With increasing data and model complexities, the time required to train neural networks has become prohibitively large. To address the exponential rise in training time, users are turning to data parallel neural networks (DPNN) to utilize large-scale distributed resources on computer clusters. Current DPNN approaches implement the network parameter updates by synchronizing and averaging gradients across all processes with blocking communication operations. This synchronization is the central algorithmic bottleneck. To combat this, we introduce the Distributed Asynchronous and Selective Optimization (DASO) method which leverages multi-GPU compute node architectures to accelerate network training. DASO uses a hierarchical and asynchronous communication scheme comprised of node-local and global networks while adjusting the global synchronization rate during the learning process. We show that DASO yields a reduction in training time of up to 34% on classical and state-of-the-art networks, as compared to other existing data parallel training methods.
Large language models have led to state-of-the-art accuracies across a range of tasks. However, training these models efficiently is challenging for two reasons: a) GPU memory capacity is limited, making it impossible to fit large models on even a multi-GPU server, and b) the number of compute operations required to train these models can result in unrealistically long training times. Consequently, new methods of model parallelism such as tensor and pipeline parallelism have been proposed. Unfortunately, naive usage of these methods leads to fundamental scaling issues at thousands of GPUs, e.g., due to expensive cross-node communication or devices spending significant time waiting on other devices to make progress. In this paper, we show how different types of parallelism methods (tensor, pipeline, and data parallelism) can be composed to scale to thousands of GPUs and models with trillions of parameters. We survey techniques for pipeline parallelism and propose a novel interleaved pipeline parallelism schedule that can improve throughput by 10+% with memory footprint comparable to existing approaches. We quantitatively study the trade-offs between tensor, pipeline, and data parallelism, and provide intuition as to how to configure distributed training of a large model. Our approach allows us to perform training iterations on a model with 1 trillion parameters at 502 petaFLOP/s on 3072 GPUs with achieved per-GPU throughput of 52% of theoretical peak. Our code is open sourced at https://github.com/nvidia/megatron-lm.
Data augmentation has been an indispensable tool to improve the performance of deep neural networks, however the augmentation can hardly transfer among different tasks and datasets. Consequently, a recent trend is to adopt AutoML technique to learn proper augmentation policy without extensive hand-crafted tuning. In this paper, we propose an efficient differentiable search algorithm called Direct Differentiable Augmentation Search (DDAS). It exploits meta-learning with one-step gradient update and continuous relaxation to the expected training loss for efficient search. Our DDAS can achieve efficient augmentation search without relying on approximations such as Gumbel Softmax or second order gradient approximation. To further reduce the adverse effect of improper augmentations, we organize the search space into a two level hierarchy, in which we first decide whether to apply augmentation, and then determine the specific augmentation policy. On standard image classification benchmarks, our DDAS achieves state-of-the-art performance and efficiency tradeoff while reducing the search cost dramatically, e.g. 0.15 GPU hours for CIFAR-10. In addition, we also use DDAS to search augmentation for object detection task and achieve comparable performance with AutoAugment, while being 1000x faster.
Making inferences with a deep neural network on a batch of states is much faster with a GPU than making inferences on one state after another. We build on this property to propose Monte Carlo Tree Search algorithms using batched inferences. Instead of using either a search tree or a transposition table we propose to use both in the same algorithm. The transposition table contains the results of the inferences while the search tree contains the statistics of Monte Carlo Tree Search. We also propose to analyze multiple heuristics that improve the search: the $\mu$ FPU, the Virtual Mean, the Last Iteration and the Second Move heuristics. They are evaluated for the game of Go using a MobileNet neural network.
This paper introduces a conditional generative adversarial network to redesign a street-level image of urban scenes by generating 1) an urban intervention policy, 2) an attention map that localises where intervention is needed, 3) a high-resolution street-level image (1024 X 1024 or 1536 X1536) after implementing the intervention. We also introduce a new dataset that comprises aligned street-level images of before and after urban interventions from real-life scenarios that make this research possible. The introduced method has been trained on different ranges of urban interventions applied to realistic images. The trained model shows strong performance in re-modelling cities, outperforming existing methods that apply image-to-image translation in other domains that is computed in a single GPU. This research opens the door for machine intelligence to play a role in re-thinking and re-designing the different attributes of cities based on adversarial learning, going beyond the mainstream of facial landmarks manipulation or image synthesis from semantic segmentation.
The objective of this work is to segment high-resolution images without overloading GPU memory usage or losing the fine details in the output segmentation map. The memory constraint means that we must either downsample the big image or divide the image into local patches for separate processing. However, the former approach would lose the fine details, while the latter can be ambiguous due to the lack of a global picture. In this work, we present MagNet, a multi-scale framework that resolves local ambiguity by looking at the image at multiple magnification levels. MagNet has multiple processing stages, where each stage corresponds to a magnification level, and the output of one stage is fed into the next stage for coarse-to-fine information propagation. Each stage analyzes the image at a higher resolution than the previous stage, recovering the previously lost details due to the lossy downsampling step, and the segmentation output is progressively refined through the processing stages. Experiments on three high-resolution datasets of urban views, aerial scenes, and medical images show that MagNet consistently outperforms the state-of-the-art methods by a significant margin.
This paper reports the first successful application of a differentiable architecture search (DARTS) approach to the deepfake and spoofing detection problems. An example of neural architecture search, DARTS operates upon a continuous, differentiable search space which enables both the architecture and parameters to be optimised via gradient descent. Solutions based on partially-connected DARTS use random channel masking in the search space to reduce GPU time and automatically learn and optimise complex neural architectures composed of convolutional operations and residual blocks. Despite being learned quickly with little human effort, the resulting networks are competitive with the best performing systems reported in the literature. Some are also far less complex, containing 85% fewer parameters than a Res2Net competitor.
In recent years, an enormous amount of fluorescence microscopy images were collected in high-throughput lab settings. Analyzing and extracting relevant information from all images in a short time is almost impossible. Detecting tiny individual cell compartments is one of many challenges faced by biologists. This paper aims at solving this problem by building an end-to-end process that employs methods from the deep learning field to automatically segment, detect and classify cell compartments of fluorescence microscopy images of yeast cells. With this intention we used Mask R-CNN to automatically segment and label a large amount of yeast cell data, and YOLOv4 to automatically detect and classify individual yeast cell compartments from these images. This fully automated end-to-end process is intended to be integrated into an interactive e-Science server in the PerICo1 project, which can be used by biologists with minimized human effort in training and operation to complete their various classification tasks. In addition, we evaluated the detection and classification performance of state-of-the-art YOLOv4 on data from the NOP1pr-GFP-SWAT yeast-cell data library. Experimental results show that by dividing original images into 4 quadrants YOLOv4 outputs good detection and classification results with an F1-score of 98% in terms of accuracy and speed, which is optimally suited for the native resolution of the microscope and current GPU memory sizes. Although the application domain is optical microscopy in yeast cells, the method is also applicable to multiple-cell images in medical applications
As GPUs scale their low precision matrix math throughput to boost deep learning (DL) performance, they upset the balance between math throughput and memory system capabilities. We demonstrate that converged GPU design trying to address diverging architectural requirements between FP32 (or larger) based HPC and FP16 (or smaller) based DL workloads results in sub-optimal configuration for either of the application domains. We argue that a Composable On-PAckage GPU (COPAGPU) architecture to provide domain-specialized GPU products is the most practical solution to these diverging requirements. A COPA-GPU leverages multi-chip-module disaggregation to support maximal design reuse, along with memory system specialization per application domain. We show how a COPA-GPU enables DL-specialized products by modular augmentation of the baseline GPU architecture with up to 4x higher off-die bandwidth, 32x larger on-package cache, 2.3x higher DRAM bandwidth and capacity, while conveniently supporting scaled-down HPC-oriented designs. This work explores the microarchitectural design necessary to enable composable GPUs and evaluates the benefits composability can provide to HPC, DL training, and DL inference. We show that when compared to a converged GPU design, a DL-optimized COPA-GPU featuring a combination of 16x larger cache capacity and 1.6x higher DRAM bandwidth scales per-GPU training and inference performance by 31% and 35% respectively and reduces the number of GPU instances by 50% in scale-out training scenarios.
We introduce the IBM Analog Hardware Acceleration Kit, a new and first of a kind open source toolkit to simulate analog crossbar arrays in a convenient fashion from within PyTorch (freely available at https://github.com/IBM/aihwkit). The toolkit is under active development and is centered around the concept of an "analog tile" which captures the computations performed on a crossbar array. Analog tiles are building blocks that can be used to extend existing network modules with analog components and compose arbitrary artificial neural networks (ANNs) using the flexibility of the PyTorch framework. Analog tiles can be conveniently configured to emulate a plethora of different analog hardware characteristics and their non-idealities, such as device-to-device and cycle-to-cycle variations, resistive device response curves, and weight and output noise. Additionally, the toolkit makes it possible to design custom unit cell configurations and to use advanced analog optimization algorithms such as Tiki-Taka. Moreover, the backward and update behavior can be set to "ideal" to enable hardware-aware training features for chips that target inference acceleration only. To evaluate the inference accuracy of such chips over time, we provide statistical programming noise and drift models calibrated on phase-change memory hardware. Our new toolkit is fully GPU accelerated and can be used to conveniently estimate the impact of material properties and non-idealities of future analog technology on the accuracy for arbitrary ANNs.
Video compression is a critical component of Internet video delivery. Recent work has shown that deep learning techniques can rival or outperform human-designed algorithms, but these methods are significantly less compute and power-efficient than existing codecs. This paper presents a new approach that augments existing codecs with a small, content-adaptive super-resolution model that significantly boosts video quality. Our method, SRVC, encodes video into two bitstreams: (i) a content stream, produced by compressing downsampled low-resolution video with the existing codec, (ii) a model stream, which encodes periodic updates to a lightweight super-resolution neural network customized for short segments of the video. SRVC decodes the video by passing the decompressed low-resolution video frames through the (time-varying) super-resolution model to reconstruct high-resolution video frames. Our results show that to achieve the same PSNR, SRVC requires 16% of the bits-per-pixel of H.265 in slow mode, and 2% of the bits-per-pixel of DVC, a recent deep learning-based video compression scheme. SRVC runs at 90 frames per second on a NVIDIA V100 GPU.
We present a novel conditional Generative Adversarial Network (cGAN) architecture that is capable of generating 3D Computed Tomography scans in voxels from noisy and/or pixelated approximations and with the potential to generate full synthetic 3D scan volumes. We believe conditional cGAN to be a tractable approach to generate 3D CT volumes, even though the problem of generating full resolution deep fakes is presently impractical due to GPU memory limitations. We present results for autoencoder, denoising, and depixelating tasks which are trained and tested on two novel COVID19 CT datasets. Our evaluation metrics, Peak Signal to Noise ratio (PSNR) range from 12.53 - 46.46 dB, and the Structural Similarity index ( SSIM) range from 0.89 to 1.
Predictive energy management of Connected and Automated Vehicles (CAVs), in particular those with multiple power sources, has the potential to significantly improve energy savings in real-world driving conditions. In particular, the eco-driving problem seeks to design optimal speed and power usage profiles based upon available information from connectivity and advanced mapping features to minimize the fuel consumption between two designated locations. In this work, the eco-driving problem is formulated as a three-state receding horizon optimal control problem and solved via Dynamic Programming (DP). The optimal solution, in terms of vehicle speed and battery State of Charge (SoC) trajectories, allows a connected and automated hybrid electric vehicle to intelligently pass the signalized intersections and minimize fuel consumption over a prescribed route. To enable real-time implementation, a parallel architecture of DP is proposed for an NVIDIA GPU with CUDA programming. Simulation results indicate that the proposed optimal controller delivers more than 15% fuel economy benefits compared to a baseline control strategy and that the solver time can be reduced by more than 90% by the parallel implementation when compared to a serial implementation.
The importance of hierarchical image organization has been witnessed by a wide spectrum of applications in computer vision and graphics. Different from image segmentation with the spatial whole-part consideration, this work designs a modern framework for disassembling an image into a family of derived signals from a scale-space perspective. Specifically, we first offer a formal definition of image disassembly. Then, by concerning desired properties, such as peeling hierarchy and structure preservation, we convert the original complex problem into a series of two-component separation sub-problems, significantly reducing the complexity. The proposed framework is flexible to both supervised and unsupervised settings. A compact recurrent network, namely hierarchical image peeling net, is customized to efficiently and effectively fulfill the task, which is about 3.5Mb in size, and can handle 1080p images in more than 60 fps per recurrence on a GTX 2080Ti GPU, making it attractive for practical use. Both theoretical findings and experimental results are provided to demonstrate the efficacy of the proposed framework, reveal its superiority over other state-of-the-art alternatives, and show its potential to various applicable scenarios. Our code is available at \url{https://github.com/ForawardStar/HIPe}.
Although neural text-to-speech (TTS) models have attracted a lot of attention and succeeded in generating human-like speech, there is still room for improvements to its naturalness and architectural efficiency. In this work, we propose a novel non-autoregressive TTS model, namely Diff-TTS, which achieves highly natural and efficient speech synthesis. Given the text, Diff-TTS exploits a denoising diffusion framework to transform the noise signal into a mel-spectrogram via diffusion time steps. In order to learn the mel-spectrogram distribution conditioned on the text, we present a likelihood-based optimization method for TTS. Furthermore, to boost up the inference speed, we leverage the accelerated sampling method that allows Diff-TTS to generate raw waveforms much faster without significantly degrading perceptual quality. Through experiments, we verified that Diff-TTS generates 28 times faster than the real-time with a single NVIDIA 2080Ti GPU.
In the field of complex action recognition in videos, the quality of the designed model plays a crucial role in the final performance. However, artificially designed network structures often rely heavily on the researchers' knowledge and experience. Accordingly, because of the automated design of its network structure, Neural architecture search (NAS) has achieved great success in the image processing field and attracted substantial research attention in recent years. Although some NAS methods have reduced the number of GPU search days required to single digits in the image field, directly using 3D convolution to extend NAS to the video field is still likely to produce a surge in computing volume. To address this challenge, we propose a new processing framework called Neural Architecture Search- Temporal Convolutional (NAS-TC). Our proposed framework is divided into two phases. In the first phase, the classical CNN network is used as the backbone network to complete the computationally intensive feature extraction task. In the second stage, a simple stitching search to the cell is used to complete the relatively lightweight long-range temporal-dependent information extraction. This ensures our method will have more reasonable parameter assignments and can handle minute-level videos. Finally, we conduct sufficient experiments on multiple benchmark datasets and obtain competitive recognition accuracy.
Robust multi-agent trajectory prediction is essential for the safe control of robotic systems. A major challenge is to efficiently learn a representation that approximates the true joint distribution of contextual, social, and temporal information to enable planning. We propose Latent Variable Sequential Set Transformers which are encoder-decoder architectures that generate scene-consistent multi-agent trajectories. We refer to these architectures as "AutoBots". The encoder is a stack of interleaved temporal and social multi-head self-attention (MHSA) modules which alternately perform equivariant processing across the temporal and social dimensions. The decoder employs learnable seed parameters in combination with temporal and social MHSA modules allowing it to perform inference over the entire future scene in a single forward pass efficiently. AutoBots can produce either the trajectory of one ego-agent or a distribution over the future trajectories for all agents in the scene. For the single-agent prediction case, our model achieves top results on the global nuScenes vehicle motion prediction leaderboard, and produces strong results on the Argoverse vehicle prediction challenge. In the multi-agent setting, we evaluate on the synthetic partition of TrajNet++ dataset to showcase the model's socially-consistent predictions. We also demonstrate our model on general sequences of sets and provide illustrative experiments modelling the sequential structure of the multiple strokes that make up symbols in the Omniglot data. A distinguishing feature of AutoBots is that all models are trainable on a single desktop GPU (1080 Ti) in under 48h.
The training of deep neural networks (DNNs) is usually memory-hungry due to the limited device memory capacity of DNN accelerators. Characterizing the memory behaviors of DNN training is critical to optimize the device memory pressures. In this work, we pinpoint the memory behaviors of each device memory block of GPU during training by instrumenting the memory allocators of the runtime system. Our results show that the memory access patterns of device memory blocks are stable and follow an iterative fashion. These observations are useful for the future optimization of memory-efficient training from the perspective of raw memory access patterns.
We propose an end-to-end speech synthesizer, Fast DCTTS, that synthesizes speech in real time on a single CPU thread. The proposed model is composed of a carefully-tuned lightweight network designed by applying multiple network reduction and fidelity improvement techniques. In addition, we propose a novel group highway activation that can compromise between computational efficiency and the regularization effect of the gating mechanism. As well, we introduce a new metric called Elastic mel-cepstral distortion (EMCD) to measure the fidelity of the output mel-spectrogram. In experiments, we analyze the effect of the acceleration techniques on speed and speech quality. Compared with the baseline model, the proposed model exhibits improved MOS from 2.62 to 2.74 with only 1.76% computation and 2.75% parameters. The speed on a single CPU thread was improved by 7.45 times, which is fast enough to produce mel-spectrogram in real time without GPU.
End-to-end neural network models achieve improved performance on various automatic speech recognition (ASR) tasks. However, these models perform poorly on edge hardware due to large memory and computation requirements. While quantizing model weights and/or activations to low-precision can be a promising solution, previous research on quantizing ASR models is limited. In particular, the previous approaches use floating-point arithmetic during inference and thus they cannot fully exploit efficient integer processing units. Moreover, they require training and/or validation data during quantization, which may not be available due to security or privacy concerns. To address these limitations, we propose an integer-only, zero-shot quantization scheme for ASR models. In particular, we generate synthetic data whose runtime statistics resemble the real data, and we use it to calibrate models during quantization. We apply our method to quantize QuartzNet, Jasper, and Conformer and show negligible WER degradation as compared to the full-precision baseline models, even without using any data. Moreover, we achieve up to 2.35x speedup on a T4 GPU and 4x compression rate, with a modest WER degradation of <1% with INT8 quantization.