mygpt.py

   1 #!/usr/bin/env python
   2
   3 # Any copyright is dedicated to the Public Domain.
   4 # https://creativecommons.org/publicdomain/zero/1.0/
   5
   6 # Written by Francois Fleuret <francois@fleuret.org>
   7
   8 import math
   9
  10 import torch
  11
  12 from torch import nn
  13 from torch.nn import functional as F
  14
  15 ######################################################################
  16
  17 # A BracketedSequence is a BxTx... tensor with a first and a nb time
  18 # steps to compute.
  19
  20 # Modules able to process it expect that they will have to process a
  21 # first bracket starting at t=0, followed by a succession of brackets
  22 # that move forward in time, do not overlap, and cover the axis T with
  23 # no holes.
  24 #
  25 # Although it is more general, for a classical prompt-conditioned
  26 # auto-regressive process it will be a first bracket starting at 0 and
  27 # of arbitrary length for the "prompt", followed by brackets of length
  28 # 1 for the successive tokens.
  29 #
  30 # Modules able to process brackets may implement a cache that is
  31 # resetted when the input bracket starts at t=0
  32
  33
  34 class BracketedSequence:
  35     def __init__(self, x, first=None, nb=None):
  36         self.x = x
  37         self.first = 0 if first is None else first
  38         self.nb = x.size(1) if nb is None else nb
  39
  40     def slice(self):
  41         return self.x[:, self.first : self.first + self.nb]
  42
  43
  44 ######################################################################
  45
  46
  47 class WithResidual(nn.Module):
  48     def __init__(self, *f):
  49         super().__init__()
  50         self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
  51
  52     def forward(self, bs):
  53         bs.x = bs.x + self.f(bs).x
  54         return bs
  55
  56
  57 ######################################################################
  58
  59
  60 class CacheWrapper(nn.Module):
  61     def __init__(self, *f):
  62         super().__init__()
  63         self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
  64
  65     def forward(self, bs):
  66         if bs.first == 0:
  67             y = self.f(bs.slice())
  68             self.cache_y = y.new(*((y.size(0), bs.x.size(1)) + y.size()[2:]))
  69             self.cache_y[:, bs.first : bs.first + bs.nb] = y
  70         else:
  71             self.cache_y[:, bs.first : bs.first + bs.nb] = self.f(bs.slice())
  72
  73         bs.x = self.cache_y
  74
  75         return bs
  76
  77
  78 ##############################
  79
  80
  81 class AddPositionalEncoding(nn.Module):
  82     def __init__(self, len_max):
  83         super().__init__()
  84         self.len_max = len_max
  85
  86     # [Vaswani et al 2018] PE_{t,2i} = sin(t/(L^{2i/D})), PE_{t,2i+1} = cos(t/(L^{2i/D}))
  87
  88     def forward(self, bs, order):  # NxTxD, T
  89         if bs.first == 0:
  90             t = (
  91                 torch.arange(bs.x.size(1) + 1, dtype=bs.x.dtype, device=bs.x.device)[
  92                     :, None
  93                 ]
  94                 - 1
  95             )
  96             j = torch.arange(bs.x.size(2) // 2, dtype=bs.x.dtype, device=bs.x.device)[
  97                 None, :
  98             ]
  99             k = j % 2
 100             pe = (
 101                 torch.sin(
 102                     t / (self.len_max ** ((j - k) / bs.x.size(2))) + math.pi / 2 * k
 103                 )
 104                 .unsqueeze(0)
 105                 .expand(bs.x.size(0), -1, -1)
 106             )
 107
 108             order_output = order + 1
 109             order_input = F.pad(order + 1, (1, -1))
 110
 111             pe_input = pe.gather(
 112                 1, order_input.unsqueeze(-1).expand(-1, -1, pe.size(-1))
 113             )
 114             pe_output = pe.gather(
 115                 1, order_output.unsqueeze(-1).expand(-1, -1, pe.size(-1))
 116             )
 117
 118             self.pe = torch.cat((pe_input, pe_output), 2)
 119             self.cache_y = bs.x.new(bs.x.size())
 120
 121         self.cache_y[:, bs.first : bs.first + bs.nb] = (
 122             bs.slice() + self.pe[:, bs.first : bs.first + bs.nb]
 123         )
 124
 125         bs.x = self.cache_y
 126
 127         return bs
 128
 129
 130 ##############################
 131
 132
 133 class QKVAttention(nn.Module):
 134     def __init__(
 135         self, dim_in, dim_qk, dim_v, nb_heads=1, causal=False, attention_dropout=0.0
 136     ):
 137         super().__init__()
 138
 139         def randw(*d):
 140             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 141
 142         self.causal = causal
 143         self.attention_dropout = attention_dropout
 144
 145         self.w_q = randw(nb_heads, dim_qk, dim_in)
 146         self.w_k = randw(nb_heads, dim_qk, dim_in)
 147         self.w_v = randw(nb_heads, dim_v, dim_in)
 148         self.w_o = randw(dim_v * nb_heads, dim_in)
 149
 150     def forward(self, bs_q):
 151         x_q = bs_q.x
 152
 153         if bs_q.first == 0:
 154             self.cache_k = x_q.new_zeros(
 155                 x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
 156             )
 157             self.cache_v = x_q.new_zeros(
 158                 x_q.size(0), self.w_v.size(0), x_q.size(1), self.w_v.size(1)
 159             )
 160             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 161
 162         q = torch.einsum(
 163             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_q
 164         )
 165         self.cache_k[:, :, bs_q.first : bs_q.first + bs_q.nb] = torch.einsum(
 166             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_k
 167         )
 168         self.cache_v[:, :, bs_q.first : bs_q.first + bs_q.nb] = torch.einsum(
 169             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_v
 170         )
 171
 172         a = torch.einsum(
 173             "nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs_q.first + bs_q.nb]
 174         ) / math.sqrt(self.w_q.size(1))
 175
 176         if self.causal:
 177             if bs_q.first == 0:
 178                 self.cache_attzero = (
 179                     torch.arange(x_q.size(1), device=q.device)[None, None, :, None]
 180                     < torch.arange(x_q.size(1), device=q.device)[None, None, None, :]
 181                 )
 182             a = a.masked_fill(
 183                 self.cache_attzero[
 184                     :, :, bs_q.first : bs_q.first + bs_q.nb, : bs_q.first + bs_q.nb
 185                 ],
 186                 float("-inf"),
 187             )
 188
 189         a = a.softmax(dim=3)
 190         a = F.dropout(a, self.attention_dropout, self.training)
 191
 192         y = torch.einsum(
 193             "nhts,nhsd->nthd", a, self.cache_v[:, :, : bs_q.first + bs_q.nb]
 194         ).flatten(2)
 195
 196         self.cache_y[:, bs_q.first : bs_q.first + bs_q.nb] = y @ self.w_o
 197
 198         bs_q.x = self.cache_y
 199
 200         return bs_q
 201
 202
 203 ##############################
 204
 205
 206 class MyGPT(nn.Module):
 207     def __init__(
 208         self,
 209         vocabulary_size,
 210         dim_model,
 211         dim_keys,
 212         dim_hidden,
 213         nb_heads,
 214         nb_blocks,
 215         causal=False,
 216         dropout=0.0,
 217         len_max=1e5,
 218     ):
 219         super().__init__()
 220
 221         assert dim_model % nb_heads == 0
 222
 223         self.embedding = CacheWrapper(
 224             nn.Embedding(vocabulary_size, dim_model), nn.Dropout(dropout)
 225         )
 226         self.pe = AddPositionalEncoding(len_max)
 227
 228         trunk_blocks = []
 229
 230         for b in range(nb_blocks):
 231             trunk_blocks += [
 232                 WithResidual(
 233                     CacheWrapper(nn.LayerNorm((dim_model,))),
 234                     QKVAttention(
 235                         dim_in=dim_model,
 236                         dim_qk=dim_keys,
 237                         dim_v=dim_model // nb_heads,
 238                         nb_heads=nb_heads,
 239                         causal=causal,
 240                         attention_dropout=dropout,
 241                     ),
 242                 ),
 243                 WithResidual(
 244                     CacheWrapper(
 245                         nn.LayerNorm((dim_model,)),
 246                         nn.Linear(in_features=dim_model, out_features=dim_hidden),
 247                         nn.ReLU(),
 248                         nn.Linear(in_features=dim_hidden, out_features=dim_model),
 249                         nn.Dropout(dropout),
 250                     ),
 251                 ),
 252             ]
 253
 254         self.trunk = nn.Sequential(*trunk_blocks)
 255
 256         self.readout = CacheWrapper(
 257             nn.Linear(in_features=dim_model, out_features=vocabulary_size)
 258         )
 259
 260         with torch.no_grad():
 261             for m in self.modules():
 262                 if isinstance(m, nn.Embedding):
 263                     m.weight.normal_(mean=0, std=2e-2)
 264                 elif isinstance(m, nn.LayerNorm):
 265                     m.bias.zero_()
 266                     m.weight.fill_(1.0)
 267
 268     def forward(self, bs, mode="standard", order=None):
 269         bs = BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb)
 270         if order is None:
 271             order = torch.arange(bs.x.size(1), device=bs.x.device)[None, :].expand_as(
 272                 bs.x
 273             )
 274         bs = self.embedding(bs)
 275         bs = self.pe(bs, order)
 276
 277         if mode == "standard":
 278             bs = self.trunk(bs)
 279             bs = self.readout(bs)
 280         elif mode == "head":
 281             bs = self.trunk(bs)
 282         elif mode == "deep":
 283             r = []
 284             for l in self.trunk:
 285                 bs = l(bs)
 286                 r += [bs.slice()]
 287             bs = BracketedSequence(torch.cat(r, -1))
 288         else:
 289             raise ValueError(f"{mode=}")
 290         return bs
 291
 292
 293 ######################################################################
 294
 295 if __name__ == "__main__":
 296     print("Basic check.")
 297
 298     vocabulary_size = 10
 299     x = torch.randint(vocabulary_size, (9, 7))
 300
 301     model = MyGPT(
 302         vocabulary_size=vocabulary_size,
 303         dim_model=18,
 304         dim_keys=50,
 305         dim_hidden=100,
 306         nb_heads=2,
 307         nb_blocks=1,
 308         dropout=0.1,
 309     )
 310
 311     model.eval()
 312
 313     y1 = model(BracketedSequence(x)).x
 314
 315     y2 = torch.randn_like(y1)
 316     for s in range(x.size(1)):
 317         z = model(BracketedSequence(x, s, 1))
 318         y2[:, s] = z.x[:, s]
 319
 320     # print(y1.max(dim = 2).values)
 321     # print(y2.max(dim = 2).values)
 322     print(f"error={((y1 - y2).norm() / (y1.norm() + y2.norm())).item()}")
 323
 324 ######################################################################