Various parallelism, such as data, tensor, and pipeline parallelism, along with memory optimizations like activation checkpointing, redundancy elimination, and offloading, have been proposed to accelerate distributed training for Large Language Models. To find the best combination of these techniques, automatic distributed training systems are proposed. However, existing systems only tune a subset of optimizations, due to the lack of overlap awareness, inability to navigate the vast search space, and ignoring the inter-microbatch imbalance, leading to sub-optimal performance. To address these shortcomings, we propose Mist, a memory, overlap, and imbalance-aware automatic distributed training system that comprehensively co-optimizes all memory footprint reduction techniques alongside parallelism. Mist is based on three key ideas: (1) fine-grained overlap-centric scheduling, orchestrating optimizations in an overlapped manner, (2) symbolic-based performance analysis that predicts runtime and memory usage using symbolic expressions for fast tuning, and (3) imbalance-aware hierarchical tuning, decoupling the process into an inter-stage imbalance and overlap aware Mixed Integer Linear Programming problem and an intra-stage Dual-Objective Constrained Optimization problem, and connecting them through Pareto frontier sampling. Our evaluation results show that Mist achieves an average of 1.28× (up to 1.73×) and 1.27× (up to 2.04×) speedup compared to state-of-the-art manual system Megatron-LM and state-of-the-art automatic system Aceso, respectively.

We introduce Ego4D, a massive-scale egocentric video dataset and benchmark suite. It offers 3,670 hours of daily-life activity video spanning hundreds of scenarios (household, outdoor, workplace, leisure, etc.) captured by 931 unique camera wearers from 74 worldwide locations and 9 different countries. The approach to collection is designed to uphold rigorous privacy and ethics standards, with consenting participants and robust de-identification procedures where relevant. Ego4D dramatically expands the volume of diverse egocentric video footage publicly available to the research community. Portions of the video are accompanied by audio, 3D meshes of the environment, eye gaze, stereo, and/or synchronized videos from multiple egocentric cameras at the same event. Furthermore, we present a host of new benchmark challenges centered around understanding the first-person visual experience in the past (querying an episodic memory), present (analyzing hand-object manipulation, audio-visual conversation, and social interactions), and future (forecasting activities). By publicly sharing this massive annotated dataset and benchmark suite, we aim to push the frontier of first-person perception.

Large pretrained models are increasingly crucial in modern computer vision tasks. These models are typically used in downstream tasks by end-to-end finetuning, which is highly memory-intensive for tasks with high-resolution data, e.g., video understanding, small object detection, and point cloud analysis. In this paper, we propose Dynamic Reversible Dual-Residual Networks, or Dr2Net, a novel family of network architectures that acts as a surrogate net-work to finetune a pretrained model with substantially re-duced memory consumption. Dr2 Net contains two types of residual connections, one maintaining the residual struc-ture in the pretrained models, and the other making the network reversible. Due to its reversibility, intermediate activations, which can be reconstructed from output, are cleared from memory during training. We use two co-efficients on either type of residual connections respec-tively, and introduce a dynamic training strategy that seam-lessly transitions the pretrained model to a reversible net-work with much higher numerical precision. We evaluate Dr2Net on various pretrained models and various tasks, and show that it can reach comparable performance to con-ventional finetuning but with significantly less memory us-age. Code will be available at https://github.com/coolbay/Dr2Net.

We introduce a novel sequential modeling approach which enables learning a Large Vision Model (LVM) without making use of any linguistic data. To do this, we define a common format, “visual sentences ”, in which we can represent raw images and videos as well as annotated data sources such as semantic segmentations and depth reconstructions with-out needing any meta-knowledge beyond the pixels. Once this wide variety of visual data (comprising 420 billion to-kens) is represented as sequences, the model can be trained to minimize a cross-entropy loss for next token prediction. By training across various scales of model architecture and data diversity, we provide empirical evidence that our models scale effectively. Many different vision tasks can be solved by designing suitable visual prompts at test time.

Aligned text-image encoders such as CLIP have become the de-facto model for vision-language tasks. Further-more, modality-specific encoders achieve impressive per-formances in their respective domains. This raises a cen-tral question: does an alignment exist between uni-modal vision and language encoders since they fundamentally rep-resent the same physical world? Analyzing the latent spaces structure of vision and language models on image-caption benchmarks using the Centered Kernel Alignment (CKA), we find that the representation spaces of unaligned and aligned encoders are semantically similar. In the absence of statistical similarity in aligned encoders like CLIP, we show that a possible matching of unaligned encoders exists with-out any training. We frame this as a seeded graph-matching problem exploiting the semantic similarity between graphs and propose two methods - a Fast Quadratic Assignment Problem optimization, and a novel localized CKA metric-based matching/retrieval. We demonstrate the effectiveness of this on several downstream tasks including cross-lingual, cross-domain caption matching and image classification. Code available at github.com/mayug/0-shot-llm-vision.

Modern computer vision pipelines handle large images in one of two sub-optimal ways: down-sampling or cropping. These two methods incur significant losses in the amount of information and context present in an image. There are many downstream applications in which global context matters as much as high frequency details, such as in real-world satellite imagery; in such cases researchers have to make the uncomfortable choice of which information to discard. We introduce *xT*, a simple framework for vision transformers which effectively aggregates global context with local details and can model large images end-to-end on contemporary GPUs. We select a set of benchmark datasets across classic vision tasks which accurately reflect a vision model's ability to understand truly large images and incorporate fine details over large scales and assess our method's improvement on them. *xT* is a streaming, two-stage architecture that adapts existing vision backbones and long sequence language models to effectively model large images without quadratic memory growth. We are able to increase accuracy by up to 8.6% on challenging classification tasks and F1 score by 11.6 on context-dependent segmentation on images as large as 29,000 x 29,000 pixels.

Temporal action localization (TAL) requires long-form reasoning to predict actions of various durations and complex content. Given limited GPU memory, training TAL end to end (i.e., from videos to predictions) on long videos is a significant challenge. Most methods can only train on pre-extracted features without optimizing them for the localization problem, consequently limiting localization performance. In this work, to extend the potential in TAL networks, we propose a novel end-to-end method Re2TAL, which rewires pretrained video backbones for reversible TAL. Re2TAL builds a backbone with reversible modules, where the input can be recovered from the output such that the bulky intermediate activations can be cleared from memory during training. Instead of designing one single type of reversible module, we propose a network rewiring mechanism, to transform any module with a residual connection to a reversible module without changing any parameters. This provides two benefits: (1) a large variety of reversible networks are easily obtained from existing and even future model designs, and (2) the reversible models require much less training effort as they reuse the pre-trained parameters of their original non-reversible versions. Re2TAL, only using the RGB modality, reaches 37.01% average mAP on ActivityNet-v1.3, a new state-of-the-art record, and mAP 64.9% at tIoU=0.5 on THUMOS-14, outperforming all other RGB-only methods. Code is available at https://github.com/coolbay/Re2TAL.

We present RAFTformer, a realtime action forecasting transformer for latency-aware realworld action forecasting. RAFTformer is a two-stage fully transformer based architecture comprising of a video transformer backbone that operates on high resolution, short-range clips, and a head transformer encoder that temporally aggregates information from multiple short-range clips to span a long-term horizon. Additionally, we propose a novel self-supervised shuffled causal masking scheme as a model level augmentation to improve forecasting fidelity. Finally, we also propose a novel realtime evaluation setting for action fore-casting that directly couples model inference latency to overall forecasting performance and brings forth a hith-erto overlooked trade-off between latency and action fore-casting performance. Our parsimonious network design fa-cilitates RAFTformer inference latency to be 9x smaller than prior works at the same forecasting accuracy. Owing to its two-staged design, RAFTformer uses 94% less training compute and 90% lesser training parameters to outperform prior state-of-the-art baselines by 4.9 points on EGTEA Gaze+ and by 1.4 points on EPIC-Kitchens-100 validation set, as measured by Top-5 recall (T5R) in the offline setting. In the realtime setting, RAFTformer outperforms prior works by an even greater margin of upto 4.4 T5R points on the EPIC-Kitchens-100 dataset. Project Webpage: https://karttikeya.github.io/publication/RAFTformer/.

There has been a longstanding belief that generation can facilitate a true understanding of visual data. In line with this, we revisit generatively pre-training visual representations in light of recent interest in denoising diffusion models. While directly pre-training with diffusion models does not produce strong representations, we condition diffusion models on masked input and formulate diffusion models as masked autoencoders (DiffMAE). Our approach is capable of (i) serving as a strong initialization for downstream recognition tasks, (ii) conducting high-quality image inpainting, and (iii) being effortlessly extended to video where it produces state-of-the-art classification accuracy. We further perform a comprehensive study on the pros and cons of design choices and build connections between diffusion models and masked autoencoders. Project page.

The recent emergence of Large Language Models based on the Transformer architecture has enabled dramatic advancements in the field of Natural Language Processing. However, these models have long inference latency, which limits their deployment and makes them prohibitively expensive for various real-time applications. The inference latency is further exacerbated by autoregressive generative tasks, as models need to run iteratively to generate tokens sequentially without leveraging token-level parallelization. To address this, we propose Big Little Decoder (BiLD), a framework that can improve inference efficiency and latency for a wide range of text generation applications. The BiLD framework contains two models with different sizes that collaboratively generate text. The small model runs autoregressively to generate text with a low inference cost, and the large model is only invoked occasionally to refine the small model’s inaccurate predictions in a non-autoregressive manner. To coordinate the small and large models, BiLD introduces two simple yet effective policies: (1) the fallback policy that determines when to hand control over to the large model; and (2) the rollback policy that determines when the large model needs to correct the small model's inaccurate predictions. To evaluate our framework across different tasks and models, we apply BiLD to various text generation scenarios encompassing machine translation on IWSLT 2017 De-En and WMT 2014 De-En, and summarization on XSUM and CNN/DailyMail. On an NVIDIA T4 GPU, our framework achieves a speedup of up to 2.12x speedup with minimal generation quality degradation. Furthermore, our framework is fully plug-and-play and can be applied without any modifications in the training process or model architecture. Our code is open-sourced.

We introduce EgoSchema, a very long-form video question-answering dataset, and benchmark to evaluate long video understanding capabilities of modern vision and language systems. Derived from Ego4D, EgoSchema consists of over 5000 human curated multiple choice question answer pairs, spanning over 250 hours of real video data, covering a very broad range of natural human activity and behavior. For each question, EgoSchema requires the correct answer to be selected between five given options based on a three-minute-long video clip. While some prior works have proposed video datasets with long clip lengths, we posit that merely the length of the video clip does not truly capture the temporal difficulty of the video task that is being considered. To remedy this, we introduce temporal certificate sets, a general notion for capturing the intrinsic temporal understanding length associated with a broad range of video understanding tasks & datasets. Based on this metric, we find EgoSchema to have intrinsic temporal lengths over 5.7x longer than the second closest dataset and 10x to 100x longer than any other video understanding dataset. Further, our evaluation of several current state-of-the-art video and language models shows them to be severely lacking in long-term video understanding capabilities. Even models with several billions of parameters achieve QA accuracy less than 33% (random is 20%) on the EgoSchema multi-choice question answering task, while humans achieve about 76% accuracy. We posit that EgoSchema, with its long intrinsic temporal structures and diverse complexity, would serve as a valuable evaluation probe for developing effective long-term video understanding systems in the future. Data and Zero-shot model evaluation code will all be open-sourced under the Ego4D license at http://egoschema.github.io.

We introduce Ego4D, a massive-scale egocentric video dataset and benchmark suite. It offers 3,670 hours of dailylife activity video spanning hundreds of scenarios (household, outdoor, workplace, leisure, etc.) captured by 931 unique camera wearers from 74 worldwide locations and 9 different countries. The approach to collection is designed to uphold rigorous privacy and ethics standards, with consenting participants and robust de-identification procedures where relevant. Ego4D dramatically expands the volume of diverse egocentric video footage publicly available to the research community. Portions of the video are accompanied by audio, 3D meshes of the environment, eye gaze, stereo, and/or synchronized videos from multiple egocentric cameras at the same event. Furthermore, we present a host of new benchmark challenges centered around understanding the first-person visual experience in the past (querying an episodic memory), present (analyzing hand-object manipulation, audio-visual conversation, and social interactions), and future (forecasting activities). By publicly sharing this massive annotated dataset and benchmark suite, we aim to push the frontier of first-person perception. Project page: https://ego4d-data.org/

Recently, video transformers have shown great success in video understanding, exceeding CNN performance; yet existing video transformer models do not explicitly model objects, although objects can be essential for recognizing actions. In this work, we present Object-Region Video Transformers (ORViT), an object-centric approach that extends video transformer layers with a block that directly incorporates object representations. The key idea is to fuse object-centric representations starting from early layers and propagate them into the transformer-layers, thus affecting the spatio-temporal representations throughout the network. Our ORViT block consists of two object-level streams: appearance and dynamics. In the appearance stream, an “Object-Region Attention” module applies self-attention over the patches and object regions. In this way, visual object regions interact with uniform patch tokens and enrich them with contextualized object information. We further model object dynamics via a separate “Object-Dynamics Module”, which captures trajectory interactions, and show how to integrate the two streams. We evaluate our model on four tasks and five datasets: compositional and few-shot action recognition on SomethingElse, spatio-temporal action detection on AVA, and standard action recognition on Something-Something V2, Diving48 and Epic-Kitchen100. We show strong performance improvement across all tasks and datasets considered, demonstrating the value of a model that incorporates object representations into a transformer architecture. For code and pretrained models, visit the project page at https://roeiherz.github.io/ORViT/

In this paper, we study Multiscale Vision Transformers (MViTv2) as a unified architecture for image and video classification, as well as object detection. We present an improved version of MViT that incorporates decomposed relative positional embeddings and residual pooling connections. We instantiate this architecture in five sizes and evaluate it for ImageNet classification, COCO detection and Kinetics video recognition where it outperforms prior work. We further compare MViTv2s' pooling attention to window attention mechanisms where it outperforms the latter in accuracy/compute. Without bells-and-whistles, MViTv2 has state-of-the-art performance in 3 domains: 88.8% accuracy on ImageNet classification, 58.7 APboxon COCO object detection as well as 86.1% on Kinetics-400 video classification. Code and models are available at https://github.com/facebookresearch/mvit.

We present Reversible Vision Transformers, a memory efficient architecture design for visual recognition. By decoupling the GPU memory footprint from the depth of the model, Reversible Vision Transformers enable memory efficient scaling of transformer architectures. We adapt two popular models, namely Vision Transformer and Multiscale Vision Transformers, to reversible variants and benchmark extensively across both model sizes and tasks of image classification, object detection and video classification. Reversible Vision Transformers achieve a reduced memory footprint of up to 15.5× at identical model complexity, parameters and accuracy, demonstrating the promise of reversible vision transformers as an efficient backbone for resource limited training regimes. Finally, we find that the additional computational burden of recomputing activations is more than overcome for deeper models, where throughput can increase up to 3.9 × over their non-reversible counterparts. Code and models are available at https://github.com/facebookresearch/mvit.

While today's video recognition systems parse snapshots or short clips accurately, they cannot connect the dots and reason across a longer range of time yet. Most existing video architectures can only process3,000% more compute to do the same. On a wide range of settings, the increased temporal support enabled by MeMViT brings large gains in recognition accuracy consistently. MeMViT obtains state-of-the-art results on the AVA, EPIC-Kitchens-100 action classification, and action anticipation datasets. Code and models will be made publicly available.

Recent action recognition models have achieved impressive results by integrating objects, their locations and interactions. However, obtaining dense structured annotations for each frame is tedious and time-consuming, making these methods expensive to train and less scalable. At the same time, if a small set of annotated images is available, either within or outside the domain of interest, how could we leverage these for a video downstream task? We propose a learning framework StructureViT (SViT for short), which demonstrates how utilizing the structure of a small number of images only available during training can improve a video model. SViT relies on two key insights. First, as both images and videos contain structured information, we enrich a transformer model with a set of object tokens that can be used across images and videos. Second, the scene representations of individual frames in video should ``align'' with those of still images. This is achieved via a Frame-Clip Consistency loss, which ensures the flow of structured information between images and videos. We explore a particular instantiation of scene structure, namely a Hand-Object Graph, consisting of hands and objects with their locations as nodes, and physical relations of contact/no-contact as edges. SViT shows strong performance improvements on multiple video understanding tasks and datasets, including the first place in the Ego4D CVPR'22 Point of No Return Temporal Localization Challenge. For code and pretrained models, visit the project page at https://eladb3.github.io/SViT/.

The recently proposed Conformer model has become the de facto backbone model for various downstream speech tasks based on its hybrid attention-convolution architecture that captures both local and global features. However, through a series of systematic studies, we find that the Conformer architecture’s design choices are not optimal. After re-examining the design choices for both the macro and micro-architecture of Conformer, we propose Squeezeformer which consistently outperforms the state-of-the-art ASR models under the same training schemes. In particular, for the macro-architecture, Squeezeformer incorporates (i) the Temporal U-Net structure which reduces the cost of the multi-head attention modules on long sequences, and (ii) a simpler block structure of multi-head attention or convolution modules followed up by feed-forward module instead of the Macaron structure proposed in Conformer. Furthermore, for the micro-architecture, Squeezeformer (i) simplifies the activations in the convolutional block, (ii) removes redundant Layer Normalization operations, and (iii) incorporates an efficient depthwise down-sampling layer to efficiently sub-sample the input signal. Squeezeformer achieves state-of-the-art results of 7.5%, 6.5%, and 6.0% word-error-rate (WER) on LibriSpeech test-other without external language models, which are 3.1%, 1.4%, and 0.6% better than Conformer-CTC with the same number of FLOPs. Our code is open-sourced and available online.

We present Multiscale Vision Transformers (MViT) for video and image recognition, by connecting the seminal idea of multiscale feature hierarchies with transformer models. Multiscale Transformers have several channel-resolution scale stages. Starting from the input resolution and a small channel dimension, the stages hierarchically expand the channel capacity while reducing the spatial resolution. This creates a multiscale pyramid of features with early layers operating at high spatial resolution to model simple low-level visual information, and deeper layers at spatially coarse, but complex, high-dimensional features. We evaluate this fundamental architectural prior for modeling the dense nature of visual signals for a variety of video recognition tasks where it outperforms concurrent vision transformers that rely on large scale external pre-training and are 5-10× more costly in computation and parameters. We further remove the temporal dimension and apply our model for image classification where it outperforms prior work on vision transformers. Code is available at: https://github.com/facebookresearch/SlowFast.

Recent advances in trajectory prediction have shown that explicit reasoning about agents’ intent is important to accurately forecast their motion. However, the current research activities are not directly applicable to intelligent and safety critical systems. This is mainly because very few public datasets are available, and they only consider pedestrian-specific intents for a short temporal horizon from a restricted egocentric view. To this end, we propose LOKI (LOng term and Key Intentions), a novel large-scale dataset that is designed to tackle joint trajectory and intention prediction for heterogeneous traffic agents (pedestrians and vehicles) in an autonomous driving setting. The LOKI dataset is created to discover several factors that may affect intention, including i) agent’s own will, ii) social interactions, iii) environmental constraints, and iv) contextual information. We also propose a model that jointly performs trajectory and intention prediction, showing that recurrently reasoning about intention can assist with trajectory prediction. We show our method outperforms state-of-the-art trajectory prediction methods by upto 27% and also provide a baseline for frame-wise intention estimation. The dataset is available at https://usa.honda-ri.com/loki

Human trajectory forecasting is an inherently multi-modal problem. Uncertainty in future trajectories stems from two sources: (a) sources that are known to the agent but unknown to the model, such as long term goals and (b) sources that are unknown to both the agent and the model, such as the intent of other agents and irreducible randomness in decisions. We propose to factorize this uncertainty into its epistemic and aleatoric sources. We model the epistemic uncertainty through multimodality in long term goals and the aleatoric uncertainty through multimodality in way-points and paths. To exemplify this dichotomy, we also propose a novel long term trajectory forecasting setting, with prediction horizons up to a minute, up to an order of magnitude longer than prior works. Finally, we present Y-net, a scene compliant trajectory forecasting network that exploits the proposed epistemic and aleatoric structure for diverse trajectory predictions across long prediction horizons. Y-net significantly improves previous state-of-the-art performance on both (a) The short prediction horizon setting on the Stanford Drone (31.7% in FDE) and ETH/UCY datasets (7.4% in FDE) and (b) The proposed long horizon setting on the re-purposed Stanford Drone and Intersection Drone datasets.

We tackle the problem of Human Locomotion Forecasting, a task for jointly predicting the spatial positions of several keypoints on human body in the near future under an egocentric setting. In contrast to the previous work that aims to solve either the task of pose prediction or trajectory forecasting in isolation, we propose a framework to unify these two problems and address the practically useful task of pedestrian locomotion prediction in the wild. Among the major challenges in solving this task is the scarcity of annotated egocentric video datasets with dense annotations for pose, depth, or egomotion. To surmount this difficulty, we use state-of-the-art models to generate (noisy) annotations and propose robust forecasting models that can learn from this noisy supervision. We present a method to disentangle the overall pedestrian motion into easier to learn subparts by uti-lizing a pose completion and a decomposition module. The completion module fills in the missing key-point annotations and the decomposition module breaks the cleaned locomotion down to global (trajectory) and local (pose keypoint movements). Further, with Quasi RNN as our backbone, we propose a novel hierarchical trajectory forecasting network that utilizes low-level vision domain specific signals like egomotion and depth to predict the global trajectory. Our method leads to state-of-the-art results for the prediction of human locomotion in the egocentric view.

We present a new task that predicts future locations of people observed in first-person videos. Consider a first-person video stream continuously recorded by a wearable camera. Given a short clip of a person that is extracted from the complete stream, we aim to predict that person's location in future frames. To facilitate this future person localization ability, we make the following three key observations: (a) First-person videos typically involve significant ego-motion which greatly affects the location of the target person in future frames; (b) Scales of the target person act as a salient cue to estimate a perspective effect in first-person videos; (c) First-person videos often capture people up-close, making it easier to leverage target poses (e.g., where they look) for predicting their future locations. We incorporate these three observations into a prediction framework with a multi-stream convolution-deconvolution architecture. Experimental results reveal our method to be effective on our new dataset as well as on a public social interaction dataset.

We investigate the effect and usefulness of spontaneity (i.e. whether a given speech is spontaneous or not) in speech in the context of emotion recognition. We hypothesize that emotional content in speech is interrelated with its spontaneity, and use spontaneity classification as an auxiliary task to the problem of emotion recognition. We propose two supervised learning settings that utilize spontaneity to improve speech emotion recognition: a hierarchical model that performs spontaneity detection before performing emotion recognition, and a multitask learning model that jointly learns to recognize both spontaneity and emotion. Through various experiments on the well known IEMOCAP database, we show that by using spontaneity detection as an additional task, significant improvement can be achieved over emotion recognition systems that are unaware of spontaneity. We achieve state-of-the-art emotion recognition accuracy (4-class, 69.1%) on the IEMOCAP database outperforming several relevant and competitive baselines.