Add model definition and pytorch model
Browse files- .gitattributes +1 -0
- README.md +3 -4
- model.py +128 -0
- model/config.json +25 -0
- model/pytorch_model.bin +3 -0
.gitattributes
CHANGED
@@ -14,3 +14,4 @@
|
|
14 |
*.pb filter=lfs diff=lfs merge=lfs -text
|
15 |
*.pt filter=lfs diff=lfs merge=lfs -text
|
16 |
*.pth filter=lfs diff=lfs merge=lfs -text
|
|
|
|
14 |
*.pb filter=lfs diff=lfs merge=lfs -text
|
15 |
*.pt filter=lfs diff=lfs merge=lfs -text
|
16 |
*.pth filter=lfs diff=lfs merge=lfs -text
|
17 |
+
model/pytorch_model.bin filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
@@ -21,10 +21,9 @@ news articles in English. It was introduced in
|
|
21 |
|
22 |
## Model description
|
23 |
|
24 |
-
|
25 |
-
|
26 |
-
## Intended users & limitations
|
27 |
-
|
28 |
|
29 |
|
30 |
## How to use
|
|
|
|
21 |
|
22 |
## Model description
|
23 |
|
24 |
+
Please find propaganda definition here:
|
25 |
+
https://propaganda.qcri.org/annotations/definitions.html
|
|
|
|
|
26 |
|
27 |
|
28 |
## How to use
|
29 |
+
|
model.py
ADDED
@@ -0,0 +1,128 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
__author__ = "Yifan Zhang (yzhang@hbku.edu.qa)"
|
2 |
+
__copyright__ = "Copyright (C) 2021, Qatar Computing Research Institute, HBKU, Doha"
|
3 |
+
|
4 |
+
|
5 |
+
from dataclasses import dataclass
|
6 |
+
from typing import Optional, Tuple
|
7 |
+
import torch
|
8 |
+
from torch import nn
|
9 |
+
from torch.nn.functional import sigmoid
|
10 |
+
from transformers import BertPreTrainedModel, BertModel
|
11 |
+
from transformers.file_utils import ModelOutput
|
12 |
+
|
13 |
+
|
14 |
+
TOKEN_TAGS = (
|
15 |
+
"<PAD>", "O",
|
16 |
+
"Name_Calling,Labeling", "Repetition", "Slogans", "Appeal_to_fear-prejudice", "Doubt",
|
17 |
+
"Exaggeration,Minimisation", "Flag-Waving", "Loaded_Language",
|
18 |
+
"Reductio_ad_hitlerum", "Bandwagon",
|
19 |
+
"Causal_Oversimplification", "Obfuscation,Intentional_Vagueness,Confusion", "Appeal_to_Authority", "Black-and-White_Fallacy",
|
20 |
+
"Thought-terminating_Cliches", "Red_Herring", "Straw_Men", "Whataboutism"
|
21 |
+
)
|
22 |
+
|
23 |
+
|
24 |
+
SEQUENCE_TAGS = ("Non-prop", "Prop")
|
25 |
+
|
26 |
+
|
27 |
+
@dataclass
|
28 |
+
class TokenAndSequenceJointClassifierOutput(ModelOutput):
|
29 |
+
loss: Optional[torch.FloatTensor] = None
|
30 |
+
token_logits: torch.FloatTensor = None
|
31 |
+
sequence_logits: torch.FloatTensor = None
|
32 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
33 |
+
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
34 |
+
|
35 |
+
|
36 |
+
class BertForTokenAndSequenceJointClassification(BertPreTrainedModel):
|
37 |
+
|
38 |
+
def __init__(self, config):
|
39 |
+
super().__init__(config)
|
40 |
+
self.num_token_labels = 20
|
41 |
+
self.num_sequence_labels = 2
|
42 |
+
|
43 |
+
self.token_tags = TOKEN_TAGS
|
44 |
+
self.sequence_tags = SEQUENCE_TAGS
|
45 |
+
|
46 |
+
self.alpha = 0.9
|
47 |
+
|
48 |
+
self.bert = BertModel(config)
|
49 |
+
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
50 |
+
self.classifier = nn.ModuleList([
|
51 |
+
nn.Linear(config.hidden_size, self.num_token_labels),
|
52 |
+
nn.Linear(config.hidden_size, self.num_sequence_labels),
|
53 |
+
])
|
54 |
+
self.masking_gate = nn.Linear(2, 1)
|
55 |
+
|
56 |
+
self.init_weights()
|
57 |
+
self.merge_classifier_1 = nn.Linear(self.num_token_labels + self.num_sequence_labels, self.num_token_labels)
|
58 |
+
|
59 |
+
def forward(
|
60 |
+
self,
|
61 |
+
input_ids=None,
|
62 |
+
attention_mask=None,
|
63 |
+
token_type_ids=None,
|
64 |
+
position_ids=None,
|
65 |
+
head_mask=None,
|
66 |
+
inputs_embeds=None,
|
67 |
+
labels=None,
|
68 |
+
output_attentions=None,
|
69 |
+
output_hidden_states=None,
|
70 |
+
return_dict=True,
|
71 |
+
):
|
72 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
73 |
+
|
74 |
+
outputs = self.bert(
|
75 |
+
input_ids,
|
76 |
+
attention_mask=attention_mask,
|
77 |
+
token_type_ids=token_type_ids,
|
78 |
+
position_ids=position_ids,
|
79 |
+
head_mask=head_mask,
|
80 |
+
inputs_embeds=inputs_embeds,
|
81 |
+
output_attentions=output_attentions,
|
82 |
+
output_hidden_states=output_hidden_states,
|
83 |
+
)
|
84 |
+
|
85 |
+
sequence_output = outputs[0]
|
86 |
+
pooler_output = outputs[1]
|
87 |
+
|
88 |
+
sequence_output = self.dropout(sequence_output)
|
89 |
+
token_logits = self.classifier[0](sequence_output)
|
90 |
+
|
91 |
+
pooler_output = self.dropout(pooler_output)
|
92 |
+
sequence_logits = self.classifier[1](pooler_output)
|
93 |
+
|
94 |
+
gate = torch.sigmoid(self.masking_gate(sequence_logits))
|
95 |
+
|
96 |
+
gates = gate.unsqueeze(1).repeat(1, token_logits.size()[1], token_logits.size()[2])
|
97 |
+
|
98 |
+
weighted_token_logits = torch.mul(gates, token_logits)
|
99 |
+
|
100 |
+
logits = [weighted_token_logits, sequence_logits]
|
101 |
+
|
102 |
+
loss = None
|
103 |
+
if labels is not None:
|
104 |
+
criterion = nn.CrossEntropyLoss(ignore_index=0)
|
105 |
+
binary_criterion = nn.BCEWithLogitsLoss(pos_weight=torch.Tensor([3932/14263]).cuda())
|
106 |
+
loss_fct = CrossEntropyLoss()
|
107 |
+
weighted_token_logits = weighted_token_logits.view(-1, weighted_token_logits.shape[-1])
|
108 |
+
sequence_logits = sequence_logits.view(-1, sequence_logits.shape[-1])
|
109 |
+
|
110 |
+
token_loss = criterion(weighted_token_logits, labels)
|
111 |
+
sequence_label = torch.LongTensor([1] if any([label > 0 for label in labels]) else [0])
|
112 |
+
sequence_loss = binary_criterion(sequence_logits, sequence_label)
|
113 |
+
|
114 |
+
loss = self.alpha*loss[0] + (1-self.alpha)*loss[1]
|
115 |
+
|
116 |
+
if not return_dict:
|
117 |
+
output = (logits,) + outputs[2:]
|
118 |
+
return ((loss,) + output) if loss is not None else output
|
119 |
+
|
120 |
+
return TokenAndSequenceJointClassifierOutput(
|
121 |
+
loss=loss,
|
122 |
+
token_logits=weighted_token_logits,
|
123 |
+
sequence_logits=sequence_logits,
|
124 |
+
hidden_states=outputs.hidden_states,
|
125 |
+
attentions=outputs.attentions,
|
126 |
+
)
|
127 |
+
|
128 |
+
|
model/config.json
ADDED
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"_name_or_path": "./model",
|
3 |
+
"architectures": [
|
4 |
+
"BertForTokenAndSequenceJointClassification"
|
5 |
+
],
|
6 |
+
"attention_probs_dropout_prob": 0.1,
|
7 |
+
"gradient_checkpointing": false,
|
8 |
+
"hidden_act": "gelu",
|
9 |
+
"hidden_dropout_prob": 0.1,
|
10 |
+
"hidden_size": 768,
|
11 |
+
"initializer_range": 0.02,
|
12 |
+
"intermediate_size": 3072,
|
13 |
+
"layer_norm_eps": 1e-12,
|
14 |
+
"max_position_embeddings": 512,
|
15 |
+
"model_type": "bert",
|
16 |
+
"num_attention_heads": 12,
|
17 |
+
"num_hidden_layers": 12,
|
18 |
+
"output_hidden_states": true,
|
19 |
+
"pad_token_id": 0,
|
20 |
+
"position_embedding_type": "absolute",
|
21 |
+
"transformers_version": "4.5.1",
|
22 |
+
"type_vocab_size": 2,
|
23 |
+
"use_cache": true,
|
24 |
+
"vocab_size": 28996
|
25 |
+
}
|
model/pytorch_model.bin
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:92c6db6f7ede8bff314c1f6324f9650579012f12b2a2478ab458ff28694b48de
|
3 |
+
size 433399567
|