philipp-zettl
commited on
Commit
•
193de8a
1
Parent(s):
a64809f
Upload folder using huggingface_hub
Browse files- .gitattributes +1 -0
- README.md +209 -0
- assets/confusion_matrix_GGU.png +0 -0
- assets/loss_plot_GGU.png +0 -0
- heads/GGU.pth +3 -0
- multi-head-sequence-classification-model-model.pth +3 -0
- pretrained/backbone/config.json +28 -0
- pretrained/backbone/model.safetensors +3 -0
- pretrained/tokenizer/special_tokens_map.json +51 -0
- pretrained/tokenizer/tokenizer.json +3 -0
- pretrained/tokenizer/tokenizer_config.json +55 -0
- requirements.txt +7 -0
- train.py +1 -0
.gitattributes
CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
36 |
+
pretrained/tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
@@ -0,0 +1,209 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
---
|
3 |
+
license: mit
|
4 |
+
language: multilingual
|
5 |
+
library_name: torch
|
6 |
+
tags: []
|
7 |
+
base_model: BAAI/bge-m3
|
8 |
+
datasets:
|
9 |
+
- philipp-zettl/GGU-xx
|
10 |
+
metrics:
|
11 |
+
|
12 |
+
- accuracy
|
13 |
+
- precision
|
14 |
+
- recall
|
15 |
+
- f1-score
|
16 |
+
model_name: Multi-Head Sequence Classification Model
|
17 |
+
pipeline_tag: text-classification
|
18 |
+
widget:
|
19 |
+
- text: "Hello, how are you?"
|
20 |
+
label: "[GGU] Greeting"
|
21 |
+
- text: "Thank you for your help"
|
22 |
+
label: "[GGU] Gratitude"
|
23 |
+
- text: "Hallo, wie geht es dir?"
|
24 |
+
label: "[GGU] Greeting (de)"
|
25 |
+
- text: "Danke dir."
|
26 |
+
label: "[GGU] Gratitude (de)"
|
27 |
+
- text: "I am not sure what you mean"
|
28 |
+
label: "[GGU] Other"
|
29 |
+
- text: "Generate me an image of a dog!"
|
30 |
+
label: "[GGU] Other"
|
31 |
+
- text: "What is the weather like today?"
|
32 |
+
label: "[GGU] Other"
|
33 |
+
- text: "Wie ist das Wetter heute?"
|
34 |
+
label: "[GGU] Other (de)"
|
35 |
+
---
|
36 |
+
# Multi-Head Sequence Classification Model
|
37 |
+
## Model description
|
38 |
+
The model is a simple sequence classification model based on hidden output layers of a pre-trained transformer model. Multiple heads are added to the output of the backbone to classify the input sequence.
|
39 |
+
|
40 |
+
### Model architecture
|
41 |
+
The model is a simple sequence classification model based on hidden output layers of a pre-trained transformer model.
|
42 |
+
|
43 |
+
The backbone of the model is BAAI/bge-m3 with 1024.
|
44 |
+
|
45 |
+
An additional layer of (GGU: 3) is added to the output of the backbone to classify the input sequence.
|
46 |
+
|
47 |
+
Using the provided implementation (in repository) of `MultiHeadClassificationTrainer`.
|
48 |
+
|
49 |
+
### Use cases
|
50 |
+
Use cases: text classification, sentiment analysis.
|
51 |
+
|
52 |
+
## Model Inference
|
53 |
+
Inference code:
|
54 |
+
|
55 |
+
```python
|
56 |
+
|
57 |
+
from transformers import AutoModel, AutoTokenizer
|
58 |
+
from .model import MultiHeadSequenceClassificationModel
|
59 |
+
import torch
|
60 |
+
|
61 |
+
model = MultiHeadSequenceClassificationModel.from_pretrained('philipp-zettl/multi-head-sequence-classification-model')
|
62 |
+
tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-m3')
|
63 |
+
|
64 |
+
def predict(text):
|
65 |
+
inputs = tokenizer([text], return_tensors="pt", padding=True, truncation=True)
|
66 |
+
outputs = model(**inputs)
|
67 |
+
return outputs
|
68 |
+
|
69 |
+
```
|
70 |
+
|
71 |
+
## Model Training
|
72 |
+
|
73 |
+
#### Confusion Matrix
|
74 |
+
**GGU**
|
75 |
+
![Confusion Matrix GGU](assets/confusion_matrix_GGU.png)
|
76 |
+
|
77 |
+
#### Training Loss
|
78 |
+
**GGU**
|
79 |
+
![Loss GGU](assets/loss_plot_GGU.png)
|
80 |
+
|
81 |
+
|
82 |
+
### Training data
|
83 |
+
The model has been trained on the following datasets:
|
84 |
+
- [philipp-zettl/GGU-xx](https://huggingface.co/datasets/philipp-zettl/GGU-xx)
|
85 |
+
using the implementation provided by MultiHeadClassificationTrainer
|
86 |
+
|
87 |
+
### Training procedure
|
88 |
+
The following code has been executed to train the model:
|
89 |
+
|
90 |
+
```python
|
91 |
+
def train_classifier():
|
92 |
+
backbone = AutoModel.from_pretrained('BAAI/bge-m3').to(torch.float16)
|
93 |
+
tokenizer = AutoTokenizer.from_pretrained('BAAI/bge-m3')
|
94 |
+
device = 'cuda' if torch.cuda.is_available() else 'cpu'
|
95 |
+
|
96 |
+
label_map = {
|
97 |
+
0: 'Greeting',
|
98 |
+
1: 'Gratitude',
|
99 |
+
2: 'Other'
|
100 |
+
}
|
101 |
+
map_label = {
|
102 |
+
label_map[i]: i
|
103 |
+
for i in label_map.keys()
|
104 |
+
}
|
105 |
+
|
106 |
+
num_labels = len(label_map.keys())
|
107 |
+
|
108 |
+
# HParams
|
109 |
+
|
110 |
+
dropout = 0.25
|
111 |
+
learning_rate = 3e-5
|
112 |
+
momentum = 0.9
|
113 |
+
l2_reg = 0.25
|
114 |
+
|
115 |
+
num_epochs = 35
|
116 |
+
l2_loss_weight = 0.25
|
117 |
+
|
118 |
+
model_conf = {
|
119 |
+
'backbone': backbone,
|
120 |
+
'head_config': {
|
121 |
+
'GGU': 3,
|
122 |
+
},
|
123 |
+
'dropout': dropout,
|
124 |
+
'l2_reg': l2_reg,
|
125 |
+
}
|
126 |
+
|
127 |
+
optimizer_conf = {
|
128 |
+
'lr': learning_rate,
|
129 |
+
'momentum': momentum
|
130 |
+
}
|
131 |
+
|
132 |
+
scheduler_conf = {
|
133 |
+
'factor': 0.2,
|
134 |
+
'patience': 3,
|
135 |
+
'min_lr': 1e-8
|
136 |
+
}
|
137 |
+
|
138 |
+
train_run = 1000
|
139 |
+
trainer = MultiHeadClassificationTrainer(
|
140 |
+
model_conf=model_conf,
|
141 |
+
optimizer_conf={**optimizer_conf, 'lr': 1e-4},
|
142 |
+
scheduler_conf=scheduler_conf,
|
143 |
+
num_epochs=1,
|
144 |
+
l2_loss_weight=l2_loss_weight,
|
145 |
+
use_lr_scheduler=True,
|
146 |
+
train_run=train_run
|
147 |
+
)
|
148 |
+
|
149 |
+
new_model, history = trainer.train(dataset_name='philipp-zettl/GGU-xx', target_heads=['GGU'])
|
150 |
+
metrics = history['metrics']
|
151 |
+
history['loss_plot'] = trainer._plot_history(**metrics)
|
152 |
+
return new_model, history, trainer, label_map
|
153 |
+
|
154 |
+
```
|
155 |
+
|
156 |
+
### Evaluation
|
157 |
+
### Evaluation data
|
158 |
+
For model evaluation, a 20% validation split was used from the training data.
|
159 |
+
|
160 |
+
### Evaluation procedure
|
161 |
+
The model was evaluated using the `eval` method provided by the `MultiHeadClassificationTrainer` class:
|
162 |
+
|
163 |
+
```python
|
164 |
+
def _eval_model(self, dataloader, label_map):
|
165 |
+
self.classifier.train(False)
|
166 |
+
eval_heads = list(label_map.keys())
|
167 |
+
y_pred = {h: [] for h in eval_heads}
|
168 |
+
y_test = {h: [] for h in eval_heads}
|
169 |
+
for sample in tqdm(dataloader, total=len(dataloader), desc='Evaluating model...'):
|
170 |
+
labels = {name: sample['label'] for name in eval_heads}
|
171 |
+
embeddings = BatchEncoding({k: torch.stack(v, dim=1).to(self.device) for k, v in sample.items() if k not in ['label', 'sample']})
|
172 |
+
output = self.classifier(embeddings.to('cuda'), head_names=eval_heads)
|
173 |
+
for head in eval_heads:
|
174 |
+
y_pred[head].extend(output[head].argmax(dim=1).cpu())
|
175 |
+
y_test[head].extend(labels[head])
|
176 |
+
torch.cuda.empty_cache()
|
177 |
+
|
178 |
+
accuracies = {h: accuracy_score(y_test[h], y_pred[h]) for h in eval_heads}
|
179 |
+
f1_scores = {h: f1_score(y_test[h], y_pred[h], average="macro") for h in eval_heads}
|
180 |
+
recalls = {h: recall_score(y_test[h], y_pred[h], average='macro') for h in eval_heads}
|
181 |
+
|
182 |
+
report = {}
|
183 |
+
for head in eval_heads:
|
184 |
+
cm = confusion_matrix(y_test[head], y_pred[head], labels=list(label_map[head].keys()))
|
185 |
+
disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=list(label_map[head].values()))
|
186 |
+
clf_report = classification_report(
|
187 |
+
y_test[head], y_pred[head], output_dict=True, target_names=list(label_map[head].values())
|
188 |
+
)
|
189 |
+
del clf_report["accuracy"]
|
190 |
+
clf_report = pd.DataFrame(clf_report).T.reset_index()
|
191 |
+
report[head] = dict(
|
192 |
+
clf_report=clf_report, confusion_matrix=disp, metrics={'accuracy': accuracies[head], 'f1': f1_scores[head], 'recall': recalls[head]}
|
193 |
+
)
|
194 |
+
return report
|
195 |
+
|
196 |
+
```
|
197 |
+
|
198 |
+
### Metrics
|
199 |
+
For evaluation, we used the following metrics: accuracy, precision, recall, f1-score. You can find a detailed classification report here:
|
200 |
+
|
201 |
+
**GGU:**
|
202 |
+
| | index | precision | recall | f1-score | support |
|
203 |
+
|---:|:-------------|------------:|---------:|-----------:|----------:|
|
204 |
+
| 0 | Greeting | 0.725 | 0.935484 | 0.816901 | 31 |
|
205 |
+
| 1 | Gratitude | 0.952381 | 0.740741 | 0.833333 | 27 |
|
206 |
+
| 2 | Other | 0.954545 | 0.893617 | 0.923077 | 47 |
|
207 |
+
| 3 | macro avg | 0.877309 | 0.856614 | 0.857771 | 105 |
|
208 |
+
| 4 | weighted avg | 0.886218 | 0.866667 | 0.868653 | 105 |
|
209 |
+
|
assets/confusion_matrix_GGU.png
ADDED
assets/loss_plot_GGU.png
ADDED
heads/GGU.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:f81781b33246b9e3a70fe2b4954bb548d29884b5902be26965e94e6653312345
|
3 |
+
size 7552
|
multi-head-sequence-classification-model-model.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:bf5bed442997472409a9e1a7115df63c9712e773362c31621f60be199f9935d3
|
3 |
+
size 1135694619
|
pretrained/backbone/config.json
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "BAAI/bge-m3",
|
3 |
+
"architectures": [
|
4 |
+
"XLMRobertaModel"
|
5 |
+
],
|
6 |
+
"attention_probs_dropout_prob": 0.1,
|
7 |
+
"bos_token_id": 0,
|
8 |
+
"classifier_dropout": null,
|
9 |
+
"eos_token_id": 2,
|
10 |
+
"hidden_act": "gelu",
|
11 |
+
"hidden_dropout_prob": 0.1,
|
12 |
+
"hidden_size": 1024,
|
13 |
+
"initializer_range": 0.02,
|
14 |
+
"intermediate_size": 4096,
|
15 |
+
"layer_norm_eps": 1e-05,
|
16 |
+
"max_position_embeddings": 8194,
|
17 |
+
"model_type": "xlm-roberta",
|
18 |
+
"num_attention_heads": 16,
|
19 |
+
"num_hidden_layers": 24,
|
20 |
+
"output_past": true,
|
21 |
+
"pad_token_id": 1,
|
22 |
+
"position_embedding_type": "absolute",
|
23 |
+
"torch_dtype": "float16",
|
24 |
+
"transformers_version": "4.41.2",
|
25 |
+
"type_vocab_size": 1,
|
26 |
+
"use_cache": true,
|
27 |
+
"vocab_size": 250002
|
28 |
+
}
|
pretrained/backbone/model.safetensors
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:68440cc1b73b9af8ab85ecdc138b51877493ffbcec92a0a16e5d7e518eb22908
|
3 |
+
size 1135554344
|
pretrained/tokenizer/special_tokens_map.json
ADDED
@@ -0,0 +1,51 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"bos_token": {
|
3 |
+
"content": "<s>",
|
4 |
+
"lstrip": false,
|
5 |
+
"normalized": false,
|
6 |
+
"rstrip": false,
|
7 |
+
"single_word": false
|
8 |
+
},
|
9 |
+
"cls_token": {
|
10 |
+
"content": "<s>",
|
11 |
+
"lstrip": false,
|
12 |
+
"normalized": false,
|
13 |
+
"rstrip": false,
|
14 |
+
"single_word": false
|
15 |
+
},
|
16 |
+
"eos_token": {
|
17 |
+
"content": "</s>",
|
18 |
+
"lstrip": false,
|
19 |
+
"normalized": false,
|
20 |
+
"rstrip": false,
|
21 |
+
"single_word": false
|
22 |
+
},
|
23 |
+
"mask_token": {
|
24 |
+
"content": "<mask>",
|
25 |
+
"lstrip": true,
|
26 |
+
"normalized": false,
|
27 |
+
"rstrip": false,
|
28 |
+
"single_word": false
|
29 |
+
},
|
30 |
+
"pad_token": {
|
31 |
+
"content": "<pad>",
|
32 |
+
"lstrip": false,
|
33 |
+
"normalized": false,
|
34 |
+
"rstrip": false,
|
35 |
+
"single_word": false
|
36 |
+
},
|
37 |
+
"sep_token": {
|
38 |
+
"content": "</s>",
|
39 |
+
"lstrip": false,
|
40 |
+
"normalized": false,
|
41 |
+
"rstrip": false,
|
42 |
+
"single_word": false
|
43 |
+
},
|
44 |
+
"unk_token": {
|
45 |
+
"content": "<unk>",
|
46 |
+
"lstrip": false,
|
47 |
+
"normalized": false,
|
48 |
+
"rstrip": false,
|
49 |
+
"single_word": false
|
50 |
+
}
|
51 |
+
}
|
pretrained/tokenizer/tokenizer.json
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:c119aa9bc83a5d76efbbc831b23e5790727c12fde474f6519dd96cde6550ffd7
|
3 |
+
size 17083052
|
pretrained/tokenizer/tokenizer_config.json
ADDED
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"added_tokens_decoder": {
|
3 |
+
"0": {
|
4 |
+
"content": "<s>",
|
5 |
+
"lstrip": false,
|
6 |
+
"normalized": false,
|
7 |
+
"rstrip": false,
|
8 |
+
"single_word": false,
|
9 |
+
"special": true
|
10 |
+
},
|
11 |
+
"1": {
|
12 |
+
"content": "<pad>",
|
13 |
+
"lstrip": false,
|
14 |
+
"normalized": false,
|
15 |
+
"rstrip": false,
|
16 |
+
"single_word": false,
|
17 |
+
"special": true
|
18 |
+
},
|
19 |
+
"2": {
|
20 |
+
"content": "</s>",
|
21 |
+
"lstrip": false,
|
22 |
+
"normalized": false,
|
23 |
+
"rstrip": false,
|
24 |
+
"single_word": false,
|
25 |
+
"special": true
|
26 |
+
},
|
27 |
+
"3": {
|
28 |
+
"content": "<unk>",
|
29 |
+
"lstrip": false,
|
30 |
+
"normalized": false,
|
31 |
+
"rstrip": false,
|
32 |
+
"single_word": false,
|
33 |
+
"special": true
|
34 |
+
},
|
35 |
+
"250001": {
|
36 |
+
"content": "<mask>",
|
37 |
+
"lstrip": true,
|
38 |
+
"normalized": false,
|
39 |
+
"rstrip": false,
|
40 |
+
"single_word": false,
|
41 |
+
"special": true
|
42 |
+
}
|
43 |
+
},
|
44 |
+
"bos_token": "<s>",
|
45 |
+
"clean_up_tokenization_spaces": true,
|
46 |
+
"cls_token": "<s>",
|
47 |
+
"eos_token": "</s>",
|
48 |
+
"mask_token": "<mask>",
|
49 |
+
"model_max_length": 128,
|
50 |
+
"pad_token": "<pad>",
|
51 |
+
"sep_token": "</s>",
|
52 |
+
"sp_model_kwargs": {},
|
53 |
+
"tokenizer_class": "XLMRobertaTokenizer",
|
54 |
+
"unk_token": "<unk>"
|
55 |
+
}
|
requirements.txt
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
transformers
|
2 |
+
accelerate
|
3 |
+
datasets
|
4 |
+
pytorch
|
5 |
+
scikit-learn
|
6 |
+
pandas
|
7 |
+
matplotlib
|
train.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
/home/phil/work/mb/easybits/model-zoo/model_zoo/train_classifier.py
|