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, nb_heads = 1, causal = False, attention_dropout = 0.0):
  45         super().__init__()
  46
  47         def randw(*d):
  48             return nn.Parameter(torch.empty(*d).normal_(0, 1 / math.sqrt(d[-1])))
  49
  50         self.w_q = randw(nb_heads, dim_qk, dim_in)
  51         self.w_k = randw(nb_heads, dim_qk, dim_in)
  52         self.w_v = randw(nb_heads, dim_v, dim_in)
  53         self.causal = causal
  54         self.attention_dropout = attention_dropout
  55
  56     def forward(self, x):
  57         q = torch.einsum('ntc,hdc->nhtd', x, self.w_q)
  58         k = torch.einsum('ntc,hdc->nhtd', x, self.w_k)
  59         v = torch.einsum('ntc,hdc->nhtd', x, self.w_v)
  60         r = math.sqrt(q.size(3))
  61         a = torch.einsum('nhtd,nhsd->nhts', q, k).div(r)
  62         if self.causal:
  63             mask = torch.tril(q.new_ones(a.size(2), a.size(3)))[None, None, :, :] == 0
  64             a = a.masked_fill(mask, float('-inf'))
  65         a = a.softmax(dim = 3)
  66         a = F.dropout(a, self.attention_dropout, self.training)
  67         y = torch.einsum('nhts,nhsd->nhtd', a, v)
  68         return y.permute(0, 2, 1, 3).flatten(2) # nhtd -> nt(hd)
  69
  70 ##############################
  71
  72 class MyGPT(nn.Module):
  73     def __init__(self,
  74                  vocabulary_size,
  75                  dim_model, dim_keys, dim_hidden,
  76                  nb_heads, nb_blocks, dropout = 0.):
  77
  78         super().__init__()
  79
  80         assert dim_model % nb_heads == 0
  81
  82         self.embedding = nn.Sequential(
  83             nn.Embedding(vocabulary_size, dim_model),
  84             nn.Dropout(dropout),
  85             PositionalEncoding(len_max = 1e5),
  86         )
  87
  88         trunk_blocks = [ ]
  89
  90         for _ in range(nb_blocks):
  91             trunk_blocks += [
  92                 Residual(
  93                     nn.LayerNorm(dim_model),
  94                     QKVAttention(
  95                         dim_in = dim_model,
  96                         dim_qk = dim_keys, dim_v = dim_model // nb_heads,
  97                         nb_heads = nb_heads,
  98                         causal = True, attention_dropout = dropout
  99                     ),
 100                     nn.Linear(in_features = dim_model, out_features = dim_model),
 101                 ),
 102                 Residual(
 103                     nn.LayerNorm(dim_model),
 104                     nn.Linear(in_features = dim_model, out_features = dim_hidden),
 105                     nn.ReLU(),
 106                     nn.Linear(in_features = dim_hidden, out_features = dim_model),
 107                     nn.Dropout(dropout),
 108                 ),
 109             ]
 110
 111         self.trunk = nn.Sequential(*trunk_blocks)
 112
 113         self.readout = nn.Linear(in_features = dim_model, out_features = vocabulary_size)
 114
 115     def forward(self, x):
 116         x = self.embedding(x)
 117         x = self.trunk(x)
 118         x = self.readout(x)
 119         return x
 120
 121 ######################################################################
 122
 123 if __name__ == '__main__':
 124     vocabulary_size = 10
 125     x = torch.randint(vocabulary_size, (25, 100))
 126
 127     model = MyGPT(
 128         vocabulary_size = vocabulary_size,
 129         dim_model = 16, dim_keys = 50, dim_hidden = 100,
 130         nb_heads = 2, nb_blocks = 3,
 131         dropout = 0.1
 132     )
 133
 134     y = model(x)
 135
 136 ######################################################################