KLD (Kullback–Leibler divergence) annealing is a technique used in training variational autoencoders (VAE), specifically those based on an autoregressive decoder (ex. RNN such as LSTM or GRU). It is used to prevent the KL divergence term from vanishing during training. During the initial stages of training, as the encoder has not yet learned a meaningful representation of the training data, the autoregressive decoder may become misled by the ineffectuality of latent encodings and learn to ignore that information entirely. This results in a highly organized latent space but a poor reconstruction efficiency of the trained model and will negatively impact the overall performance of the autoencoder.
This repository contains a simple implementation of KLD annealing in PyTorch.
VAELoss is a subclass of torch.nn.Module. It is a standard loss function for training VAE (variational autoencoder) neural networks. It is callable and takes four arguments:
- x (torch.Tensor): reconstructed input tensor
- x0 (torch.Tensor): input tensor
- mu (torch.Tensor): mean of the latent space
- logvar (torch.Tensor): log variance of the latent space
When called, it returns two instances of single-valued torch.Tensor, which can be passed to an optimizer:
- bce (recon loss component)
- kld (divergence loss component)
Annealer is a class. An instance is created by passing two parameters:
- total_steps (int): Number of epochs (steps) to reach full KL divergence weight
- shape (str): Shape of the annealing function. Can be 'linear', 'cosine', or 'logistic'.
Annealer can be instantiated with disable=True parameter. This way an instance of Annealer can be still called, but the argument passed to the call() method will be returned unchanged.
An instance of Annealer is callable.
Args:
- kld (torch.tensor): Kullback–Leibler divergence loss (or any other loss object for which multiplication by a scalar is defined).
Returns:
- torch.tensor: KL divergence loss multiplied by the value of the annealing function
An instance of the Annealer class stores total_steps and current_steps integer attributes and uses them to calculate the value of the annealing coefficient at each step. Annealer is to be instantiated before the training loop, ex:
from loss import VAELoss
from annealing import Annealer
annealing_agent = Annealer(total_steps, shape='cosine') # instantiating annealing agent
criterion = VAELoss() # instantiating VAELoss
for epoch in range(100): # training loop starts here
x, mu, logvar = model(x0)
bce, kld = criterion(x, x0, mu, logvar)
kld = annealing_agent(kld)
optimizer.zero_grad()
(bcd + kld).backward
optimizer.step()
annealing_agent.step()
After each epoch annealing_agent.step() method is to be called. This increases the current_step attribute by 1 until the value matches the total_steps attribute. Based on those values, the slope of an annealing function is calculated. The initial value of current_step is always 0.
A method proposed by Fu et al. in their work Cyclical Annealing Schedule: A Simple Approach to Mitigating KL Vanishing
Cyclical annealing is enabled by instantiating the Annealer with cyclical=True parameter. When this is done, every time the step counter (current_step) reaches total_steps, its value is set to zero. The annealing cycle starts over again until it is disabled.
The cyclical annealing functionality can be disabled (or enabled) during training by passing False (or True) value to the Annealer.cyclical_setter() method. An example of cyclical annealing being disabled at epoch 30 and enabled again at epoch 70 is illustrated below:
annealing_agent = Annealer(total_steps, shape='cosine', cyclical=True)
# Instantiating annealing agent with cyclical annealing functionality
for epoch in range(100):
... # training script
if epoch == 30:
annealing_agent.cyclical_setter(False)
if epoch == 70:
annealing_agent.cyclical_setter(True)
The expected outcome:
