Permalink
Fetching contributors…
Cannot retrieve contributors at this time
89 lines (75 sloc) 4.69 KB

Unorthodox Design

There are a couple design decisions about DyNet that are different from the way things are implemented in other libraries, or different from the way you might expect things to be implemented. The items below are a list of these unorthodox design decisions, which you should read to avoid being surprised. We also try to give some justification for these decisions (although we realize that this is not the only right way to do things).

Sparse Updates

By default, DyNet parameter optimizers perform sparse updates over LookupParameters. This means that if you have a LookupParameters object, use a certain subset of indices, then perform a parameter update, the optimizer will loop over the used subset, and not perform any updates over the unused values. This can improve efficiency in some cases: e.g. if you have embeddings for a vocabulary of 100,000 words and you only use 5 of them in a particular update, this will avoid doing updates over all 100,000. However, there are two things to be careful of. First, this means that some update rules such as ones using momentum such as MomentumSGDTrainer and AdamTrainer are not strictly correct (these could be made correct with some effort, but this would complicate the programming interface, which we have opted against). Also, on GPUs, because large operations are relatively cheap, it can sometimes be faster to just perform a single operation over all of the parameters, as opposed to multiple small operations. In this case, you can set the sparse_updates_enabled variable of your Trainer to false, and DyNet will perform a standard dense update, which is guaranteed to be exactly correct, and potentially faster on GPU.

Weight Decay

As described in the :ref:`command-line-options`, weight decay is implemented through the option --dynet-weight-decay. If this value is set to wd, each parameter in the model is multiplied by (1-wd) after every parameter update. This weight decay is similar to L2 regularization, and is equivalent in the case of using simple SGD (SimpleSGDTrainer), but it is not the same when using any other optimizers such as AdagradTrainer or AdamTrainer. You can still try to use weight decay with these optimizers, and it might work, but if you really want to correctly apply L2 regularization with these optimizers, you will have to directly calculate the L2 norm of each of the parameters and add it to the objective function before performing your update.

Minibatching Implementation

:ref:`minibatching` in DyNet is different than how it is implemented in other libraries. In other libraries, you can create minibatches by explicitly adding another dimension to each of the variables that you want to process, and managing them yourself. Instead, DyNet provides special :ref:`operations` that allow you to perform input, lookup, or loss calculation over mini-batched input, then DyNet will handle the rest. The programming paradigm is a bit different from other toolkits, and may take a bit of getting used to, but is often more convenient once you're used to it.

LSTM Implementation

The implementation of LSTMs in LSTMBuilder is not the canonical implementation, but an implementation using coupled input and forget gates, as described in "LSTM: A Search Space Odyssey" (https://arxiv.org/abs/1503.04069). In other words, if the value of the input gate is i, the forget gate is 1-i. This reduces the number of parameters in the model and speeds training a little, and in many cases the accuracy is the same or better. If you want to try the standard version of the LSTM, use the VanillaLSTMBuilder class.

Dropout Scaling

When using dropout to help prevent overfitting, dropout is generally applied at training time, then at test time all the nodes in the neural net are used to make the final decision, increasing robustness. However, because there is a disconnect between the number of nodes being used in each situation, it is important to scale the values of the output to ensure that they match in both situations. There are two ways to do this:

  • Vanilla Dropout: At training time, perform dropout with probability p. At test time, scale the outputs of each node by p.
  • Inverted Dropout: At training time, perform dropout with probability p, and scale the outputs by 1/p. At test time, use the outputs as-is.

The first is perhaps more common, but the second is convenient, because we only need to think about dropout at training time, and thus DyNet opts to use the latter. See here for more details on these two methods.