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