In recent years, Transformer-based language models have become the standard approach for natural language processing tasks. However, stringent throughput and latency requirements in industrial applications are limiting their adoption. To mitigate the gap, model compression techniques such as structured pruning are being used to improve inference efficiency. However, most existing neural network inference runtimes lack adequate support for structured sparsity. In this paper, we propose an efficient sparse deep learning inference software stack for Transformer-based language models where the weights are pruned with constant block size. Our sparse software accelerator leverages Intel Deep Learning Boost to maximize the performance of sparse matrix - dense matrix multiplication (commonly abbreviated as SpMM) on CPUs. Our SpMM kernel outperforms the existing sparse libraries (oneMKL, TVM, and LIBXSMM) by an order of magnitude on a wide range of GEMM shapes under 5 representative sparsity ratios (70%, 75%, 80%, 85%, 90%). Moreover, our SpMM kernel shows up to 5x speedup over dense GEMM kernel of oneDNN, a well-optimized dense library widely used in industry. We apply our sparse accelerator on widely-used Transformer-based language models including Bert-Mini, DistilBERT, Bert-Base, and BERT-Large. Our sparse inference software shows up to 1.5x speedup over Neural Magic's Deepsparse under same configurations on Xeon on Amazon Web Services under proxy production latency constraints. We also compare our solution with two framework-based inference solutions, ONNX Runtime and PyTorch, and demonstrate up to 37x speedup over ONNX Runtime and 345x over PyTorch on Xeon under the latency constraints. All the source code is publicly available on Github: https://github.com/intel/intel-extension-for-transformers.
The components underpinning PLMs -- large weight matrices -- were shown to bear considerable redundancy. Matrix factorization, a well-established technique from matrix theory, has been utilized to reduce the number of parameters in PLM. However, it fails to retain satisfactory performance under moderate to high compression rate. In this paper, we identify the \textit{full-rankness} of fine-tuned PLM as the fundamental bottleneck for the failure of matrix factorization and explore the use of network pruning to extract low-rank sparsity pattern desirable to matrix factorization. We find such low-rank sparsity pattern exclusively exists in models generated by first-order pruning, which motivates us to unite the two approaches and achieve more effective model compression. We further propose two techniques: sparsity-aware SVD and mixed-rank fine-tuning, which improve the initialization and training of the compression procedure, respectively. Experiments on GLUE and question-answering tasks show that the proposed method has superior compression-performance trade-off compared to existing approaches.
Deploying pre-trained transformer models like BERT on downstream tasks in resource-constrained scenarios is challenging due to their high inference cost, which grows rapidly with input sequence length. In this work, we propose a constraint-aware and ranking-distilled token pruning method ToP, which selectively removes unnecessary tokens as input sequence passes through layers, allowing the model to improve online inference speed while preserving accuracy. ToP overcomes the limitation of inaccurate token importance ranking in the conventional self-attention mechanism through a ranking-distilled token distillation technique, which distills effective token rankings from the final layer of unpruned models to early layers of pruned models. Then, ToP introduces a coarse-to-fine pruning approach that automatically selects the optimal subset of transformer layers and optimizes token pruning decisions within these layers through improved $L_0$ regularization. Extensive experiments on GLUE benchmark and SQuAD tasks demonstrate that ToP outperforms state-of-the-art token pruning and model compression methods with improved accuracy and speedups. ToP reduces the average FLOPs of BERT by 8.1x while achieving competitive accuracy on GLUE, and provides a real latency speedup of up to 7.4x on an Intel CPU.
Due to the point cloud's irregular and unordered geometry structure, conventional knowledge distillation technology lost a lot of information when directly used on point cloud tasks. In this paper, we propose Feature Adversarial Distillation (FAD) method, a generic adversarial loss function in point cloud distillation, to reduce loss during knowledge transfer. In the feature extraction stage, the features extracted by the teacher are used as the discriminator, and the students continuously generate new features in the training stage. The feature of the student is obtained by attacking the feedback from the teacher and getting a score to judge whether the student has learned the knowledge well or not. In experiments on standard point cloud classification on ModelNet40 and ScanObjectNN datasets, our method reduced the information loss of knowledge transfer in distillation in 40x model compression while maintaining competitive performance.
Compactness in deep learning can be critical to a model's viability in low-resource applications, and a common approach to extreme model compression is quantization. We consider Iterative Product Quantization (iPQ) with Quant-Noise to be state-of-the-art in this area, but this quantization framework suffers from preventable inference quality degradation due to prevalent empty clusters. In this paper, we propose several novel enhancements aiming to improve the accuracy of iPQ with Quant-Noise by focusing on resolving empty clusters. Our contribution, which we call Partitioning-Guided k-means (PG k-means), is a heavily augmented k-means implementation composed of three main components. First, we propose a partitioning-based pre-assignment strategy that ensures no initial empty clusters and encourages an even weight-to-cluster distribution. Second, we propose an empirically superior empty cluster resolution heuristic executed via cautious partitioning of large clusters. Finally, we construct an optional optimization step that consolidates intuitively dense clusters of weights to ensure shared representation. The proposed approach consistently reduces the number of empty clusters in iPQ with Quant-Noise by 100x on average, uses 8x fewer iterations during empty cluster resolution, and improves overall model accuracy by up to 12%, when applied to RoBERTa on a variety of tasks in the GLUE benchmark.
Model compression techniques reduce the computational load and memory consumption of deep neural networks. After the compression operation, e.g. parameter pruning, the model is normally fine-tuned on the original training dataset to recover from the performance drop caused by compression. However, the training data is not always available due to privacy issues or other factors. In this work, we present a data-free fine-tuning approach for pruning the backbone of deep neural networks. In particular, the pruned network backbone is trained with synthetically generated images, and our proposed intermediate supervision to mimic the unpruned backbone's output feature map. Afterwards, the pruned backbone can be combined with the original network head to make predictions. We generate synthetic images by back-propagating gradients to noise images while relying on L1-pruning for the backbone pruning. In our experiments, we show that our approach is task-independent due to pruning only the backbone. By evaluating our approach on 2D human pose estimation, object detection, and image classification, we demonstrate promising performance compared to the unpruned model. Our code is available at https://github.com/holzbock/dfbf.
With the increase in the scale of Deep Learning (DL) training workloads in terms of compute resources and time consumption, the likelihood of encountering in-training failures rises substantially, leading to lost work and resource wastage. Such failures are typically offset by a checkpointing mechanism, which comes at the cost of storage and network bandwidth overhead. State-of-the-art approaches involve lossy model compression mechanisms, which induce a tradeoff between the resulting model quality (accuracy) and compression ratio. Delta compression is then also used to further reduce the overhead by only storing the difference between consecutive checkpoints. We make a key enabling observation that the sensitivity of model weights to compression varies during training, and different weights benefit from different quantization levels (ranging from retaining full precision to pruning). We propose (1) a non-uniform quantization scheme that leverages this variation, (2) an efficient search mechanism to dynamically adjust to the best quantization configurations, and (3) a quantization-aware delta compression mechanism that rearranges weights to minimize checkpoint differences, thereby maximizing compression. We instantiate these contributions in DynaQuant - a framework for DL workload checkpoint compression. Our experiments show that DynaQuant consistently achieves better tradeoff between accuracy and compression ratios compared to prior works, enabling a compression ratio up to 39x and withstanding up to 10 restores with negligible accuracy impact for fault-tolerant training. DynaQuant achieves at least an order of magnitude reduction in checkpoint storage overhead for training failure recovery as well as transfer learning use cases without any loss of accuracy
Transformer models have achieved remarkable results in various natural language tasks, but they are often prohibitively large, requiring massive memories and computational resources. To reduce the size and complexity of these models, we propose LoSparse (Low-Rank and Sparse approximation), a novel model compression technique that approximates a weight matrix by the sum of a low-rank matrix and a sparse matrix. Our method combines the advantages of both low-rank approximations and pruning, while avoiding their limitations. Low-rank approximation compresses the coherent and expressive parts in neurons, while pruning removes the incoherent and non-expressive parts in neurons. Pruning enhances the diversity of low-rank approximations, and low-rank approximation prevents pruning from losing too many expressive neurons. We evaluate our method on natural language understanding, question answering, and natural language generation tasks. We show that it significantly outperforms existing compression methods.
Knowledge Distillation (KD) has been validated as an effective model compression technique for learning compact object detectors. Existing state-of-the-art KD methods for object detection are mostly based on feature imitation. In this paper, we present a general and effective prediction mimicking distillation scheme, called CrossKD, which delivers the intermediate features of the student's detection head to the teacher's detection head. The resulting cross-head predictions are then forced to mimic the teacher's predictions. This manner relieves the student's head from receiving contradictory supervision signals from the annotations and the teacher's predictions, greatly improving the student's detection performance. Moreover, as mimicking the teacher's predictions is the target of KD, CrossKD offers more task-oriented information in contrast with feature imitation. On MS COCO, with only prediction mimicking losses applied, our CrossKD boosts the average precision of GFL ResNet-50 with 1x training schedule from 40.2 to 43.7, outperforming all existing KD methods. In addition, our method also works well when distilling detectors with heterogeneous backbones. Code is available at https://github.com/jbwang1997/CrossKD.
Learning-based video compression is currently a popular research topic, offering the potential to compete with conventional standard video codecs. In this context, Implicit Neural Representations (INRs) have previously been used to represent and compress image and video content, demonstrating relatively high decoding speed compared to other methods. However, existing INR-based methods have failed to deliver rate quality performance comparable with the state of the art in video compression. This is mainly due to the simplicity of the employed network architectures, which limit their representation capability. In this paper, we propose HiNeRV, an INR that combines light weight layers with novel hierarchical positional encodings. We employs depth-wise convolutional, MLP and interpolation layers to build the deep and wide network architecture with high capacity. HiNeRV is also a unified representation encoding videos in both frames and patches at the same time, which offers higher performance and flexibility than existing methods. We further build a video codec based on HiNeRV and a refined pipeline for training, pruning and quantization that can better preserve HiNeRV's performance during lossy model compression. The proposed method has been evaluated on both UVG and MCL-JCV datasets for video compression, demonstrating significant improvement over all existing INRs baselines and competitive performance when compared to learning-based codecs (72.3% overall bit rate saving over HNeRV and 43.4% over DCVC on the UVG dataset, measured in PSNR).
Quantization and pruning are two effective Deep Neural Networks model compression methods. In this paper, we propose Automatic Prune Binarization (APB), a novel compression technique combining quantization with pruning. APB enhances the representational capability of binary networks using a few full-precision weights. Our technique jointly maximizes the accuracy of the network while minimizing its memory impact by deciding whether each weight should be binarized or kept in full precision. We show how to efficiently perform a forward pass through layers compressed using APB by decomposing it into a binary and a sparse-dense matrix multiplication. Moreover, we design two novel efficient algorithms for extremely quantized matrix multiplication on CPU, leveraging highly efficient bitwise operations. The proposed algorithms are 6.9x and 1.5x faster than available state-of-the-art solutions. We extensively evaluate APB on two widely adopted model compression datasets, namely CIFAR10 and ImageNet. APB delivers better accuracy/memory trade-off compared to state-of-the-art methods based on i) quantization, ii) pruning, and iii) combination of pruning and quantization. APB outperforms quantization in the accuracy/efficiency trade-off, being up to 2x faster than the 2-bit quantized model with no loss in accuracy.
Hypernetworks, or hypernets for short, are neural networks that generate weights for another neural network, known as the target network. They have emerged as a powerful deep learning technique that allows for greater flexibility, adaptability, dynamism, faster training, information sharing, and model compression. Hypernets have shown promising results in a variety of deep learning problems, including continual learning, causal inference, transfer learning, weight pruning, uncertainty quantification, zero-shot learning, natural language processing, and reinforcement learning. Despite their success across different problem settings, there is currently no comprehensive review available to inform researchers about the latest developments and to assist in utilizing hypernets. To fill this gap, we review the progress in hypernets. We present an illustrative example of training deep neural networks using hypernets and propose categorizing hypernets based on five design criteria: inputs, outputs, variability of inputs and outputs, and the architecture of hypernets. We also review applications of hypernets across different deep learning problem settings, followed by a discussion of general scenarios where hypernets can be effectively employed. Finally, we discuss the challenges and future directions that remain underexplored in the field of hypernets. We believe that hypernetworks have the potential to revolutionize the field of deep learning. They offer a new way to design and train neural networks, and they have the potential to improve the performance of deep learning models on a variety of tasks. Through this review, we aim to inspire further advancements in deep learning through hypernetworks.
Quantization-aware training (QAT) is a representative model compression method to reduce redundancy in weights and activations. However, most existing QAT methods require end-to-end training on the entire dataset, which suffers from long training time and high energy costs. In addition, the potential label noise in the training data undermines the robustness of QAT. We propose two metrics based on analysis of loss and gradient of quantized weights: error vector score and disagreement score, to quantify the importance of each sample during training. Guided by these two metrics, we proposed a quantization-aware Adaptive Coreset Selection (ACS) method to select the data for the current training epoch. We evaluate our method on various networks (ResNet-18, MobileNetV2, RetinaNet), datasets(CIFAR-10, CIFAR-100, ImageNet-1K, COCO), and under different quantization settings. Specifically, our method can achieve an accuracy of 68.39\% of 4-bit quantized ResNet-18 on the ImageNet-1K dataset with only a 10\% subset, which has an absolute gain of 4.24\% compared to the baseline. Our method can also improve the robustness of QAT by removing noisy samples in the training set.
Natural language understanding (NLU) tasks face a non-trivial amount of ambiguous samples where veracity of their labels is debatable among annotators. NLU models should thus account for such ambiguity, but they approximate the human opinion distributions quite poorly and tend to produce over-confident predictions. To address this problem, we must consider how to exactly capture the degree of relationship between each sample and its candidate classes. In this work, we propose a novel method with deep model compression and show how such relationship can be accounted for. We see that more reasonably represented relationships can be discovered in the lower layers and that validation accuracies are converging at these layers, which naturally leads to layer pruning. We also see that distilling the relationship knowledge from a lower layer helps models produce better distribution. Experimental results demonstrate that our method makes substantial improvement on quantifying ambiguity without gold distribution labels. As positive side-effects, our method is found to reduce the model size significantly and improve latency, both attractive aspects of NLU products.
Deploying NMT models on mobile devices is essential for privacy, low latency, and offline scenarios. For high model capacity, NMT models are rather large. Running these models on devices is challenging with limited storage, memory, computation, and power consumption. Existing work either only focuses on a single metric such as FLOPs or general engine which is not good at auto-regressive decoding. In this paper, we present MobileNMT, a system that can translate in 15MB and 30ms on devices. We propose a series of principles for model compression when combined with quantization. Further, we implement an engine that is friendly to INT8 and decoding. With the co-design of model and engine, compared with the existing system, we speed up 47.0x and save 99.5% of memory with only 11.6% loss of BLEU. The code is publicly available at https://github.com/zjersey/Lightseq-ARM.
We present a work on low-complexity acoustic scene classification (ASC) with multiple devices, namely the subtask A of Task 1 of the DCASE2021 challenge. This subtask focuses on classifying audio samples of multiple devices with a low-complexity model, where two main difficulties need to be overcome. First, the audio samples are recorded by different devices, and there is mismatch of recording devices in audio samples. We reduce the negative impact of the mismatch of recording devices by using some effective strategies, including data augmentation (e.g., mix-up, spectrum correction, pitch shift), usages of multi-patch network structure and channel attention. Second, the model size should be smaller than a threshold (e.g., 128 KB required by the DCASE2021 challenge). To meet this condition, we adopt a ResNet with both depthwise separable convolution and channel attention as the backbone network, and perform model compression. In summary, we propose a low-complexity ASC method using data augmentation and a lightweight ResNet. Evaluated on the official development and evaluation datasets, our method obtains classification accuracy scores of 71.6% and 66.7%, respectively; and obtains Log-loss scores of 1.038 and 1.136, respectively. Our final model size is 110.3 KB which is smaller than the maximum of 128 KB.
Pre-trained Transformer models like T5 and BART have advanced the state of the art on a wide range of text generation tasks. Compressing these models into smaller ones has become critically important for practical use. Common neural network compression techniques such as knowledge distillation or quantization are limited to static compression where the compression ratio is fixed. In this paper, we introduce Modular Transformers, a modularized encoder-decoder framework for flexible sequence-to-sequence model compression. Modular Transformers train modularized layers that have the same function of two or more consecutive layers in the original model via module replacing and knowledge distillation. After training, the modularized layers can be flexibly assembled into sequence-to-sequence models that meet different performance-efficiency trade-offs. Experimental results show that after a single training phase, by simply varying the assembling strategy, Modular Transformers can achieve flexible compression ratios from 1.1x to 6x with little to moderate relative performance drop.
Low-rank model compression is a widely used technique for reducing the computational load when training machine learning models. However, existing methods often rely on relaxing the low-rank constraint of the model weights using a regularized nuclear norm penalty, which requires an appropriate hyperparameter that can be difficult to determine in practice. Furthermore, existing compression techniques are not directly applicable to efficient over-the-air (OTA) aggregation in federated learning (FL) systems for distributed Internet-of-Things (IoT) scenarios. In this paper, we propose a novel manifold optimization formulation for low-rank model compression in FL that does not relax the low-rank constraint. Our optimization is conducted directly over the low-rank manifold, guaranteeing that the model is exactly low-rank. We also introduce a consensus penalty in the optimization formulation to support OTA aggregation. Based on our optimization formulation, we propose an alternating Riemannian optimization algorithm with a precoder that enables efficient OTA aggregation of low-rank local models without sacrificing training performance. Additionally, we provide convergence analysis in terms of key system parameters and conduct extensive experiments with real-world datasets to demonstrate the effectiveness of our proposed Riemannian low-rank model compression scheme compared to various state-of-the-art baselines.
Self-supervised pre-trained models such as Wav2vec2, Hubert, and WavLM have been shown to significantly improve many speech tasks. However, their large memory and strong computational requirements hinder their industrial applicability. Structured pruning is a hardware-friendly model compression technique but usually results in a larger loss of accuracy. In this paper, we propose a fine-grained attention head pruning method to compensate for the performance degradation. In addition, we also introduce the straight through estimator into the L0 regularization to further accelerate the pruned model. Experiments on the SUPERB benchmark show that our model can achieve comparable performance to the dense model in multiple tasks and outperforms the Wav2vec 2.0 base model on average, with 72% fewer parameters and 2 times faster inference speed.
Due to the over-parameterization of neural networks, many model compression methods based on pruning and quantization have emerged. They are remarkable in reducing the size, parameter number, and computational complexity of the model. However, most of the models compressed by such methods need the support of special hardware and software, which increases the deployment cost. Moreover, these methods are mainly used in classification tasks, and rarely directly used in detection tasks. To address these issues, for the object detection network we introduce a three-stage model compression method: dynamic sparse training, group channel pruning, and spatial attention distilling. Firstly, to select out the unimportant channels in the network and maintain a good balance between sparsity and accuracy, we put forward a dynamic sparse training method, which introduces a variable sparse rate, and the sparse rate will change with the training process of the network. Secondly, to reduce the effect of pruning on network accuracy, we propose a novel pruning method called group channel pruning. In particular, we divide the network into multiple groups according to the scales of the feature layer and the similarity of module structure in the network, and then we use different pruning thresholds to prune the channels in each group. Finally, to recover the accuracy of the pruned network, we use an improved knowledge distillation method for the pruned network. Especially, we extract spatial attention information from the feature maps of specific scales in each group as knowledge for distillation. In the experiments, we use YOLOv4 as the object detection network and PASCAL VOC as the training dataset. Our method reduces the parameters of the model by 64.7 % and the calculation by 34.9%.