Spaces:
Running on Zero
Running on Zero
File size: 4,374 Bytes
83e59db | 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 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 | """
Transformer with rotary position embedding, adapted from Phil Wang's repository
at https://github.com/lucidrains/BS-RoFormer (under MIT License).
"""
import torch
import torch.nn.functional as F
from einops import rearrange
from torch import nn
from torch.nn import Module, ModuleList
# helper functions
def exists(val):
return val is not None
# norm
class RMSNorm(Module):
def __init__(self, size, dim=-1):
super().__init__()
self.scale = size**0.5
if dim >= 0:
raise ValueError(f"dim must be negative, got {dim}")
self.gamma = nn.Parameter(torch.ones((size,) + (1,) * (abs(dim) - 1)))
self.dim = dim
def forward(self, x):
return F.normalize(x, dim=self.dim) * self.scale * self.gamma
# feedforward
class FeedForward(Module):
def __init__(
self,
dim,
mult=4,
dropout=0.0,
dim_out=None,
):
super().__init__()
if dim_out is None:
dim_out = dim
dim_inner = int(dim * mult)
self.activation = nn.GELU()
self.net = nn.Sequential(
RMSNorm(dim),
nn.Linear(dim, dim_inner),
self.activation,
nn.Dropout(dropout),
nn.Linear(dim_inner, dim_out),
nn.Dropout(dropout),
)
def forward(self, x):
return self.net(x)
# attention
class Attend(nn.Module):
def __init__(self, dropout=0.0, scale=None):
super().__init__()
self.dropout = dropout
self.scale = scale
def forward(self, q, k, v):
if exists(self.scale):
default_scale = q.shape[-1] ** -0.5
q = q * (self.scale / default_scale)
return F.scaled_dot_product_attention(
q, k, v, dropout_p=self.dropout if self.training else 0.0
)
class Attention(Module):
def __init__(
self,
dim,
heads=8,
dim_head=64,
dropout=0.0,
rotary_embed=None,
gating=True,
):
super().__init__()
self.heads = heads
self.scale = dim_head**-0.5
dim_inner = heads * dim_head
self.rotary_embed = rotary_embed
self.attend = Attend(dropout=dropout)
self.norm = RMSNorm(dim)
self.to_qkv = nn.Linear(dim, dim_inner * 3, bias=False)
if gating:
self.to_gates = nn.Linear(dim, heads)
else:
self.to_gates = None
self.to_out = nn.Sequential(
nn.Linear(dim_inner, dim, bias=False), nn.Dropout(dropout)
)
def forward(self, x):
x = self.norm(x)
q, k, v = rearrange(
self.to_qkv(x), "b n (qkv h d) -> qkv b h n d", qkv=3, h=self.heads
)
if exists(self.rotary_embed):
q = self.rotary_embed.rotate_queries_or_keys(q)
k = self.rotary_embed.rotate_queries_or_keys(k)
out = self.attend(q, k, v)
if exists(self.to_gates):
gates = self.to_gates(x)
out = out * rearrange(gates, "b n h -> b h n 1").sigmoid()
out = rearrange(out, "b h n d -> b n (h d)")
return self.to_out(out)
# Roformer
class Transformer(Module):
def __init__(
self,
*,
dim,
depth,
dim_head=32,
heads=16,
attn_dropout=0.1,
ff_dropout=0.1,
ff_mult=4,
norm_output=True,
rotary_embed=None,
gating=True,
):
super().__init__()
self.layers = ModuleList([])
for _ in range(depth):
ff = FeedForward(dim=dim, mult=ff_mult, dropout=ff_dropout)
self.layers.append(
ModuleList(
[
Attention(
dim=dim,
dim_head=dim_head,
heads=heads,
dropout=attn_dropout,
rotary_embed=rotary_embed,
gating=gating,
),
ff,
]
)
)
self.norm = RMSNorm(dim) if norm_output else nn.Identity()
def forward(self, x):
for attn, ff in self.layers:
x = attn(x) + x
x = ff(x) + x
x = self.norm(x)
return x
|