Here is my Transformer
import math
import torch
from torch import nn
from torch.nn import TransformerEncoder, TransformerEncoderLayer, TransformerDecoder, TransformerDecoderLayer
from torch.nn.modules.normalization import LayerNorm
class Transformer(nn.Module):
def __init__(self, input_dim: int, output_dim: int, d_model: int = 512, num_head: int = 8, num_e_layer: int = 6,
num_d_layer: int = 6, ff_dim: int = 2048, drop_out: float = 0.1):
'''
Args:
input_dim: Size of the vocab of the input
output_dim: Size of the vocab for output
num_head: Number of heads in mutliheaded attention models
num_e_layer: Number of sub-encoder layers
num_d_layer: Number of sub-decoder layers
ff_dim: Dimension of feedforward network in mulihead models
d_model: The dimension to embed input and output features into
drop_out: The drop out percentage
'''
super(Transformer, self).__init__()
self.decoder = TDecoder(output_dim, d_model, num_head, ff_dim, num_d_layer, drop_out)
self.encoder = TEncoder(input_dim, d_model, num_head, ff_dim, num_e_layer, drop_out)
def forward(self, src: torch.Tensor, trg: torch.Tensor)
enc_out = self.encoder(src)
return self.decoder(trg, enc_out)
class TEncoder(nn.Module):
def __init__(self, input_dim: int, d_model: int, num_head: int, ff_dim: int, num_layers: int, drop_out: float):
super(TEncoder, self).__init__()
self.pos_encoder = PositionalEncoding(d_model, drop_out)
self.embed = nn.Embedding(input_dim, d_model)
# Encoder layer is the multiheaded attention part of the transformer
# Norm is the normalization after every multiheaded portion
# ff_dim is the dimension for feedforward network at the end
layer = TransformerEncoderLayer(d_model, num_head, ff_dim, drop_out)
norm = LayerNorm(d_model)
self.encoder = TransformerEncoder(layer, num_layers, norm)
def forward(self, src: torch.Tensor):
src_embed = self.pos_encoder(self.embed(src))
return self.encoder(src_embed)
class TDecoder(nn.Module)
def __init__(self, input_dim: int, d_model: int, num_head: int, ff_dim: int, num_layers: int, drop_out: float):
super().__init__()
self.pos_encoder = PositionalEncoding(d_model, drop_out)
self.embed = nn.Embedding(input_dim, d_model)
# Same as Encoder
layer = TransformerDecoderLayer(d_model, num_head, ff_dim, drop_out)
norm = LayerNorm(d_model)
self.decoder = TransformerDecoder(layer, num_layers, norm)
# Ends with a linear layer and a softmax
self.linear = nn.Linear(d_model, input_dim)
self.softmax = nn.Softmax(dim=2)
def forward(self, trg: torch.Tensor, encoder_output: torch.Tensor):
dec_mask = self._generate_square_subsequent_mask(len(trg))
trg_embed = self.pos_encoder(self.embed(trg))
dec_out = self.decoder(trg_embed, encoder_output, tgt_mask=dec_mask)
output = self.linear(dec_out)
return self.softmax(output)
def _generate_square_subsequent_mask(self, sz):
mask = (torch.triu(torch.ones(sz, sz)) == 1).transpose(0, 1)
mask = mask.float().masked_fill(mask == 0, float('-inf')).masked_fill(mask == 1, float(0.0))
return mask
class PositionalEncoding(nn.Module):
def __init__(self, d_model, drop_out=0.1, max_len=200):
super(PositionalEncoding, self).__init__()
self.drop_out = nn.Dropout(p=drop_out)
pe = torch.zeros(max_len, d_model)
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
pe[:, 0::2] = torch.sin(position * div_term)
pe[:, 1::2] = torch.cos(position * div_term)
pe = pe.unsqueeze(0).transpose(0, 1)
self.register_buffer('pe', pe)
def forward(self, x):
x = x + self.pe[:x.size(0), :]
return self.drop_out(x)
Just to make sure I implemented it correctly I’m testing to see if I can put just a made up seq of length 5 and put that through the encoder, then pretend it generated 2 characters already, with which I already took the best probs from both and appending them then put them through the decoder.
input = torch.ones(5).type(torch.LongTensor)
trg = torch.zeros(4).type(torch.LongTensor)
transformer = Transformer(2, 1)
output = transformer.forward(input, trg)
In this example the encoder input dim is 2 and decoder input dim is 1. For some reason I get the above error.