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         pe = torch.sin(t / (self.len_max ** ((j - k) / x.size(2))) + math.pi/2 * k)
  40         return x + pe # Let broadcasting to its job
  41
  42 ##############################
  43
  44 class QKVAttention(nn.Module):
  45     def __init__(
  46             self,
  47             dim_in, dim_qk, dim_v,
  48             nb_heads = 1, causal = False, attention_dropout = 0.0
  49     ):
  50         super().__init__()
  51
  52         def randw(*d):
  53             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
  54
  55         self.causal = causal
  56         self.attention_dropout = attention_dropout
  57
  58         self.w_q = randw(nb_heads, dim_qk, dim_in)
  59         self.w_k = randw(nb_heads, dim_qk, dim_in)
  60         self.w_v = randw(nb_heads, dim_v, dim_in)
  61         self.w_o = randw(dim_v * nb_heads, dim_in)
  62
  63     def forward(self, x_q, x_kv = None):
  64         if x_kv is None: x_kv = x_q
  65
  66         q = torch.einsum('ntc,hdc->nhtd', x_q, self.w_q)
  67         k = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_k)
  68         v = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_v)
  69
  70         a = torch.einsum('nhtd,nhsd->nhts', q, k) / math.sqrt(q.size(3))
  71
  72         if self.causal:
  73             mask = torch.arange(a.size(2), device = q.device)[None, None, :, None] \
  74                    < torch.arange(a.size(3), device = q.device)[None, None, None, :]
  75             a = a.masked_fill(mask, float('-inf'))
  76
  77         a = a.softmax(dim = 3)
  78         a = F.dropout(a, self.attention_dropout, self.training)
  79         y = torch.einsum('nhts,nhsd->nthd', a, v).flatten(2)
  80
  81         y = y @ self.w_o
  82
  83         return y
  84
  85 ##############################
  86
  87 class MyGPT(nn.Module):
  88     def __init__(self,
  89                  vocabulary_size,
  90                  dim_model, dim_keys, dim_hidden,
  91                  nb_heads, nb_blocks, dropout = 0.):
  92
  93         super().__init__()
  94
  95         assert dim_model % nb_heads == 0
  96
  97         self.embedding = nn.Sequential(
  98             nn.Embedding(vocabulary_size, dim_model),
  99             nn.Dropout(dropout),
 100             PositionalEncoding(len_max = 1e5),
 101         )
 102
 103         trunk_blocks = [ ]
 104
 105         for _ in range(nb_blocks):
 106             trunk_blocks += [
 107                 Residual(
 108                     nn.LayerNorm(dim_model),
 109                     QKVAttention(
 110                         dim_in = dim_model,
 111                         dim_qk = dim_keys,
 112                         dim_v = dim_model // nb_heads,
 113                         nb_heads = nb_heads,
 114                         causal = True, attention_dropout = dropout
 115                     ),
 116                 ),
 117                 Residual(
 118                     nn.LayerNorm(dim_model),
 119                     nn.Linear(in_features = dim_model, out_features = dim_hidden),
 120                     nn.ReLU(),
 121                     nn.Linear(in_features = dim_hidden, out_features = dim_model),
 122                     nn.Dropout(dropout),
 123                 ),
 124             ]
 125
 126         self.trunk = nn.Sequential(*trunk_blocks)
 127
 128         self.readout = nn.Linear(in_features = dim_model, out_features = vocabulary_size)
 129
 130     def forward(self, x):
 131         x = F.pad(x, (1, 0))
 132         x = self.embedding(x)
 133         x = self.trunk(x)
 134         x = self.readout(x)
 135         return x[:, :-1]
 136
 137 ######################################################################
 138
 139 if __name__ == '__main__':
 140     print('Basic check.')
 141
 142     vocabulary_size = 10
 143     x = torch.randint(vocabulary_size, (25, 100))
 144
 145     model = MyGPT(
 146         vocabulary_size = vocabulary_size,
 147         dim_model = 18, dim_keys = 50, dim_hidden = 100,
 148         nb_heads = 2, nb_blocks = 3,
 149         dropout = 0.1
 150     )
 151
 152     y = model(x)
 153
 154 ######################################################################