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__(self, 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.empty(*d).normal_(0, 1 / math.sqrt(d[-1])))
  50
  51         self.w_q = randw(nb_heads, dim_qk, dim_in)
  52         self.w_k = randw(nb_heads, dim_qk, dim_in)
  53         self.w_v = randw(nb_heads, dim_v, dim_in)
  54         self.w_o = randw(nb_heads, dim_in, dim_v)
  55         self.causal = causal
  56         self.attention_dropout = attention_dropout
  57
  58     def forward(self, x_q, x_kv = None):
  59         if x_kv is None: x_kv = x_q
  60         q = torch.einsum('ntc,hdc->nhtd', x_q, self.w_q)
  61         k = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_k)
  62         v = torch.einsum('ntc,hdc->nhtd', x_kv, self.w_v)
  63         a = torch.einsum('nhtd,nhsd->nhts', q, k) / math.sqrt(q.size(3))
  64         if self.causal:
  65             mask = torch.arange(a.size(2), device = q.device)[None, None, :, None] \
  66                    < torch.arange(a.size(3), device = q.device)[None, None, None, :]
  67             a = a.masked_fill(mask, float('-inf'))
  68         a = a.softmax(dim = 3)
  69         a = F.dropout(a, self.attention_dropout, self.training)
  70         y = torch.einsum('nhts,nhsd->nhtd', a, v)
  71         y = torch.einsum('nhtd,hcd->ntc', y, self.w_o)
  72
  73         return y
  74
  75 ##############################
  76
  77 class MyGPT(nn.Module):
  78     def __init__(self,
  79                  vocabulary_size,
  80                  dim_model, dim_keys, dim_hidden,
  81                  nb_heads, nb_blocks, dropout = 0.):
  82
  83         super().__init__()
  84
  85         assert dim_model % nb_heads == 0
  86
  87         self.embedding = nn.Sequential(
  88             nn.Embedding(vocabulary_size, dim_model),
  89             nn.Dropout(dropout),
  90             PositionalEncoding(len_max = 1e5),
  91         )
  92
  93         trunk_blocks = [ ]
  94
  95         for _ in range(nb_blocks):
  96             trunk_blocks += [
  97                 Residual(
  98                     nn.LayerNorm(dim_model),
  99                     QKVAttention(
 100                         dim_in = dim_model,
 101                         dim_qk = dim_keys, dim_v = dim_model // nb_heads,
 102                         nb_heads = nb_heads,
 103                         causal = True, attention_dropout = dropout
 104                     ),
 105                     nn.Linear(in_features = dim_model, out_features = dim_model),
 106                 ),
 107                 Residual(
 108                     nn.LayerNorm(dim_model),
 109                     nn.Linear(in_features = dim_model, out_features = dim_hidden),
 110                     nn.ReLU(),
 111                     nn.Linear(in_features = dim_hidden, out_features = dim_model),
 112                     nn.Dropout(dropout),
 113                 ),
 114             ]
 115
 116         self.trunk = nn.Sequential(*trunk_blocks)
 117
 118         self.readout = nn.Linear(in_features = dim_model, out_features = vocabulary_size)
 119
 120     def forward(self, x):
 121         x = self.embedding(x)
 122         x = self.trunk(x)
 123         x = self.readout(x)
 124         return x
 125
 126 ######################################################################
 127
 128 if __name__ == '__main__':
 129     print('Basic check.')
 130
 131     vocabulary_size = 10
 132     x = torch.randint(vocabulary_size, (25, 100))
 133
 134     model = MyGPT(
 135         vocabulary_size = vocabulary_size,
 136         dim_model = 16, dim_keys = 50, dim_hidden = 100,
 137         nb_heads = 2, nb_blocks = 3,
 138         dropout = 0.1
 139     )
 140
 141     y = model(x)
 142
 143 ######################################################################