File size: 2,841 Bytes
c700a08
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
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
from transformers import PretrainedConfig
from transformers.utils import logging

logger = logging.get_logger(__name__)


class PureDebertaConfig(PretrainedConfig):

    model_type = "pure_deberta"

    def __init__(
        self,
        vocab_size=50265,
        hidden_size=768,
        num_hidden_layers=12,
        num_attention_heads=12,
        intermediate_size=3072,
        hidden_act="gelu",
        hidden_dropout_prob=0.1,
        attention_probs_dropout_prob=0.1,
        max_position_embeddings=512,
        type_vocab_size=0,
        initializer_range=0.02,
        layer_norm_eps=1e-7,
        relative_attention=False,
        max_relative_positions=-1,
        pad_token_id=0,
        position_biased_input=True,
        pos_att_type=None,
        pooler_dropout=0,
        pooler_hidden_act="gelu",
        svd_rank=5, # A slightly overestimated rank of token embedding matrix
        num_pc_to_remove=1, # Number of principal component to remove
        center=False, # If True, centre the input token embedding matrix
        num_iters=2, # Number of subspace iterations to conduct
        alpha=1, # Feature expression factor in parameter-free self-attention module
        disable_pcr=False, 
        disable_pfsa=False, 
        disable_covariance=True, 
        **kwargs,
    ):
        super().__init__(**kwargs)

        self.hidden_size = hidden_size
        self.num_hidden_layers = num_hidden_layers
        self.num_attention_heads = num_attention_heads
        self.intermediate_size = intermediate_size
        self.hidden_act = hidden_act
        self.hidden_dropout_prob = hidden_dropout_prob
        self.attention_probs_dropout_prob = attention_probs_dropout_prob
        self.max_position_embeddings = max_position_embeddings
        self.type_vocab_size = type_vocab_size
        self.initializer_range = initializer_range
        self.relative_attention = relative_attention
        self.max_relative_positions = max_relative_positions
        self.pad_token_id = pad_token_id
        self.position_biased_input = position_biased_input

        # Backwards compatibility
        if isinstance(pos_att_type, str):
            pos_att_type = [x.strip() for x in pos_att_type.lower().split("|")]

        self.pos_att_type = pos_att_type
        self.vocab_size = vocab_size
        self.layer_norm_eps = layer_norm_eps

        self.pooler_hidden_size = kwargs.get("pooler_hidden_size", hidden_size)
        self.pooler_dropout = pooler_dropout
        self.pooler_hidden_act = pooler_hidden_act

        self.svd_rank = svd_rank
        self.num_pc_to_remove = num_pc_to_remove
        self.center = center
        self.num_iters = num_iters
        self.alpha = alpha
        self.disable_pcr = disable_pcr
        self.disable_pfsa = disable_pfsa
        self.disable_covariance = disable_covariance