```
|
30 |
|
31 |
+
```py
|
32 |
+
import torch
|
33 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM, GenerationConfig
|
34 |
|
35 |
+
MODEL_NAME = "p1atdev/dart-v1-sft"
|
36 |
|
37 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True) # trust_remote_code is required for tokenizer
|
38 |
+
model = AutoModelForCausalLM.from_pretrained(MODEL_NAME, torch_dtype=torch.bfloat16)
|
39 |
|
40 |
+
prompt = "<|bos|><rating>rating:sfw, rating:general</rating><copyright>original</copyright><character></character><general>1girl, "
|
41 |
+
inputs = tokenizer(prompt, return_tensors="pt").input_ids
|
42 |
|
43 |
+
with torch.no_grad():
|
44 |
+
outputs = model.generate(inputs, generation_config=generation_config)
|
45 |
|
46 |
+
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
47 |
+
# rating:sfw, rating:general, original, 1girl, ahoge, black hair, blue eyes, blush, closed mouth, ear piercing, earrings, jewelry, looking at viewer, mole, mole under eye, piercing, portrait, shirt, short hair, solo, white shirt
|
48 |
+
```
|
49 |
|
50 |
+
#### Flash attention (optional)
|
51 |
|
52 |
+
Using flash attention can optimize computations, but it is currently only compatible with Linux.
|
53 |
|
54 |
+
```bash
|
55 |
+
pip install flash_attn
|
56 |
+
```
|
57 |
|
58 |
+
### Accelerate with ORTModel
|
59 |
|
60 |
+
🤗 Optimum library is also compatible, for the high performance inference using ONNX.
|
61 |
|
62 |
+
```bash
|
63 |
+
pip install "optimum[onnxruntime]"
|
64 |
+
```
|
65 |
|
66 |
+
Two ONNX models are provided:
|
67 |
|
68 |
+
- [Normal](./model.onnx)
|
69 |
+
- [Quantized](./model_quantized.onnx)
|
70 |
|
71 |
+
Both can be utilized based on the following code:
|
72 |
|
73 |
+
```py
|
74 |
+
import torch
|
75 |
+
from transformers import AutoTokenizer, GenerationConfig
|
76 |
+
from optimum.onnxruntime import ORTModelForCausalLM
|
77 |
|
78 |
+
MODEL_NAME = "p1atdev/dart-v1-sft"
|
79 |
|
80 |
+
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
|
81 |
|
82 |
+
# normal version
|
83 |
+
ort_model = ORTModelForCausalLM.from_pretrained(MODEL_NAME)
|
84 |
|
85 |
+
# qunatized version
|
86 |
+
# ort_model = ORTModelForCausalLM.from_pretrained(MODEL_NAME, file_name="model_quantized.onnx")
|
87 |
|
88 |
+
prompt = "<|bos|><rating>rating:sfw, rating:general</rating><copyright>original</copyright><character></character><general>1girl, "
|
89 |
+
inputs = tokenizer(prompt, return_tensors="pt").input_ids
|
90 |
|
91 |
+
with torch.no_grad():
|
92 |
+
outputs = model.generate(inputs, generation_config=generation_config)
|
93 |
|
94 |
+
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
|
95 |
+
```
|
96 |
|
97 |
+
### Prompt guidde
|
98 |
|
99 |
+
Due to training with a specialized prompt format, **natural language is not supported**.
|
100 |
|
101 |
+
The trained sentences are essentially composed of the following elements, arranged in the strict order shown below:
|
102 |
|
103 |
+
- `<|bos|>`: The bos (begin of sentence) token
|
104 |
+
- `<rating>[RATING_PARENT], [RATING_CHILD]</rating>`: The block of rating tags
|
105 |
+
- [RATING_PARENT]: `rating:sfw`, `rating:nsfw`
|
106 |
+
- [RATING_CHILD]:
|
107 |
+
- if `[RATING_PARENT]` is `rating:sfw`: `rating:general`, `rating:sensitive`
|
108 |
+
- else: `rating:questionable`, `rating:explicit`
|
109 |
+
- `<copyright>[COPYRIGHT, ...]</copyright>`: The block of copyright tags.
|
110 |
+
- [COPYRIGHT, ...]: All supported copyright tags can be seen in [TODO]()
|
111 |
+
- `<character>[CHARACTER, ...]</character>`: The block of character tags.
|
112 |
+
- [CHARACTER, ...]: All supported character tags can be seen in [TODO]()
|
113 |
+
- `<general>[LENGTH_TOKEN][GENERAL, ...]<|input_end|>[COMPLETION]</general>`: The block of general tags.
|
114 |
+
- [LENGTH_TOKEN]: A token to specify **total** amount of general tags.
|
115 |
+
- Avaiable:
|
116 |
+
- `<|very_short|>`: less than 10 tags
|
117 |
+
- `<|short|>`: less than 20 tags
|
118 |
+
- `<|long|>`: less than 40 tags (recommended)
|
119 |
+
- `<|very_long|>`: more than 40 tags
|
120 |
+
- [GENERAL, ...]: All supported general tags can be seen in [TODO]()
|
121 |
+
- `<|input_end|>`: A tag to show the end of input. Set this token at last of prompt.
|
122 |
+
- [COMPLETION]: The model complete tags in alphabetical order.
|
123 |
+
- `<|eos|>`: The eos (end of sentence) token
|
124 |
|
125 |
+
- Tags other than special tokens are separated by commas.
|
126 |
+
- You can place tags in any order you like in each block.
|
127 |
|
128 |
+
Example sentence:
|
129 |
|
130 |
+
```
|
131 |
+
<|bos|><rating>rating:sfw, rating:general</rating><copyright>vocaloid</copyright><character>hatsune miku</character><general><|long|>solo, 1girl, very long hair<|input_end|>blue hair, cowboy shot, ...</general><|eos|>
|
132 |
+
```
|
133 |
|
134 |
+
Therefore, to complete the tags, the input prompt should be as follows:
|
135 |
|
136 |
+
1. without any copyright and character tags
|
137 |
|
138 |
+
```
|
139 |
+
<|bos|><rating>rating:sfw, rating:general</rating><copyright></copyright><character></character><general><|very_long|>1girl, solo, cat ears<|input_end|>
|
140 |
+
```
|
141 |
|
142 |
+
2. specifing copyright and character tags
|
143 |
|
144 |
+
```
|
145 |
+
<|bos|><rating>rating:sfw, rating:general</rating><copyright>sousou no frieren</copyright><character>frieren</character><general><|long|>1girl, solo, from side<|input_end|>
|
146 |
+
```
|
147 |
|
148 |
+
## Model Details
|
|
|
|
|
|
|
|
|
|
|
|
|
149 |
|
150 |
+
### Model Description
|
151 |
|
152 |
+
- **Developed by:** Plat
|
153 |
+
- **Model type:** Causal language model
|
154 |
+
- **Language(s) (NLP):** Danbooru tags
|
155 |
+
- **License:** Apache-2.0
|
156 |
|
157 |
+
- **Demo:** Avaiable on [🤗Space](https://huggingface.co/spaces/p1atdev/danbooru-tags-transformer)
|
158 |
|
159 |
+
## Bias, Risks, and Limitations
|
160 |
|
161 |
+
Since this model is a pre-trained model, it cannot accommodate flexible specifications.
|
162 |
|
163 |
+
## Training Details
|
164 |
|
165 |
+
### Training Data
|
166 |
|
167 |
+
This model was trained with:
|
168 |
|
169 |
+
- [isek-ai/danbooru-tags-2023](https://huggingface.co/datasets/isek-ai/danbooru-tags-2023): 6M size of danbooru tags dataset since 2005 to 2023
|
170 |
|
171 |
+
Only data from 2020 onwards was used for SFT.
|
172 |
|
173 |
+
### Training Procedure
|
174 |
|
175 |
+
Trained using 🤗 transformers' trainer.
|
176 |
|
177 |
+
#### Preprocessing
|
178 |
|
179 |
+
Preprocessing was conducted through the following process:
|
180 |
|
181 |
+
1. Remove data where `general` tags is null.
|
182 |
+
2. Remove `general` tags that appear less than 100 times.
|
183 |
+
3. Remove undesirable tags such as `watermark` and `bad anatomy`.
|
184 |
+
4. Remove based on the number of tags attached to a single post (following rules):
|
185 |
+
- Remove if more than 100 for `general` tags.
|
186 |
+
- Remove if more than 5 for `copyright` tags.
|
187 |
+
- Remove if more than 10 for `character` tags.
|
188 |
+
5. Remove posts created before 2020
|
189 |
+
6. Set length token according to each tags length
|
190 |
+
7. Shuffle some tags in the following rule:
|
191 |
+
- Include people tags (e.g. `1girl`, `no humans`) tags in the shuffle-group with a 95% probability, and do not do so with a 5% probability.
|
192 |
+
- Get tags at a random percentage between 0% and 75% to create a shuffle-group.
|
193 |
+
- Shuffle tags in shuffle-group and concatnate with `<|input_end|>` token and remains in alphabetical order.
|
194 |
+
8. Concatnate all categories
|
195 |
|
196 |
+
#### Training Hyperparameters
|
197 |
|
198 |
+
The following hyperparameters were used during training:
|
199 |
+
- learning_rate: 0.0001
|
200 |
+
- train_batch_size: 32
|
201 |
+
- eval_batch_size: 8
|
202 |
+
- seed: 42
|
203 |
+
- gradient_accumulation_steps: 2
|
204 |
+
- total_train_batch_size: 64
|
205 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
206 |
+
- lr_scheduler_type: linear
|
207 |
+
- lr_scheduler_warmup_steps: 500
|
208 |
+
- num_epochs: 1
|
209 |
|
|
|
210 |
|
211 |
+
## Evaluation
|
212 |
|
213 |
+
Evaluation has not been done yet and it needs to evaluate.
|
|
|
|
|
|
|
|
|
214 |
|
215 |
+
## Technical Specifications
|
216 |
|
217 |
### Model Architecture and Objective
|
218 |
|
219 |
+
The architecture of this model is [OPT (Open Pretrained Transformer)](https://huggingface.co/docs/transformers/model_doc/opt), but the position embeddings was not trained.
|
220 |
|
221 |
### Compute Infrastructure
|
222 |
|
223 |
+
In house
|
224 |
|
225 |
#### Hardware
|
226 |
|
227 |
+
1x RTX 3070 Ti
|
228 |
|
229 |
#### Software
|
230 |
|
231 |
+
- Dataset processing: [🤗 Datasets](https://github.com/huggingface/datasets)
|
232 |
+
- Training: [🤗 Transformers](https://github.com/huggingface/transformers)
|
233 |
+
- Optimizing: [🤗 Optimum](https://github.com/huggingface/optimum)
|
234 |
+
- SFT: [🤗 TRL](https://github.com/huggingface/trl)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
235 |
|
236 |
## More Information [optional]
|
237 |
|
238 |
+
[More Information Needed]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|