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 = torch.cat(
 110                 (order.new_zeros(order.size(0), 1), order[:, :-1] + 1), 1
 111             )
 112
 113             self.pe = torch.cat(
 114                 (
 115                     pe.gather(1, order_input.unsqueeze(-1).expand(-1, -1, pe.size(-1))),
 116                     pe.gather(
 117                         1, order_output.unsqueeze(-1).expand(-1, -1, pe.size(-1))
 118                     ),
 119                 ),
 120                 2,
 121             )
 122
 123             self.cache_y = bs.x.new(bs.x.size())
 124
 125         self.cache_y[:, bs.first : bs.first + bs.nb] = (
 126             bs.slice() + self.pe[:, bs.first : bs.first + bs.nb]
 127         )
 128
 129         bs.x = self.cache_y
 130
 131         return bs
 132
 133
 134 ##############################
 135
 136
 137 class QKVAttention(nn.Module):
 138     def __init__(
 139         self, dim_in, dim_qk, dim_v, nb_heads=1, causal=False, attention_dropout=0.0
 140     ):
 141         super().__init__()
 142
 143         def randw(*d):
 144             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
 145
 146         self.causal = causal
 147         self.attention_dropout = attention_dropout
 148
 149         self.w_q = randw(nb_heads, dim_qk, dim_in)
 150         self.w_k = randw(nb_heads, dim_qk, dim_in)
 151         self.w_v = randw(nb_heads, dim_v, dim_in)
 152         self.w_o = randw(dim_v * nb_heads, dim_in)
 153
 154     def forward(self, bs_q):
 155         x_q = bs_q.x
 156
 157         if bs_q.first == 0:
 158             self.cache_k = x_q.new_zeros(
 159                 x_q.size(0), self.w_k.size(0), x_q.size(1), self.w_k.size(1)
 160             )
 161             self.cache_v = x_q.new_zeros(
 162                 x_q.size(0), self.w_v.size(0), x_q.size(1), self.w_v.size(1)
 163             )
 164             self.cache_y = x_q.new_zeros(x_q.size(0), x_q.size(1), self.w_o.size(1))
 165
 166         q = torch.einsum(
 167             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_q
 168         )
 169         self.cache_k[:, :, bs_q.first : bs_q.first + bs_q.nb] = torch.einsum(
 170             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_k
 171         )
 172         self.cache_v[:, :, bs_q.first : bs_q.first + bs_q.nb] = torch.einsum(
 173             "ntc,hdc->nhtd", x_q[:, bs_q.first : bs_q.first + bs_q.nb], self.w_v
 174         )
 175
 176         a = torch.einsum(
 177             "nhtd,nhsd->nhts", q, self.cache_k[:, :, : bs_q.first + bs_q.nb]
 178         ) / math.sqrt(self.w_q.size(1))
 179
 180         if self.causal:
 181             if bs_q.first == 0:
 182                 self.cache_attzero = (
 183                     torch.arange(x_q.size(1), device=q.device)[None, None, :, None]
 184                     < torch.arange(x_q.size(1), device=q.device)[None, None, None, :]
 185                 )
 186             a = a.masked_fill(
 187                 self.cache_attzero[
 188                     :, :, bs_q.first : bs_q.first + bs_q.nb, : bs_q.first + bs_q.nb
 189                 ],
 190                 float("-inf"),
 191             )
 192
 193         a = a.softmax(dim=3)
 194         a = F.dropout(a, self.attention_dropout, self.training)
 195
 196         y = torch.einsum(
 197             "nhts,nhsd->nthd", a, self.cache_v[:, :, : bs_q.first + bs_q.nb]
 198         ).flatten(2)
 199
 200         self.cache_y[:, bs_q.first : bs_q.first + bs_q.nb] = y @ self.w_o
 201
 202         bs_q.x = self.cache_y
 203
 204         return bs_q
 205
 206
 207 ##############################
 208
 209
 210 class MyGPT(nn.Module):
 211     def __init__(
 212         self,
 213         vocabulary_size,
 214         dim_model,
 215         dim_keys,
 216         dim_hidden,
 217         nb_heads,
 218         nb_blocks,
 219         causal=False,
 220         dropout=0.0,
 221         len_max=1e5,
 222     ):
 223         super().__init__()
 224
 225         assert dim_model % nb_heads == 0
 226
 227         self.embedding = CacheWrapper(
 228             nn.Embedding(vocabulary_size, dim_model), nn.Dropout(dropout)
 229         )
 230         self.pe = AddPositionalEncoding(len_max)
 231
 232         trunk_blocks = []
 233
 234         for b in range(nb_blocks):
 235             trunk_blocks += [
 236                 WithResidual(
 237                     CacheWrapper(nn.LayerNorm((dim_model,))),
 238                     QKVAttention(
 239                         dim_in=dim_model,
 240                         dim_qk=dim_keys,
 241                         dim_v=dim_model // nb_heads,
 242                         nb_heads=nb_heads,
 243                         causal=causal,
 244                         attention_dropout=dropout,
 245                     ),
 246                 ),
 247                 WithResidual(
 248                     CacheWrapper(
 249                         nn.LayerNorm((dim_model,)),
 250                         nn.Linear(in_features=dim_model, out_features=dim_hidden),
 251                         nn.ReLU(),
 252                         nn.Linear(in_features=dim_hidden, out_features=dim_model),
 253                         nn.Dropout(dropout),
 254                     ),
 255                 ),
 256             ]
 257
 258         self.trunk = nn.Sequential(*trunk_blocks)
 259
 260         self.readout = CacheWrapper(
 261             nn.Linear(in_features=dim_model, out_features=vocabulary_size)
 262         )
 263
 264         with torch.no_grad():
 265             for m in self.modules():
 266                 if isinstance(m, nn.Embedding):
 267                     m.weight.normal_(mean=0, std=2e-2)
 268                 elif isinstance(m, nn.LayerNorm):
 269                     m.bias.zero_()
 270                     m.weight.fill_(1.0)
 271
 272     def forward(self, bs, mode="standard", order=None):
 273         bs = BracketedSequence(F.pad(bs.x, (1, -1)), bs.first, bs.nb)
 274         if order is None:
 275             order = torch.arange(bs.x.size(1), device=bs.x.device)[None, :].expand_as(
 276                 bs.x
 277             )
 278         bs = self.embedding(bs)
 279         bs = self.pe(bs, order)
 280
 281         if mode == "standard":
 282             bs = self.trunk(bs)
 283             bs = self.readout(bs)
 284         elif mode == "head":
 285             bs = self.trunk(bs)
 286         elif mode == "deep":
 287             r = []
 288             for l in self.trunk:
 289                 bs = l(bs)
 290                 r += [bs.slice()]
 291             bs = BracketedSequence(torch.cat(r, -1))
 292         else:
 293             raise ValueError(f"{mode=}")
 294         return bs
 295
 296
 297 ######################################################################
 298
 299 if __name__ == "__main__":
 300     print("Basic check.")
 301
 302     vocabulary_size = 10
 303     x = torch.randint(vocabulary_size, (9, 7))
 304
 305     model = MyGPT(
 306         vocabulary_size=vocabulary_size,
 307         dim_model=18,
 308         dim_keys=50,
 309         dim_hidden=100,
 310         nb_heads=2,
 311         nb_blocks=1,
 312         dropout=0.1,
 313     )
 314
 315     model.eval()
 316
 317     y1 = model(BracketedSequence(x)).x
 318
 319     y2 = torch.randn_like(y1)
 320     for s in range(x.size(1)):
 321         z = model(BracketedSequence(x, s, 1))
 322         y2[:, s] = z.x[:, s]
 323
 324     # print(y1.max(dim = 2).values)
 325     # print(y2.max(dim = 2).values)
 326     print(f"error={((y1 - y2).norm() / (y1.norm() + y2.norm())).item()}")
 327
 328 ######################################################################