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 WithResidual(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 AddPositionalEncoding(nn.Module):
  28     def __init__(self, len_max):
  29         super().__init__()
  30         self.len_max = len_max
  31
  32     # [Vaswani et al 2018] PE_{t,2i} = sin(t/(L^{2i/D})), PE_{t,2i+1} = cos(t/(L^{2i/D}))
  33     def forward(self, x):
  34         t = torch.arange(x.size(1), dtype = x.dtype, device = x.device)[:, None]
  35         j = torch.arange(x.size(2), dtype = x.dtype, device = x.device)[None, :]
  36         k = j%2
  37         pe = torch.sin(t / (self.len_max ** ((j - k) / x.size(2))) + math.pi/2 * k)
  38         return x + pe
  39
  40 ##############################
  41
  42 class QKVAttention(nn.Module):
  43     def __init__(self,
  44                  dim_in, dim_qk, dim_v,
  45                  nb_heads = 1, causal = False, attention_dropout = 0.0):
  46         super().__init__()
  47
  48         def randw(*d):
  49             return nn.Parameter(torch.randn(*d) / math.sqrt(d[-1]))
  50
  51         self.causal = causal
  52         self.attention_dropout = attention_dropout
  53
  54         self.w_q = randw(nb_heads, dim_qk, dim_in)
  55         self.w_k = randw(nb_heads, dim_qk, dim_in)
  56         self.w_v = randw(nb_heads, dim_v, dim_in)
  57         self.w_o = randw(dim_v * nb_heads, dim_in)
  58
  59     def forward(self, x_q, x_kv = None):
  60         if x_kv is None: x_kv = x_q
  61
  62         q = torch.einsum('ntc,hdc->nhtd', x_q, self.w_q)
  63         k = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_k)
  64         v = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_v)
  65
  66         a = torch.einsum('nhtd,nhsd->nhts', q, k) / math.sqrt(q.size(3))
  67
  68         if self.causal:
  69             mask = torch.arange(a.size(2), device = q.device)[None, None, :, None] \
  70                    < torch.arange(a.size(3), device = q.device)[None, None, None, :]
  71             a = a.masked_fill(mask, float('-inf'))
  72
  73         a = a.softmax(dim = 3)
  74         a = F.dropout(a, self.attention_dropout, self.training)
  75         y = torch.einsum('nhts,nhsd->nthd', a, v).flatten(2)
  76
  77         y = y @ self.w_o
  78
  79         return y
  80
  81 ##############################
  82
  83 class MyGPT(nn.Module):
  84     def __init__(self,
  85                  vocabulary_size,
  86                  dim_model, dim_keys, dim_hidden,
  87                  nb_heads, nb_blocks,
  88                  dropout = 0.0, len_max = 1e5):
  89
  90         super().__init__()
  91
  92         assert dim_model % nb_heads == 0
  93
  94         self.embedding = nn.Sequential(
  95             nn.Embedding(vocabulary_size, dim_model),
  96             nn.Dropout(dropout),
  97             AddPositionalEncoding(len_max),
  98         )
  99
 100         trunk_blocks = [ ]
 101
 102         for _ in range(nb_blocks):
 103             trunk_blocks += [
 104                 WithResidual(
 105                     nn.LayerNorm((dim_model,)),
 106                     QKVAttention(
 107                         dim_in = dim_model,
 108                         dim_qk = dim_keys,
 109                         dim_v = dim_model // nb_heads,
 110                         nb_heads = nb_heads,
 111                         causal = True, attention_dropout = dropout
 112                     ),
 113                 ),
 114                 WithResidual(
 115                     nn.LayerNorm((dim_model,)),
 116                     nn.Linear(in_features = dim_model, out_features = dim_hidden),
 117                     nn.ReLU(),
 118                     nn.Linear(in_features = dim_hidden, out_features = dim_model),
 119                     nn.Dropout(dropout),
 120                 ),
 121             ]
 122
 123         self.trunk = nn.Sequential(*trunk_blocks)
 124
 125         self.readout = nn.Linear(in_features = dim_model, out_features = vocabulary_size)
 126
 127         with torch.no_grad():
 128             for m in self.modules():
 129                 if isinstance(m, nn.Embedding):
 130                     m.weight.normal_(mean = 0, std = 2e-2)
 131                 elif isinstance(m, nn.LayerNorm):
 132                     m.bias.zero_()
 133                     m.weight.fill_(1.0)
 134
 135     def forward(self, x):
 136         x = F.pad(x, (1, -1))
 137         x = self.embedding(x)
 138         x = self.trunk(x)
 139         x = self.readout(x)
 140         return x
 141
 142 ######################################################################
 143
 144 if __name__ == '__main__':
 145     print('Basic check.')
 146
 147     vocabulary_size = 10
 148     x = torch.randint(vocabulary_size, (25, 100))
 149
 150     model = MyGPT(
 151         vocabulary_size = vocabulary_size,
 152         dim_model = 18, dim_keys = 50, dim_hidden = 100,
 153         nb_heads = 2, nb_blocks = 3,
 154         dropout = 0.1
 155     )
 156
 157     y = model(x)
 158
 159 ######################################################################