File size: 2,405 Bytes
0a2b9f1 |
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 |
from transformers import BertConfig
class BertConfig(BertConfig):
def __init__(
self,
alibi_starting_size: int = 512,
attention_probs_dropout_prob: float = 0.0,
# mlp
use_glu_mlp: bool = True,
use_monarch_mlp: bool = False,
monarch_mlp_nblocks: int = 4,
# position
use_positional_encodings: bool = False,
max_position_embeddings: int = 512,
# architecture selection
residual_long_conv: bool = False,
# hyena and long conv hyperparameters
bidirectional: bool = True,
hyena_w_mod: int = 1,
hyena_filter_dropout: float = 0.2,
hyena_filter_order: int = 64,
hyena_training_additions: bool = False,
# efficiency
use_flash_mm: bool = False,
# average pooling instead of CLS token
pool_all: bool = False,
**kwargs,
):
"""Configuration class for MosaicBert.
Args:
alibi_starting_size (int): Use `alibi_starting_size` to determine how large of an alibi tensor to
create when initializing the model. You should be able to ignore this parameter in most cases.
Defaults to 512.
attention_probs_dropout_prob (float): By default, turn off attention dropout in Mosaic BERT.
Defaults to 0.0.
"""
super().__init__(
attention_probs_dropout_prob=attention_probs_dropout_prob, **kwargs)
self.alibi_starting_size = alibi_starting_size
# mlp
self.use_glu_mlp = use_glu_mlp
self.use_monarch_mlp = use_monarch_mlp
self.monarch_mlp_nblocks = monarch_mlp_nblocks
# positional encodings
self.use_positional_encodings = use_positional_encodings
self.max_position_embeddings = max_position_embeddings
# architecture
self.residual_long_conv = residual_long_conv
# hyena and long conv hyperparameters
self.bidirectional = bidirectional
self.hyena_w_mod = hyena_w_mod
self.hyena_filter_dropout = hyena_filter_dropout
self.hyena_filter_order = hyena_filter_order
self.hyena_training_additions = hyena_training_additions
# efficiency
self.use_flash_mm = use_flash_mm
# average pooling instead of CLS token
self.pool_all = pool_all
|