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 class Residual(nn.Module):
  18     def __init__(self, *f):
  19         super().__init__()
  20         self.f = f[0] if len(f) == 1 else nn.Sequential(*f)
  21
  22     def forward(self, x):
  23         return x + self.f(x)
  24
  25 ##############################
  26
  27 class PositionalEncoding(nn.Module):
  28     def __init__(self, len_max):
  29         super().__init__()
  30         self.len_max = len_max
  31
  32     # From Vaswani et al 2018
  33     # PE_{t,2i}   = sin(t/(L^{2i/D}))
  34     # PE_{t,2i+1} = cos(t/(L^{2i/D}))
  35     def forward(self, x):
  36         t = torch.arange(x.size(1), dtype = x.dtype, device = x.device)[:, None]
  37         j = torch.arange(x.size(2), dtype = x.dtype, device = x.device)[None, :]
  38         k = j%2
  39         return x + torch.sin(t / (self.len_max ** ((j - k) / x.size(2))) + math.pi/2 * k)[None, :, :]
  40
  41 ##############################
  42
  43 class QKVAttention(nn.Module):
  44     def __init__(
  45             self,
  46             dim_in, dim_qk, dim_v,
  47             nb_heads = 1, causal = False, attention_dropout = 0.0
  48     ):
  49         super().__init__()
  50
  51         def randw(*d):
  52             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
  53
  54         self.causal = causal
  55         self.attention_dropout = attention_dropout
  56
  57         self.w_q = randw(nb_heads, dim_qk, dim_in)
  58         self.w_k = randw(nb_heads, dim_qk, dim_in)
  59         self.w_v = randw(nb_heads, dim_v, dim_in)
  60         self.w_o = randw(dim_in, dim_v * nb_heads)
  61
  62     def forward(self, x_q, x_kv = None):
  63         if x_kv is None: x_kv = x_q
  64
  65         q = torch.einsum('ntc,hdc->nhtd', x_q, self.w_q)
  66         k = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_k)
  67         v = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_v)
  68
  69         a = torch.einsum('nhtd,nhsd->nhts', q, k) / math.sqrt(q.size(3))
  70
  71         if self.causal:
  72             mask = torch.arange(a.size(2), device = q.device)[None, None, :, None] \
  73                    < torch.arange(a.size(3), device = q.device)[None, None, None, :]
  74             a = a.masked_fill(mask, float('-inf'))
  75
  76         a = a.softmax(dim = 3)
  77         a = F.dropout(a, self.attention_dropout, self.training)
  78         y = torch.einsum('nhts,nhsd->nthd', a, v).flatten(2)
  79
  80         y = y @ self.w_o
  81
  82         return y
  83
  84 ##############################
  85
  86 class MyGPT(nn.Module):
  87     def __init__(self,
  88                  vocabulary_size,
  89                  dim_model, dim_keys, dim_hidden,
  90                  nb_heads, nb_blocks, dropout = 0.):
  91
  92         super().__init__()
  93
  94         assert dim_model % nb_heads == 0
  95
  96         self.embedding = nn.Sequential(
  97             nn.Embedding(vocabulary_size, dim_model),
  98             nn.Dropout(dropout),
  99             PositionalEncoding(len_max = 1e5),
 100         )
 101
 102         trunk_blocks = [ ]
 103
 104         for _ in range(nb_blocks):
 105             trunk_blocks += [
 106                 Residual(
 107                     nn.LayerNorm(dim_model),
 108                     QKVAttention(
 109                         dim_in = dim_model,
 110                         dim_qk = dim_keys, dim_v = dim_model // nb_heads,
 111                         nb_heads = nb_heads,
 112                         causal = True, attention_dropout = dropout
 113                     ),
 114                 ),
 115                 Residual(
 116                     nn.LayerNorm(dim_model),
 117                     nn.Linear(in_features = dim_model, out_features = dim_hidden),
 118                     nn.ReLU(),
 119                     nn.Linear(in_features = dim_hidden, out_features = dim_model),
 120                     nn.Dropout(dropout),
 121                 ),
 122             ]
 123
 124         self.trunk = nn.Sequential(*trunk_blocks)
 125
 126         self.readout = nn.Linear(in_features = dim_model, out_features = vocabulary_size)
 127
 128     def forward(self, x):
 129         x = torch.cat((x.new_zeros(x.size(0), 1), x), 1)
 130         x = self.embedding(x)
 131         x = self.trunk(x)
 132         x = self.readout(x)
 133         return x[:, :-1]
 134
 135 ######################################################################
 136
 137 if __name__ == '__main__':
 138     print('Basic check.')
 139
 140     vocabulary_size = 10
 141     x = torch.randint(vocabulary_size, (25, 100))
 142
 143     model = MyGPT(
 144         vocabulary_size = vocabulary_size,
 145         dim_model = 16, dim_keys = 50, dim_hidden = 100,
 146         nb_heads = 2, nb_blocks = 3,
 147         dropout = 0.1
 148     )
 149
 150     y = model(x)
 151
 152 ######################################################################