In this paper, we propose an open source, production first, and production ready speech recognition toolkit called WeNet in which a new two-pass approach is implemented to unify streaming and non-streaming end-to-end (E2E) speech recognition in a single model. The main motivation of WeNet is to close the gap between the research and the production of E2E speechrecognition models. WeNet provides an efficient way to ship ASR applications in several real-world scenarios, which is the main difference and advantage to other open source E2E speech recognition toolkits. In our toolkit, a new two-pass method is implemented. Our method propose a dynamic chunk-based attention strategy of the the transformer layers to allow arbitrary right context length modifies in hybrid CTC/attention architecture. The inference latency could be easily controlled by only changing the chunk size. The CTC hypotheses are then rescored by the attention decoder to get the final result. Our experiments on the AISHELL-1 dataset using WeNet show that, our model achieves 5.03\% relative character error rate (CER) reduction in non-streaming ASR compared to a standard non-streaming transformer. After model quantification, our model perform reasonable RTF and latency.
We introduce Generative Spoken Language Modeling, the task of learning the acoustic and linguistic characteristics of a language from raw audio (no text, no labels), and a set of metrics to automatically evaluate the learned representations at acoustic and linguistic levels for both encoding and generation. We set up baseline systems consisting of a discrete speech encoder (returning pseudo-text units), a generative language model (trained on pseudo-text), and a speech decoder (generating a waveform from pseudo-text) all trained without supervision and validate the proposed metrics with human evaluation. Across 3 speech encoders (CPC, wav2vec 2.0, HuBERT), we find that the number of discrete units (50, 100, or 200) matters in a task-dependent and encoder-dependent way, and that some combinations approach text-based systems.
Numerical weather forecasting using high-resolution physical models often requires extensive computational resources on supercomputers, which diminishes their wide usage in most real-life applications. As a remedy, applying deep learning methods has revealed innovative solutions within this field. To this end, we introduce a novel deep learning architecture for forecasting high-resolution spatio-temporal weather data. Our approach extends the conventional encoder-decoder structure by integrating Convolutional Long-short Term Memory and Convolutional Neural Networks. In addition, we incorporate attention and context matcher mechanisms into the model architecture. Our Weather Model achieves significant performance improvements compared to baseline deep learning models, including ConvLSTM, TrajGRU, and U-Net. Our experimental evaluation involves high-scale, real-world benchmark numerical weather datasets, namely the ERA5 hourly dataset on pressure levels and WeatherBench. Our results demonstrate substantial improvements in identifying spatial and temporal correlations with attention matrices focusing on distinct parts of the input series to model atmospheric circulations. We also compare our model with high-resolution physical models using the benchmark metrics and show that our Weather Model is accurate and easy to interpret.
Notwithstanding recent advances, syntactic generalization remains a challenge for text decoders. While some studies showed gains from incorporating source-side symbolic syntactic and semantic structure into text generation Transformers, very little work addressed the decoding of such structure. We propose a general approach for tree decoding using a transition-based approach. Examining the challenging test case of incorporating Universal Dependencies syntax into machine translation, we present substantial improvements on test sets that focus on syntactic generalization, while presenting improved or comparable performance on standard MT benchmarks. Further qualitative analysis addresses cases where syntactic generalization in the vanilla Transformer decoder is inadequate and demonstrates the advantages afforded by integrating syntactic information.
This paper introduces the unsupervised learning problem of playable video generation (PVG). In PVG, we aim at allowing a user to control the generated video by selecting a discrete action at every time step as when playing a video game. The difficulty of the task lies both in learning semantically consistent actions and in generating realistic videos conditioned on the user input. We propose a novel framework for PVG that is trained in a self-supervised manner on a large dataset of unlabelled videos. We employ an encoder-decoder architecture where the predicted action labels act as bottleneck. The network is constrained to learn a rich action space using, as main driving loss, a reconstruction loss on the generated video. We demonstrate the effectiveness of the proposed approach on several datasets with wide environment variety. Further details, code and examples are available on our project page willi-menapace.github.io/playable-video-generation-website.
We propose a novel hybrid approach to lemmatization that enhances the seq2seq neural model with additional lemmas extracted from an external lexicon or a rule-based system. During training, the enhanced lemmatizer learns both to generate lemmas via a sequential decoder and copy the lemma characters from the external candidates supplied during run-time. Our lemmatizer enhanced with candidates extracted from the Apertium morphological analyzer achieves statistically significant improvements compared to baseline models not utilizing additional lemma information, achieves an average accuracy of 97.25% on a set of 23 UD languages, which is 0.55% higher than obtained with the Stanford Stanza model on the same set of languages. We also compare with other methods of integrating external data into lemmatization and show that our enhanced system performs considerably better than a simple lexicon extension method based on the Stanza system, and it achieves complementary improvements w.r.t. the data augmentation method.
The COVID-19 pandemic continues to have major impact to health and medical infrastructure, economy, and agriculture. Prominent computational and mathematical models have been unreliable due to the complexity of the spread of infections. Moreover, lack of data collection and reporting makes modelling attempts difficult and unreliable. Hence, we need to re-look at the situation with reliable data sources and innovative forecasting models. Deep learning models such as recurrent neural networks are well suited for modelling spatiotemporal sequences. In this paper, we apply recurrent neural networks such as long short term memory (LSTM), bidirectional LSTM, and encoder-decoder LSTM models for multi-step (short-term) COVID-19 infection forecasting. We select Indian states with COVID-19 hotpots and capture the first (2020) and second (2021) wave of infections and provide two months ahead forecast. Our model predicts that the likelihood of another wave of infections in October and November 2021 is low; however, the authorities need to be vigilant given emerging variants of the virus. The accuracy of the predictions motivate the application of the method in other countries and regions. Nevertheless, the challenges in modelling remain due to the reliability of data and difficulties in capturing factors such as population density, logistics, and social aspects such as culture and lifestyle.
Despite having impressive vision-language (VL) pretraining with BERT-based encoder for VL understanding, the pretraining of a universal encoder-decoder for both VL understanding and generation remains challenging. The difficulty originates from the inherently different peculiarities of the two disciplines, e.g., VL understanding tasks capitalize on the unrestricted message passing across modalities, while generation tasks only employ visual-to-textual message passing. In this paper, we start with a two-stream decoupled design of encoder-decoder structure, in which two decoupled cross-modal encoder and decoder are involved to separately perform each type of proxy tasks, for simultaneous VL understanding and generation pretraining. Moreover, for VL pretraining, the dominant way is to replace some input visual/word tokens with mask tokens and enforce the multi-modal encoder/decoder to reconstruct the original tokens, but no mask token is involved when fine-tuning on downstream tasks. As an alternative, we propose a primary scheduled sampling strategy that elegantly mitigates such discrepancy via pretraining encoder-decoder in a two-pass manner. Extensive experiments demonstrate the compelling generalizability of our pretrained encoder-decoder by fine-tuning on four VL understanding and generation downstream tasks. Source code is available at \url{https://github.com/YehLi/TDEN}.
Paraphrase generation plays an essential role in natural language process (NLP), and it has many downstream applications. However, training supervised paraphrase models requires many annotated paraphrase pairs, which are usually costly to obtain. On the other hand, the paraphrases generated by existing unsupervised approaches are usually syntactically similar to the source sentences and are limited in diversity. In this paper, we demonstrate that it is possible to generate syntactically various paraphrases without the need for annotated paraphrase pairs. We propose Syntactically controlled Paraphrase Generator (SynPG), an encoder-decoder based model that learns to disentangle the semantics and the syntax of a sentence from a collection of unannotated texts. The disentanglement enables SynPG to control the syntax of output paraphrases by manipulating the embedding in the syntactic space. Extensive experiments using automatic metrics and human evaluation show that SynPG performs better syntactic control than unsupervised baselines, while the quality of the generated paraphrases is competitive. We also demonstrate that the performance of SynPG is competitive or even better than supervised models when the unannotated data is large. Finally, we show that the syntactically controlled paraphrases generated by SynPG can be utilized for data augmentation to improve the robustness of NLP models.
RGB-D salient object detection (SOD) recently has attracted increasing research interest and many deep learning methods based on encoder-decoder architectures have emerged. However, most existing RGB-D SOD models conduct feature fusion either in the single encoder or the decoder stage, which hardly guarantees sufficient cross-modal fusion ability. In this paper, we make the first attempt in addressing RGB-D SOD through 3D convolutional neural networks. The proposed model, named RD3D, aims at pre-fusion in the encoder stage and in-depth fusion in the decoder stage to effectively promote the full integration of RGB and depth streams. Specifically, RD3D first conducts pre-fusion across RGB and depth modalities through an inflated 3D encoder, and later provides in-depth feature fusion by designing a 3D decoder equipped with rich back-projection paths (RBPP) for leveraging the extensive aggregation ability of 3D convolutions. With such a progressive fusion strategy involving both the encoder and decoder, effective and thorough interaction between the two modalities can be exploited and boost the detection accuracy. Extensive experiments on six widely used benchmark datasets demonstrate that RD3D performs favorably against 14 state-of-the-art RGB-D SOD approaches in terms of four key evaluation metrics. Our code will be made publicly available: https://github.com/PPOLYpubki/RD3D.
In recent years, layered image compression is demonstrated to be a promising direction, which encodes a compact representation of the input image and apply an up-sampling network to reconstruct the image. To further improve the quality of the reconstructed image, some works transmit the semantic segment together with the compressed image data. Consequently, the compression ratio is also decreased because extra bits are required for transmitting the semantic segment. To solve this problem, we propose a new layered image compression framework with encoder-decoder matched semantic segmentation (EDMS). And then, followed by the semantic segmentation, a special convolution neural network is used to enhance the inaccurate semantic segment. As a result, the accurate semantic segment can be obtained in the decoder without requiring extra bits. The experimental results show that the proposed EDMS framework can get up to 35.31% BD-rate reduction over the HEVC-based (BPG) codec, 5% bitrate, and 24% encoding time saving compare to the state-of-the-art semantic-based image codec.
This thesis is a proof of concept for the potential of Variational Auto-Encoder (VAE) on representation learning of real-world Knowledge Graphs (KG). Inspired by successful approaches to the generation of molecular graphs, we evaluate the capabilities of our model, the Relational Graph Variational Auto-Encoder (RGVAE). The impact of the modular hyperparameter choices, encoding through graph convolutions, graph matching and latent space prior, is compared. The RGVAE is first evaluated on link prediction. The mean reciprocal rank (MRR) scores on the two datasets FB15K-237 and WN18RR are compared to the embedding-based model DistMult. A variational DistMult and a RGVAE without latent space prior constraint are implemented as control models. The results show that between different settings, the RGVAE with relaxed latent space, scores highest on both datasets, yet does not outperform the DistMult. Further, we investigate the latent space in a twofold experiment: first, linear interpolation between the latent representation of two triples, then the exploration of each latent dimension in a $95\%$ confidence interval. Both interpolations show that the RGVAE learns to reconstruct the adjacency matrix but fails to disentangle. For the last experiment we introduce a new validation method for the FB15K-237 data set. The relation type-constrains of generated triples are filtered and matched with entity types. The observed rate of valid generated triples is insignificantly higher than the random threshold. All generated and valid triples are unseen. A comparison between different latent space priors, using the $\delta$-VAE method, reveals a decoder collapse. Finally we analyze the limiting factors of our approach compared to molecule generation and propose solutions for the decoder collapse and successful representation learning of multi-relational KGs.
This work presents a new fine-grained transparent object segmentation dataset, termed Trans10K-v2, extending Trans10K-v1, the first large-scale transparent object segmentation dataset. Unlike Trans10K-v1 that only has two limited categories, our new dataset has several appealing benefits. (1) It has 11 fine-grained categories of transparent objects, commonly occurring in the human domestic environment, making it more practical for real-world application. (2) Trans10K-v2 brings more challenges for the current advanced segmentation methods than its former version. Furthermore, a novel transformer-based segmentation pipeline termed Trans2Seg is proposed. Firstly, the transformer encoder of Trans2Seg provides the global receptive field in contrast to CNN's local receptive field, which shows excellent advantages over pure CNN architectures. Secondly, by formulating semantic segmentation as a problem of dictionary look-up, we design a set of learnable prototypes as the query of Trans2Seg's transformer decoder, where each prototype learns the statistics of one category in the whole dataset. We benchmark more than 20 recent semantic segmentation methods, demonstrating that Trans2Seg significantly outperforms all the CNN-based methods, showing the proposed algorithm's potential ability to solve transparent object segmentation.
The recently proposed Detection Transformer (DETR) model successfully applies Transformer to objects detection and achieves comparable performance with two-stage object detection frameworks, such as Faster-RCNN. However, DETR suffers from its slow convergence. Training DETR \cite{carion2020end} from scratch needs 500 epochs to achieve a high accuracy. To accelerate its convergence, we propose a simple yet effective scheme for improving the DETR framework, namely Spatially Modulated Co-Attention (SMCA) mechanism. The core idea of SMCA is to conduct regression-aware co-attention in DETR by constraining co-attention responses to be high near initially estimated bounding box locations. Our proposed SMCA increases DETR's convergence speed by replacing the original co-attention mechanism in the decoder while keeping other operations in DETR unchanged. Furthermore, by integrating multi-head and scale-selection attention designs into SMCA, our fully-fledged SMCA can achieve better performance compared to DETR with a dilated convolution-based backbone (45.6 mAP at 108 epochs vs. 43.3 mAP at 500 epochs). We perform extensive ablation studies on COCO dataset to validate the effectiveness of the proposed SMCA.
We propose a new convolution neural network called HarDNet-MSEG for polyp segmentation. It achieves SOTA in both accuracy and inference speed on five popular datasets. For Kvasir-SEG, HarDNet-MSEG delivers 0.904 mean Dice running at 86.7 FPS on a GeForce RTX 2080 Ti GPU. It consists of a backbone and a decoder. The backbone is a low memory traffic CNN called HarDNet68, which has been successfully applied to various CV tasks including image classification, object detection, multi-object tracking and semantic segmentation, etc. The decoder part is inspired by the Cascaded Partial Decoder, known for fast and accurate salient object detection. We have evaluated HarDNet-MSEG using those five popular datasets. The code and all experiment details are available at Github. https://github.com/james128333/HarDNet-MSEG
Metric learning has received conflicting assessments concerning its suitability for solving instance segmentation tasks. It has been dismissed as theoretically flawed due to the shift equivariance of the employed CNNs and their respective inability to distinguish same-looking objects. Yet it has been shown to yield state of the art results for a variety of tasks, and practical issues have mainly been reported in the context of tile-and-stitch approaches, where discontinuities at tile boundaries have been observed. To date, neither of the reported issues have undergone thorough formal analysis. In our work, we contribute a comprehensive formal analysis of the shift equivariance properties of encoder-decoder-style CNNs, which yields a clear picture of what can and cannot be achieved with metric learning in the face of same-looking objects. In particular, we prove that a standard encoder-decoder network that takes $d$-dimensional images as input, with $l$ pooling layers and pooling factor $f$, has the capacity to distinguish at most $f^{dl}$ same-looking objects, and we show that this upper limit can be reached. Furthermore, we show that to avoid discontinuities in a tile-and-stitch approach, assuming standard batch size 1, it is necessary to employ valid convolutions in combination with a training output window size strictly greater than $f^l$, while at test-time it is necessary to crop tiles to size $n\cdot f^l$ before stitching, with $n\geq 1$. We complement these theoretical findings by discussing a number of insightful special cases for which we show empirical results on synthetic data.
The difficulty of obtaining paired data remains a major bottleneck for learning image restoration and enhancement models for real-world applications. Current strategies aim to synthesize realistic training data by modeling noise and degradations that appear in real-world settings. We propose DeFlow, a method for learning stochastic image degradations from unpaired data. Our approach is based on a novel unpaired learning formulation for conditional normalizing flows. We model the degradation process in the latent space of a shared flow encoder-decoder network. This allows us to learn the conditional distribution of a noisy image given the clean input by solely minimizing the negative log-likelihood of the marginal distributions. We validate our DeFlow formulation on the task of joint image restoration and super-resolution. The models trained with the synthetic data generated by DeFlow outperform previous learnable approaches on three recent datasets. Code and trained models are available at: https://github.com/volflow/DeFlow
This work presents a non-intrusive surrogate modeling scheme based on machine learning technology for predictive modeling of complex systems, described by parametrized time-dependent PDEs. For these problems, typical finite element approaches involve the spatiotemporal discretization of the PDE and the solution of the corresponding linear system of equations at each time step. Instead, the proposed method utilizes a convolutional autoencoder in conjunction with a feed forward neural network to establish a low-cost and accurate mapping from the problem's parametric space to its solution space. For this purpose, time history response data are collected by solving the high-fidelity model via FEM for a reduced set of parameter values. Then, by applying the convolutional autoencoder to this data set, a low-dimensional representation of the high-dimensional solution matrices is provided by the encoder, while the reconstruction map is obtained by the decoder. Using the latent representation given by the encoder, a feed-forward neural network is efficiently trained to map points from the problem's parametric space to the compressed version of the respective solution matrices. This way, the encoded response of the system at new parameter values is given by the neural network, while the entire response is delivered by the decoder. This approach effectively bypasses the need to serially formulate and solve the system's governing equations at each time increment, thus resulting in a significant cost reduction and rendering the method ideal for problems requiring repeated model evaluations or 'real-time' computations. The elaborated methodology is demonstrated on the stochastic analysis of time-dependent PDEs solved with the Monte Carlo method, however, it can be straightforwardly applied to other similar-type problems, such as sensitivity analysis, design optimization, etc.
Although deep models have greatly improved the accuracy and robustness of image segmentation, obtaining segmentation results with highly accurate boundaries and fine structures is still a challenging problem. In this paper, we propose a simple yet powerful Boundary-Aware Segmentation Network (BASNet), which comprises a predict-refine architecture and a hybrid loss, for highly accurate image segmentation. The predict-refine architecture consists of a densely supervised encoder-decoder network and a residual refinement module, which are respectively used to predict and refine a segmentation probability map. The hybrid loss is a combination of the binary cross entropy, structural similarity and intersection-over-union losses, which guide the network to learn three-level (ie, pixel-, patch- and map- level) hierarchy representations. We evaluate our BASNet on two reverse tasks including salient object segmentation, camouflaged object segmentation, showing that it achieves very competitive performance with sharp segmentation boundaries. Importantly, BASNet runs at over 70 fps on a single GPU which benefits many potential real applications. Based on BASNet, we further developed two (close to) commercial applications: AR COPY & PASTE, in which BASNet is integrated with augmented reality for "COPYING" and "PASTING" real-world objects, and OBJECT CUT, which is a web-based tool for automatic object background removal. Both applications have already drawn huge amount of attention and have important real-world impacts. The code and two applications will be publicly available at: https://github.com/NathanUA/BASNet.
Accurately forecasting air quality is critical to protecting general public from lung and heart diseases. This is a challenging task due to the complicated interactions among distinct pollution sources and various other influencing factors. Existing air quality forecasting methods cannot effectively model the diffusion processes of air pollutants between cities and monitoring stations, which may suddenly deteriorate the air quality of a region. In this paper, we propose HighAir, i.e., a hierarchical graph neural network-based air quality forecasting method, which adopts an encoder-decoder architecture and considers complex air quality influencing factors, e.g., weather and land usage. Specifically, we construct a city-level graph and station-level graphs from a hierarchical perspective, which can consider city-level and station-level patterns, respectively. We design two strategies, i.e., upper delivery and lower updating, to implement the inter-level interactions, and introduce message passing mechanism to implement the intra-level interactions. We dynamically adjust edge weights based on wind direction to model the correlations between dynamic factors and air quality. We compare HighAir with the state-of-the-art air quality forecasting methods on the dataset of Yangtze River Delta city group, which covers 10 major cities within 61,500 km2. The experimental results show that HighAir significantly outperforms other methods.