ofir1080/Transformer-Encoder

Implementation of Transformer encoder in PyTorch

Transformer Encoder

This repository provides a pytorch implementation of the encoder of Transformer.

Getting started

Build a transformer encoder

from transformer_encoder import TransformerEncoder

encoder = TransformerEncoder(d_model=512, d_ff=2048, n_heads=8, n_layers=6, dropout=0.1)

input_seqs = ...
mask = ...
out = encoder(input_seqs, mask)

Add positional encoding to input embeddings

import torch.nn as nn
from transformer_encoder.utils import PositionalEncoding

input_layer = nn.Sequential(
    nn.Embedding(num_embeddings=10000, embedding_dim=512),
    PositionalEncoding(d_model=512, dropout=0.1, max_len=5000)
)

Optimize model with the warming up strategy

import torch.optim as optim
from transformer_encoder.utils import WarmupOptimizer

model = ...

base_optimizer = optim.Adam(model.parameters(), lr=1e-3)
optimizer = WarmupOptimizer(base_optimizer, d_model=512, scale_factor=1, warmup_steps=100)

optimizer.zero_grad()
loss = ...
loss.backward()
optimizer.step()

API Reference

transformer_encoder.TransformerEncoder(d_model, d_ff, n_heads=1, n_layers=1, dropout=0.1)

• d_model: dimension of each word vector
• d_ff: hidden dimension of feed forward layer
• n_heads: number of heads in self-attention (defaults to 1)
• n_layers: number of stacked layers of encoder (defaults to 1)
• dropout: dropout rate (defaults to 0.1)

transformer_encoder.TransformerEncoder.forward(x, mask)

• x (~torch.FloatTensor): shape (batch_size, max_seq_len, d_model)
• mask (~torch.ByteTensor): shape (batch_size, max_seq_len)

transformer_encoder.utils.PositionalEncoding(d_model, dropout=0.1, max_len=5000)

• d_model: same as TransformerEncoder
• dropout: dropout rate (defaults to 0.1)
• max_len: max sequence length (defaults to 5000)

transformer_encoder.utils.PositionalEncoding.forward(x)

• x (~torch.FloatTensor): shape (batch_size, max_seq_len, d_model)

transformer_encoder.utils.WarmupOptimizer(base_optimizer, d_model, scale_factor, warmup_steps)

• base_optimizer (~torch.optim.Optimzier): e.g. adam optimzier
• d_model: equals d_model in TransformerEncoder
• scale_factor: scale factor of learning rate
• warmup_steps: warming up steps

Installation

Requires python 3.5+, pytorch 1.0.0+

pip install transformer_encoder