ELECTRA¶
Overview¶
The ELECTRA model was proposed in the paper ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators. ELECTRA is a new pretraining approach which trains two transformer models: the generator and the discriminator. The generator’s role is to replace tokens in a sequence, and is therefore trained as a masked language model. The discriminator, which is the model we’re interested in, tries to identify which tokens were replaced by the generator in the sequence.
The abstract from the paper is the following:
Masked language modeling (MLM) pretraining methods such as BERT corrupt the input by replacing some tokens with [MASK] and then train a model to reconstruct the original tokens. While they produce good results when transferred to downstream NLP tasks, they generally require large amounts of compute to be effective. As an alternative, we propose a more sample-efficient pretraining task called replaced token detection. Instead of masking the input, our approach corrupts it by replacing some tokens with plausible alternatives sampled from a small generator network. Then, instead of training a model that predicts the original identities of the corrupted tokens, we train a discriminative model that predicts whether each token in the corrupted input was replaced by a generator sample or not. Thorough experiments demonstrate this new pretraining task is more efficient than MLM because the task is defined over all input tokens rather than just the small subset that was masked out. As a result, the contextual representations learned by our approach substantially outperform the ones learned by BERT given the same model size, data, and compute. The gains are particularly strong for small models; for example, we train a model on one GPU for 4 days that outperforms GPT (trained using 30x more compute) on the GLUE natural language understanding benchmark. Our approach also works well at scale, where it performs comparably to RoBERTa and XLNet while using less than 1/4 of their compute and outperforms them when using the same amount of compute.
Tips:
ELECTRA is the pretraining approach, therefore there is nearly no changes done to the underlying model: BERT. The only change is the separation of the embedding size and the hidden size: the embedding size is generally smaller, while the hidden size is larger. An additional projection layer (linear) is used to project the embeddings from their embedding size to the hidden size. In the case where the embedding size is the same as the hidden size, no projection layer is used.
The ELECTRA checkpoints saved using Google Research’s implementation contain both the generator and discriminator. The conversion script requires the user to name which model to export into the correct architecture. Once converted to the HuggingFace format, these checkpoints may be loaded into all available ELECTRA models, however. This means that the discriminator may be loaded in the
ElectraForMaskedLMmodel, and the generator may be loaded in theElectraForPreTrainingmodel (the classification head will be randomly initialized as it doesn’t exist in the generator).
The original code can be found here.