Transformer-based models, such as BERT and ViT, have achieved state-of-the-art results across different natural language processing (NLP) and computer vision (CV) tasks. However, these models are extremely memory intensive during their fine-tuning process, making them difficult to deploy on GPUs with limited memory resources. To address this issue, we introduce a new tool called SlimFit that reduces the memory requirements of these models by dynamically analyzing their training dynamics and freezing less-contributory layers during fine-tuning. The layers to freeze are chosen using a runtime inter-layer scheduling algorithm. This allows SlimFit to freeze up to 95% of layers and reduce the overall on-device GPU memory usage of transformer-based models such as ViT and BERT by an average of 2.2x, across different NLP and CV benchmarks/datasets such as GLUE, SQuAD 2.0, CIFAR-10, CIFAR-100 and ImageNet with an average degradation of 0.2% in accuracy. For such NLP and CV tasks, SlimFit can reduce up to 3.1x the total on-device memory usage with an accuracy degradation of only up to 0.4%. As a result, while fine-tuning of ViT on ImageNet and BERT on SQuAD 2.0 with a batch size of 128 requires 3 and 2 32GB GPUs, respectively, SlimFit enables fine-tuning them on a single 32GB GPU without any significant accuracy degradation. The code of SlimFit is available at https://github.com/arashardakani/SlimFit.

In the past few years, pre-trained BERT has become one of the most popular deep-learning language models due to their remarkable performance in natural language processing (NLP) tasks. However, the superior performance of BERT comes at the cost of high computational and memory complexity, hindering its envisioned widespread deployment in edge devices with limited computing resources. Binarization can alleviate these limitations by reducing storage requirements and improving computing performance. However, obtaining a comparable accuracy performance for binary BERT w.r.t. its full-precision counterpart is still a difficult task. We observe that direct binarization of pre-trained BERT provides a poor initialization during the fine-tuning phase, making the model incapable of achieving a decent accuracy on downstream tasks. Based on this observation, we put forward the following hypothesis: partially randomly-initialized BERT with binary weights and activations can reach to a decent accuracy performance by distilling knowledge from the its full-precision counterpart. We show that BERT with pre-trained embedding layer and randomly-initialized encoder is a smoking gun for this hypothesis. We identify the smoking gun through a series of experiments and show that it yields a new set of state-of-the-art results on the GLUE and SQuAD benchmarks.

The success of large BERT models has raised the demand for model compression methods to reduce model size and computational cost. Quantization can reduce the model size and inference latency, making inference more efficient, without changing its stucture, but it comes at the cost of performance degradation. Due to the complex loss landscape of ternarized/binarized BERT, we present an efficient two-stage progressive quantization method in which we fine tune the model with quantized weights and progressively lower its bits, and then we fine tune the model with quantized weights and activations. At the same time, we strategically choose which bitwidth to fine-tune on and to initialize from, and which bitwidth to fine-tune under augmented data to outperform the existing BERT binarization methods without adding an extra module, compressing the binary model 18% more than previous binarization methods or compressing BERT by 31x w.r.t. to the full-precision model. Our method without data augmentation can outperform existing BERT ternarization methods.