Search for: All records

It has been shown by many researchers that transformers perform as well as convolutional neural networks in many computer vision tasks. Meanwhile, the large computational costs of its attention module hinder further studies and applications on edge devices. Some pruning methods have been developed to construct efficient vision transformers, but most of them have considered image classification tasks only. Inspired by these results, we propose SiDT, a method for pruning vision transformer backbones on more complicated vision tasks like object detection, based on the search of transformer dimensions. Experiments on CIFAR-100 and COCO datasets show that the backbones with 20% or 40% dimensions/parameters pruned can have similar or even better performance than the unpruned models. Moreover, we have also provided the complexity analysis and comparisons with the previous pruning methods. 

Multi-resolution paths and multi-scale feature representation are key elements of semantic segmentation networks. We develop two techniques for efficient networks based on the recent FasterSeg network architecture. One is to use a state-of-the-art high resolution network (e.g. HRNet) as a teacher to distill a light weight student network. Due to dissimilar structures in the teacher and student networks, distillation is not effective to be carried out directly in a standard way. To solve this problem, we introduce a tutor network with an added high resolution path to help distill a student network which improves FasterSeg student while maintaining its parameter/FLOPs counts. The other finding is to replace standard bilinear interpolation in the upscaling module of FasterSeg student net by a depth-wise separable convolution and a Pixel Shuffle module which leads to 1.9% (1.4%) mIoU improvements on low (high) input image sizes without increasing model size. A combination of these techniques will be pursued in future works. 

Convolutional neural networks (CNN) have been hugely successful recently with superior accuracy and performance in various imaging applications, such as classification, object detection, and segmentation. However, a highly accurate CNN model requires millions of parameters to be trained and utilized. Even to increase its performance slightly would require significantly more parameters due to adding more layers and/or increasing the number of filters per layer. Apparently, many of these weight parameters turn out to be redundant and extraneous, so the original, dense model can be replaced by its compressed version attained by imposing inter- and intra-group sparsity onto the layer weights during training. In this paper, we propose a nonconvex family of sparse group lasso that blends nonconvex regularization (e.g., transformed ℓ1, ℓ1 − ℓ2, and ℓ0) that induces sparsity onto the individual weights and ℓ2,1 regularization onto the output channels of a layer. We apply variable splitting onto the proposed regularization to develop an algorithm that consists of two steps per iteration: gradient descent and thresholding. Numerical experiments are demonstrated on various CNN architectures showcasing the effectiveness of the nonconvex family of sparse group lasso in network sparsification and test accuracy on par with the current state of the art. 

In the last decade, convolutional neural networks (CNNs) have evolved to become the dominant models for various computer vision tasks, but they cannot be deployed in low-memory devices due to its high memory requirement and computational cost. One popular, straightforward approach to compressing CNNs is network slimming, which imposes an ℓ1 penalty on the channel-associated scaling factors in the batch normalization layers during training. In this way, channels with low scaling factors are identified to be insignificant and are pruned in the models. In this paper, we propose replacing the ℓ1 penalty with the ℓp and transformed ℓ1 (T ℓ1 ) penalties since these nonconvex penalties outperformed ℓ1 in yielding sparser satisfactory solutions in various compressed sensing problems. In our numerical experiments, we demonstrate network slimming with ℓp and T ℓ1 penalties on VGGNet and Densenet trained on CIFAR 10/100. The results demonstrate that the nonconvex penalties compress CNNs better than ℓ1 . In addition, T ℓ1 preserves the model accuracy after channel pruning, and ℓ1/2,3/4 yield compressed models with similar accuracies as ℓ1 after retraining. 

ShuffleNet is a state-of-the-art light weight convolutional neural network architecture. Its basic operations include group, channelwise convolution and channel shuffling. However, channel shuffling is manually designed on empirical grounds. Mathematically, shuffling is a multiplication by a permutation matrix. In this paper, we propose to automate channel shuffling by learning permutation matrices in network training. We introduce an exact Lipschitz continuous non-convex penalty so that it can be incorporated in the stochastic gradient descent to approximate permutation at high precision. Exact permutations are obtained by simple rounding at the end of training and are used in inference. The resulting network, referred to as AutoShuffleNet, achieved improved classification accuracies on data from CIFAR-10, CIFAR-100 and ImageNet while preserving the inference costs of ShuffleNet. In addition, we found experimentally that the standard convex relaxation of permutation matrices into stochastic matrices leads to poor performance. We prove theoretically the exactness (error bounds) in recovering permutation matrices when our penalty function is zero (very small). We present examples of permutation optimization through graph matching and two-layer neural network models where the loss functions are calculated in closed analytical form. In the examples, convex relaxation failed to capture permutations whereas our penalty succeeded. 

We propose a multistage differentiable method to select convolutional channels and construct light neural networks from a heavy network for inference on a subset of a big data set. The selection proceeds backward in layers and utilizes sparse penalty to diversify channel scores. The resulting light network gains sizable accuracy over the baseline heavy network. 

A relaxed groupwise splitting method (RGSM) is developed and evaluated for channel pruning of deep neural net- works. Experiments with VGG-16 and ResNet-18 architectures on CIFAR-10/100 image data show that RGSM can achieve much higher channel sparsity than group Lasso method, while keeping comparable accuracy.