Package model correctly
#2
by
tcapelle
- opened
- README.md +21 -6
- config.json +18 -3
- configuration_deberta_multi.py +7 -0
- custom_pipeline.py +29 -0
- model.py +0 -20
- modelling_deberta_multi.py +31 -0
- special_tokens_map.json +42 -6
- tokenizer.json +0 -0
- tokenizer_config.json +8 -1
README.md
CHANGED
@@ -33,15 +33,14 @@ For more detailed code regarding generating the annotations in Toxic Commons, tr
|
|
33 |
|
34 |
# How to Use
|
35 |
|
36 |
-
```
|
37 |
-
from transformers import AutoTokenizer
|
38 |
-
from celadon.model import MultiHeadDebertaForSequenceClassification
|
39 |
|
40 |
-
|
41 |
-
|
42 |
model.eval()
|
43 |
|
44 |
-
sample_text = "
|
45 |
|
46 |
inputs = tokenizer(sample_text, return_tensors="pt", padding=True, truncation=True)
|
47 |
outputs = model(input_ids=inputs['input_ids'], attention_mask=inputs['attention_mask'])
|
@@ -53,6 +52,22 @@ predictions = outputs.argmax(dim=-1).squeeze().tolist()
|
|
53 |
print(f"Text: {sample_text}")
|
54 |
for i, category in enumerate(categories):
|
55 |
print(f"Prediction for Category {category}: {predictions[i]}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
56 |
```
|
57 |
|
58 |
# How to Cite
|
|
|
33 |
|
34 |
# How to Use
|
35 |
|
36 |
+
```py
|
37 |
+
from transformers import AutoTokenizer, AutoModelForSequenceClassification
|
|
|
38 |
|
39 |
+
model = AutoModelForSequenceClassification.from_pretrained("PleIAs/celadon", trust_remote_code=True)
|
40 |
+
tokenizer = AutoTokenizer.from_pretrained("PleIAs/celadon", trust_remote_code=True)
|
41 |
model.eval()
|
42 |
|
43 |
+
sample_text = "A very gender inappropriate comment"
|
44 |
|
45 |
inputs = tokenizer(sample_text, return_tensors="pt", padding=True, truncation=True)
|
46 |
outputs = model(input_ids=inputs['input_ids'], attention_mask=inputs['attention_mask'])
|
|
|
52 |
print(f"Text: {sample_text}")
|
53 |
for i, category in enumerate(categories):
|
54 |
print(f"Prediction for Category {category}: {predictions[i]}")
|
55 |
+
# Text: A very gender inappropriate comment
|
56 |
+
# Prediction for Category Race/Origin: 0
|
57 |
+
# Prediction for Category Gender/Sex: 3
|
58 |
+
# Prediction for Category Religion: 0
|
59 |
+
# Prediction for Category Ability: 0
|
60 |
+
# Prediction for Category Violence: 0
|
61 |
+
```
|
62 |
+
|
63 |
+
you can also use transformers pipelines to get a more streamlined experience
|
64 |
+
|
65 |
+
```py
|
66 |
+
from transformers import pipeline
|
67 |
+
pipe = pipeline("text-classification", model="PleIAs/celadon", trust_remote_code=True)
|
68 |
+
result = pipe("This is an example of a normal sentence")
|
69 |
+
print(result)
|
70 |
+
# [{'Race/Origin': 0, 'Gender/Sex': 3, 'Religion': 0, 'Ability': 0, 'Violence': 0}]
|
71 |
```
|
72 |
|
73 |
# How to Cite
|
config.json
CHANGED
@@ -1,8 +1,22 @@
|
|
1 |
{
|
|
|
2 |
"architectures": [
|
3 |
-
"
|
4 |
],
|
5 |
"attention_probs_dropout_prob": 0.1,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
"hidden_act": "gelu",
|
7 |
"hidden_dropout_prob": 0.1,
|
8 |
"hidden_size": 768,
|
@@ -11,9 +25,10 @@
|
|
11 |
"layer_norm_eps": 1e-07,
|
12 |
"max_position_embeddings": 512,
|
13 |
"max_relative_positions": -1,
|
14 |
-
"model_type": "deberta-
|
15 |
"norm_rel_ebd": "layer_norm",
|
16 |
"num_attention_heads": 12,
|
|
|
17 |
"num_hidden_layers": 6,
|
18 |
"pad_token_id": 0,
|
19 |
"pooler_dropout": 0,
|
@@ -28,7 +43,7 @@
|
|
28 |
"relative_attention": true,
|
29 |
"share_att_key": true,
|
30 |
"torch_dtype": "float32",
|
31 |
-
"transformers_version": "4.
|
32 |
"type_vocab_size": 0,
|
33 |
"vocab_size": 128100
|
34 |
}
|
|
|
1 |
{
|
2 |
+
"_name_or_path": "./celadon",
|
3 |
"architectures": [
|
4 |
+
"MultiHeadDebertaForSequenceClassificationModel"
|
5 |
],
|
6 |
"attention_probs_dropout_prob": 0.1,
|
7 |
+
"auto_map": {
|
8 |
+
"AutoConfig": "configuration_deberta_multi.MultiHeadDebertaV2Config",
|
9 |
+
"AutoModelForSequenceClassification": "modelling_deberta_multi.MultiHeadDebertaForSequenceClassificationModel"
|
10 |
+
},
|
11 |
+
"custom_pipelines": {
|
12 |
+
"text-classification": {
|
13 |
+
"impl": "custom_pipeline.CustomTextClassificationPipeline",
|
14 |
+
"pt": [
|
15 |
+
"AutoModelForSequenceClassification"
|
16 |
+
],
|
17 |
+
"tf": []
|
18 |
+
}
|
19 |
+
},
|
20 |
"hidden_act": "gelu",
|
21 |
"hidden_dropout_prob": 0.1,
|
22 |
"hidden_size": 768,
|
|
|
25 |
"layer_norm_eps": 1e-07,
|
26 |
"max_position_embeddings": 512,
|
27 |
"max_relative_positions": -1,
|
28 |
+
"model_type": "multi-head-deberta-for-sequence-classification",
|
29 |
"norm_rel_ebd": "layer_norm",
|
30 |
"num_attention_heads": 12,
|
31 |
+
"num_heads": 5,
|
32 |
"num_hidden_layers": 6,
|
33 |
"pad_token_id": 0,
|
34 |
"pooler_dropout": 0,
|
|
|
43 |
"relative_attention": true,
|
44 |
"share_att_key": true,
|
45 |
"torch_dtype": "float32",
|
46 |
+
"transformers_version": "4.46.2",
|
47 |
"type_vocab_size": 0,
|
48 |
"vocab_size": 128100
|
49 |
}
|
configuration_deberta_multi.py
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import DebertaV2Config
|
2 |
+
|
3 |
+
class MultiHeadDebertaV2Config(DebertaV2Config):
|
4 |
+
model_type = "multi-head-deberta-for-sequence-classification"
|
5 |
+
def __init__(self, num_heads=5, **kwargs):
|
6 |
+
self.num_heads = num_heads
|
7 |
+
super().__init__(**kwargs)
|
custom_pipeline.py
ADDED
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import TextClassificationPipeline, AutoTokenizer
|
2 |
+
|
3 |
+
class CustomTextClassificationPipeline(TextClassificationPipeline):
|
4 |
+
def __init__(self, model, tokenizer=None, **kwargs):
|
5 |
+
# Initialize tokenizer first
|
6 |
+
if tokenizer is None:
|
7 |
+
tokenizer = AutoTokenizer.from_pretrained(model.config._name_or_path)
|
8 |
+
# Make sure we store the tokenizer before calling super().__init__
|
9 |
+
self.tokenizer = tokenizer
|
10 |
+
super().__init__(model=model, tokenizer=tokenizer, **kwargs)
|
11 |
+
|
12 |
+
def _sanitize_parameters(self, **kwargs):
|
13 |
+
preprocess_kwargs = {}
|
14 |
+
return preprocess_kwargs, {}, {}
|
15 |
+
|
16 |
+
def preprocess(self, inputs):
|
17 |
+
return self.tokenizer(inputs, return_tensors='pt', truncation=False)
|
18 |
+
|
19 |
+
def _forward(self, model_inputs):
|
20 |
+
input_ids = model_inputs['input_ids']
|
21 |
+
attention_mask = (input_ids != 0).long()
|
22 |
+
outputs = self.model(input_ids=input_ids, attention_mask=attention_mask)
|
23 |
+
return outputs
|
24 |
+
|
25 |
+
def postprocess(self, model_outputs):
|
26 |
+
predictions = model_outputs.logits.argmax(dim=-1).squeeze().tolist()
|
27 |
+
categories = ["Race/Origin", "Gender/Sex", "Religion", "Ability", "Violence", "Other"]
|
28 |
+
return dict(zip(categories, predictions))
|
29 |
+
|
model.py
DELETED
@@ -1,20 +0,0 @@
|
|
1 |
-
import torch
|
2 |
-
import torch.nn as nn
|
3 |
-
from transformers import DebertaV2Model, DebertaV2PreTrainedModel
|
4 |
-
|
5 |
-
class MultiHeadDebertaForSequenceClassification(DebertaV2PreTrainedModel):
|
6 |
-
def __init__(self, config, num_heads=5):
|
7 |
-
super().__init__(config)
|
8 |
-
self.num_heads = num_heads
|
9 |
-
self.deberta = DebertaV2Model(config)
|
10 |
-
self.heads = nn.ModuleList([nn.Linear(config.hidden_size, 4) for _ in range(num_heads)])
|
11 |
-
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
12 |
-
self.post_init()
|
13 |
-
|
14 |
-
def forward(self, input_ids=None, attention_mask=None):
|
15 |
-
outputs = self.deberta(input_ids=input_ids, attention_mask=attention_mask)
|
16 |
-
sequence_output = outputs[0]
|
17 |
-
logits_list = [head(self.dropout(sequence_output[:, 0, :])) for head in self.heads]
|
18 |
-
logits = torch.stack(logits_list, dim=1)
|
19 |
-
return logits
|
20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
modelling_deberta_multi.py
ADDED
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
from torch import nn, Tensor
|
3 |
+
from typing import Optional
|
4 |
+
from transformers import DebertaV2PreTrainedModel, DebertaV2Model
|
5 |
+
from .configuration_deberta_multi import MultiHeadDebertaV2Config
|
6 |
+
|
7 |
+
class MultiHeadDebertaForSequenceClassificationModel(DebertaV2PreTrainedModel):
|
8 |
+
|
9 |
+
config_class = MultiHeadDebertaV2Config
|
10 |
+
def __init__(self, config): # type: ignore
|
11 |
+
super().__init__(config)
|
12 |
+
self.deberta = DebertaV2Model(config)
|
13 |
+
self.heads = nn.ModuleList(
|
14 |
+
[nn.Linear(config.hidden_size, 4) for _ in range(config.num_heads)]
|
15 |
+
)
|
16 |
+
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
17 |
+
self.post_init()
|
18 |
+
|
19 |
+
def forward(
|
20 |
+
self,
|
21 |
+
input_ids: Optional["Tensor"] = None,
|
22 |
+
attention_mask: Optional["Tensor"] = None,
|
23 |
+
) -> "Tensor":
|
24 |
+
outputs = self.deberta(input_ids=input_ids, attention_mask=attention_mask)
|
25 |
+
sequence_output = outputs[0]
|
26 |
+
logits_list = [
|
27 |
+
head(self.dropout(sequence_output[:, 0, :])) for head in self.heads
|
28 |
+
]
|
29 |
+
logits = torch.stack(logits_list, dim=1)
|
30 |
+
outputs["logits"] = logits
|
31 |
+
return outputs
|
special_tokens_map.json
CHANGED
@@ -1,10 +1,46 @@
|
|
1 |
{
|
2 |
-
"bos_token":
|
3 |
-
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
8 |
"unk_token": {
|
9 |
"content": "[UNK]",
|
10 |
"lstrip": false,
|
|
|
1 |
{
|
2 |
+
"bos_token": {
|
3 |
+
"content": "[CLS]",
|
4 |
+
"lstrip": false,
|
5 |
+
"normalized": false,
|
6 |
+
"rstrip": false,
|
7 |
+
"single_word": false
|
8 |
+
},
|
9 |
+
"cls_token": {
|
10 |
+
"content": "[CLS]",
|
11 |
+
"lstrip": false,
|
12 |
+
"normalized": false,
|
13 |
+
"rstrip": false,
|
14 |
+
"single_word": false
|
15 |
+
},
|
16 |
+
"eos_token": {
|
17 |
+
"content": "[SEP]",
|
18 |
+
"lstrip": false,
|
19 |
+
"normalized": false,
|
20 |
+
"rstrip": false,
|
21 |
+
"single_word": false
|
22 |
+
},
|
23 |
+
"mask_token": {
|
24 |
+
"content": "[MASK]",
|
25 |
+
"lstrip": false,
|
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": "[SEP]",
|
39 |
+
"lstrip": false,
|
40 |
+
"normalized": false,
|
41 |
+
"rstrip": false,
|
42 |
+
"single_word": false
|
43 |
+
},
|
44 |
"unk_token": {
|
45 |
"content": "[UNK]",
|
46 |
"lstrip": false,
|
tokenizer.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
tokenizer_config.json
CHANGED
@@ -47,12 +47,19 @@
|
|
47 |
"do_lower_case": false,
|
48 |
"eos_token": "[SEP]",
|
49 |
"mask_token": "[MASK]",
|
50 |
-
"
|
|
|
|
|
51 |
"pad_token": "[PAD]",
|
|
|
|
|
52 |
"sep_token": "[SEP]",
|
53 |
"sp_model_kwargs": {},
|
54 |
"split_by_punct": false,
|
|
|
55 |
"tokenizer_class": "DebertaV2Tokenizer",
|
|
|
|
|
56 |
"unk_token": "[UNK]",
|
57 |
"vocab_type": "spm"
|
58 |
}
|
|
|
47 |
"do_lower_case": false,
|
48 |
"eos_token": "[SEP]",
|
49 |
"mask_token": "[MASK]",
|
50 |
+
"max_length": 512,
|
51 |
+
"model_max_length": 512,
|
52 |
+
"pad_to_multiple_of": null,
|
53 |
"pad_token": "[PAD]",
|
54 |
+
"pad_token_type_id": 0,
|
55 |
+
"padding_side": "right",
|
56 |
"sep_token": "[SEP]",
|
57 |
"sp_model_kwargs": {},
|
58 |
"split_by_punct": false,
|
59 |
+
"stride": 0,
|
60 |
"tokenizer_class": "DebertaV2Tokenizer",
|
61 |
+
"truncation_side": "right",
|
62 |
+
"truncation_strategy": "longest_first",
|
63 |
"unk_token": "[UNK]",
|
64 |
"vocab_type": "spm"
|
65 |
}
|