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