Weight subcloning: direct initialization of transformers using larger pretrained ones
This addresses a practical bottleneck for researchers and practitioners needing efficient training of smaller transformers, though it is an incremental improvement over existing transfer learning methods.
The paper tackles the problem of training smaller transformer models when no pretrained model of the required size is available, by introducing weight subcloning to initialize scaled-down transformers from larger pretrained ones, resulting in up to 4x faster training speed for vision and language tasks.
Training large transformer models from scratch for a target task requires lots of data and is computationally demanding. The usual practice of transfer learning overcomes this challenge by initializing the model with weights of a pretrained model of the same size and specification to increase the convergence and training speed. However, what if no pretrained model of the required size is available? In this paper, we introduce a simple yet effective technique to transfer the knowledge of a pretrained model to smaller variants. Our approach called weight subcloning expedites the training of scaled-down transformers by initializing their weights from larger pretrained models. Weight subcloning involves an operation on the pretrained model to obtain the equivalent initialized scaled-down model. It consists of two key steps: first, we introduce neuron importance ranking to decrease the embedding dimension per layer in the pretrained model. Then, we remove blocks from the transformer model to match the number of layers in the scaled-down network. The result is a network ready to undergo training, which gains significant improvements in training speed compared to random initialization. For instance, we achieve 4x faster training for vision transformers in image classification and language models designed for next token prediction.