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=None):
  89         if bs.first == 0:
  90             t = torch.arange(bs.x.size(1), dtype=bs.x.dtype, device=bs.x.device)[
  91                 :, None
  92             ]
  93             j = torch.arange(bs.x.size(2), dtype=bs.x.dtype, device=bs.x.device)[
  94                 None, :
  95             ]
  96             k = j % 2
  97             self.pe = torch.sin(
  98                 t / (self.len_max ** ((j - k) / bs.x.size(2))) + math.pi / 2 * k
  99             )
 100
 101             if order is not None:
 102                 self.pe = self.pe.gather(1, order.unsqueeze(-1).expand_as(self.pe))
 103
 104             self.cache_y = bs.x.new(bs.x.size())
 105
 106         self.cache_y[:, bs.first : bs.first + bs.nb] = (
 107             bs.slice() + self.pe[bs.first : bs.first + bs.nb]
 108         )
 109
 110         bs.x = self.cache_y
 111
 112         return bs
 113
 114
 115 ##############################
 116
 117
 118 class QKVAttention(nn.Module):
 119     def __init__(
 120         self, dim_in, dim_qk, dim_v, nb_heads=1, causal=False, attention_dropout=0.0
 121     ):
 122         super().__init__()
 123
 124         def randw(*d):
 125             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 126
 127         self.causal = causal
 128         self.attention_dropout = attention_dropout
 129
 130         self.w_q = randw(nb_heads, dim_qk, dim_in)
 131         self.w_k = randw(nb_heads, dim_qk, dim_in)
 132         self.w_v = randw(nb_heads, dim_v, dim_in)
 133         self.w_o = randw(dim_v * nb_heads, dim_in)
 134
 135     def forward(self, bs_q):
 136         x_q = bs_q.x
 137
 138         if bs_q.first == 0:
 139             self.cache_k = x_q.new_zeros(
 140                 x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
 141             )
 142             self.cache_v = x_q.new_zeros(
 143                 x_q.size(0), self.w_v.size(0), x_q.size(1), self.w_v.size(1)
 144             )
 145             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 146
 147         q = torch.einsum(
 148             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_q
 149         )
 150         self.cache_k[:, :, bs_q.first : bs_q.first + bs_q.nb] = torch.einsum(
 151             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_k
 152         )
 153         self.cache_v[:, :, bs_q.first : bs_q.first + bs_q.nb] = torch.einsum(
 154             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_v
 155         )
 156
 157         a = torch.einsum(
 158             "nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs_q.first + bs_q.nb]
 159         ) / math.sqrt(self.w_q.size(1))
 160
 161         if self.causal:
 162             if bs_q.first == 0:
 163                 self.cache_attzero = (
 164                     torch.arange(x_q.size(1), device=q.device)[None, None, :, None]
 165                     < torch.arange(x_q.size(1), device=q.device)[None, None, None, :]
 166                 )
 167             a = a.masked_fill(
 168                 self.cache_attzero[
 169                     :, :, bs_q.first : bs_q.first + bs_q.nb, : bs_q.first + bs_q.nb
 170                 ],
 171                 float("-inf"),
 172             )
 173
 174         a = a.softmax(dim=3)
 175         a = F.dropout(a, self.attention_dropout, self.training)
 176
 177         y = torch.einsum(
 178             "nhts,nhsd->nthd", a, self.cache_v[:, :, : bs_q.first + bs_q.nb]
 179         ).flatten(2)
 180
 181         self.cache_y[:, bs_q.first : bs_q.first + bs_q.nb] = y @ self.w_o
 182
 183         bs_q.x = self.cache_y
 184
 185         return bs_q
 186
 187
 188 ##############################
 189
 190
 191 class MyGPT(nn.Module):
 192     def __init__(
 193         self,
 194         vocabulary_size,
 195         dim_model,
 196         dim_keys,
 197         dim_hidden,
 198         nb_heads,
 199         nb_blocks,
 200         causal=False,
 201         dropout=0.0,
 202         len_max=1e5,
 203     ):
 204         super().__init__()
 205
 206         assert dim_model % nb_heads == 0
 207
 208         self.embedding = CacheWrapper(
 209             nn.Embedding(vocabulary_size, dim_model), nn.Dropout(dropout)
 210         )
 211         self.pe = AddPositionalEncoding(len_max)
 212
 213         trunk_blocks = []
 214
 215         for b in range(nb_blocks):
 216             trunk_blocks += [
 217                 WithResidual(
 218                     CacheWrapper(nn.LayerNorm((dim_model,))),
 219                     QKVAttention(
 220                         dim_in=dim_model,
 221                         dim_qk=dim_keys,
 222                         dim_v=dim_model // nb_heads,
 223                         nb_heads=nb_heads,
 224                         causal=causal,
 225                         attention_dropout=dropout,
 226                     ),
 227                 ),
 228                 WithResidual(
 229                     CacheWrapper(
 230                         nn.LayerNorm((dim_model,)),
 231                         nn.Linear(in_features=dim_model, out_features=dim_hidden),
 232                         nn.ReLU(),
 233                         nn.Linear(in_features=dim_hidden, out_features=dim_model),
 234                         nn.Dropout(dropout),
 235                     ),
 236                 ),
 237             ]
 238
 239         self.trunk = nn.Sequential(*trunk_blocks)
 240
 241         self.readout = CacheWrapper(
 242             nn.Linear(in_features=dim_model, out_features=vocabulary_size)
 243         )
 244
 245         with torch.no_grad():
 246             for m in self.modules():
 247                 if isinstance(m, nn.Embedding):
 248                     m.weight.normal_(mean=0, std=2e-2)
 249                 elif isinstance(m, nn.LayerNorm):
 250                     m.bias.zero_()
 251                     m.weight.fill_(1.0)
 252
 253     def forward(self, bs, mode="standard", order=None):
 254         bs = BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb)
 255         if order is not None:
 256             order = F.pad(order + 1, (1, -1))
 257         bs = self.embedding(bs)
 258         bs = self.pe(bs, order)
 259
 260         if mode == "standard":
 261             bs = self.trunk(bs)
 262             bs = self.readout(bs)
 263         elif mode == "head":
 264             bs = self.trunk(bs)
 265         elif mode == "deep":
 266             r = []
 267             for l in self.trunk:
 268                 bs = l(bs)
 269                 r += [bs.slice()]
 270             bs = BracketedSequence(torch.cat(r, -1))
 271         else:
 272             raise ValueError
 273         return bs
 274
 275
 276 ######################################################################
 277
 278 if __name__ == "__main__":
 279     print("Basic check.")
 280
 281     vocabulary_size = 10
 282     x = torch.randint(vocabulary_size, (9, 7))
 283
 284     model = MyGPT(
 285         vocabulary_size=vocabulary_size,
 286         dim_model=18,
 287         dim_keys=50,
 288         dim_hidden=100,
 289         nb_heads=2,
 290         nb_blocks=1,
 291         dropout=0.1,
 292     )
 293
 294     model.eval()
 295
 296     y1 = model(BracketedSequence(x)).x
 297
 298     y2 = torch.randn_like(y1)
 299     for s in range(x.size(1)):
 300         z = model(BracketedSequence(x, s, 1))
 301         y2[:, s] = z.x[:, s]
 302
 303     # print(y1.max(dim = 2).values)
 304     # print(y2.max(dim = 2).values)
 305     print(f"error={((y1 - y2).norm() / (y1.norm() + y2.norm())).item()}")
 306
 307 ######################################################################