Search results for: PyTorch

Authors: Elias Nemer, Greg Vines

Abstract:

Time-frequency transformation and spectral representations of audio signals are commonly used in various machine learning applications. Training networks on frequency features such as the Mel-Spectrogram or Cochleogram have been proven more effective and convenient than training on-time samples. In practical realizations, these features are created on a different processor and/or pre-computed and stored on disk, requiring additional efforts and making it difficult to experiment with different features. In this paper, we provide a PyTorch framework for creating various spectral features as well as time-frequency transformation and time-domain filter-banks using the built-in trainable conv1d() layer. This allows computing these features on the fly as part of a larger network and enabling easier experimentation with various combinations and parameters. Our work extends the work in the literature developed for that end: First, by adding more of these features and also by allowing the possibility of either starting from initialized kernels or training them from random values. The code is written as a template of classes and scripts that users may integrate into their own PyTorch classes or simply use as is and add more layers for various applications.

Keywords: neural networks Mel-Spectrogram, chromagram, cochleogram, discrete Fourrier transform, PyTorch conv1d()

Procedia PDF Downloads 195

Authors: Raghavi C. Janaswamy

Abstract:

In the healthcare sector, heterogenous data elements such as patients, diagnosis, symptoms, conditions, observation text from physician notes, and prescriptions form the essentials of the Electronic Health Record (EHR). The data in the form of clear text and images are stored or processed in a relational format in most systems. However, the intrinsic structure restrictions and complex joins of relational databases limit the widespread utility. In this regard, the design and development of realistic mapping and deep connections as real-time objects offer unparallel advantages. Herein, a graph neural network-based classification of EHR data has been developed. The patient conditions have been predicted as a node classification task using a graph-based open source EHR data, Synthea Database, stored in Tigergraph. The Synthea DB dataset is leveraged due to its closer representation of the real-time data and being voluminous. The graph model is built from the EHR heterogeneous data using python modules, namely, pyTigerGraph to get nodes and edges from the Tigergraph database, PyTorch to tensorize the nodes and edges, PyTorch-Geometric (PyG) to train the Graph Neural Network (GNN) and adopt the self-supervised learning techniques with the AutoEncoders to generate the node embeddings and eventually perform the node classifications using the node embeddings. The model predicts patient conditions ranging from common to rare situations. The outcome is deemed to open up opportunities for data querying toward better predictions and accuracy.

Keywords: electronic health record, graph neural network, heterogeneous data, prediction

Procedia PDF Downloads 60

Authors: Albert Bou, Sebastian Dittert, Gianni de Fabritiis

Abstract:

Advancing reinforcement learning (RL) requires tools that are flexible enough to easily prototype new methods while avoiding impractically slow experimental turnaround times. To match the first requirement, the most popular RL libraries advocate for highly modular agent composability, which facilitates experimentation and development. To solve challenging environments within reasonable time frames, scaling RL to large sampling and computing resources has proved a successful strategy. However, this capability has been so far difficult to combine with modularity. In this work, we explore design choices to allow agent composability both at a local and distributed level of execution. We propose a versatile approach that allows the definition of RL agents at different scales through independent, reusable components. We demonstrate experimentally that our design choices allow us to reproduce classical benchmarks, explore multiple distributed architectures, and solve novel and complex environments while giving full control to the user in the agent definition and training scheme definition. We believe this work can provide useful insights to the next generation of RL libraries.

Keywords: deep reinforcement learning, Python, PyTorch, distributed training, modularity, library

Procedia PDF Downloads 53

Authors: Megha Gupta, Nupur Prakash

Abstract:

Identification of plant diseases has been performed using machine learning and deep learning models on the datasets containing images of healthy and diseased plant leaves. The current study carries out an evaluation of some of the deep learning models based on convolutional neural network (CNN) architectures for identification of plant diseases. For this purpose, the publicly available New Plant Diseases Dataset, an augmented version of PlantVillage dataset, available on Kaggle platform, containing 87,900 images has been used. The dataset contained images of 26 diseases of 14 different plants and images of 12 healthy plants. The CNN models selected for the study presented in this paper are AlexNet, ZFNet, VGGNet (four models), GoogLeNet, and ResNet (three models). The selected models are trained using PyTorch, an open-source machine learning library, on Google Colaboratory. A comparative study has been carried out to analyze the high degree of accuracy achieved using these models. The highest test accuracy and F1-score of 99.59% and 0.996, respectively, were achieved by using GoogLeNet with Mini-batch momentum based gradient descent learning algorithm.

Keywords: comparative analysis, convolutional neural networks, deep learning, plant disease identification

Procedia PDF Downloads 159

Authors: Nurul Bin Ibrahim

Abstract:

Advancements in neural networks have offered valuable insights into Fluid Dynamics, notably in addressing turbulence-related challenges. In this research, we introduce multiple applications of models of neural networks, namely Feed-Forward and Recurrent Neural Networks, to explore the relationship between jet formations and stratified turbulence within stochastically excited Boussinesq systems. Using machine learning tools like TensorFlow and PyTorch, the study has created models that effectively mimic and show the underlying features of the complex patterns of jet formation and stratified turbulence. These models do more than just help us understand these patterns; they also offer a faster way to solve problems in stochastic systems, improving upon traditional numerical techniques to solve stochastic differential equations such as the Euler-Maruyama method. In addition, the research includes a thorough comparison with the Statistical State Dynamics (SSD) approach, which is a well-established method for studying chaotic systems. This comparison helps evaluate how well neural networks can help us understand the complex relationship between jet formations and stratified turbulence. The results of this study underscore the potential of neural networks in computational physics and fluid dynamics, opening up new possibilities for more efficient and accurate simulations in these fields.

Keywords: neural networks, machine learning, computational fluid dynamics, stochastic systems, simulation, stratified turbulence

Procedia PDF Downloads 38

Authors: Virender Singh, Mathew Rees, Simon Hampton, Sivaram Annadurai

Abstract:

Plant identification is a challenging task that aims to identify the family, genus, and species according to plant morphological features. Automated deep learning-based computer vision algorithms are widely used for identifying plants and can help users narrow down the possibilities. However, numerous morphological similarities between and within species render correct classification difficult. In this paper, we tested custom convolution neural network (CNN) and vision transformer (ViT) based models using the PyTorch framework to classify plants. We used a large dataset of 88,000 provided by the Royal Horticultural Society (RHS) and a smaller dataset of 16,000 images from the PlantClef 2015 dataset for classifying plants at genus and species levels, respectively. Our results show that for classifying plants at the genus level, ViT models perform better compared to CNN-based models ResNet50 and ResNet-RS-420 and other state-of-the-art CNN-based models suggested in previous studies on a similar dataset. ViT model achieved top accuracy of 83.3% for classifying plants at the genus level. For classifying plants at the species level, ViT models perform better compared to CNN-based models ResNet50 and ResNet-RS-420, with a top accuracy of 92.5%. We show that the correct set of augmentation techniques plays an important role in classification success. In conclusion, these results could help end users, professionals and the general public alike in identifying plants quicker and with improved accuracy.

Keywords: plant identification, CNN, image processing, vision transformer, classification

Procedia PDF Downloads 60

Authors: Susan Diamond

Abstract:

Machine Learning workloads have traditionally been run in high-performance computing (HPC) environments, where users log in to dedicated machines and utilize the attached GPUs to run training jobs on huge datasets. Training of large neural network models is very resource intensive, and even after exploiting parallelism and accelerators such as GPUs, a single training job can still take days. Consequently, the cost of hardware is a barrier to entry. Even when upfront cost is not a concern, the lead time to set up such an HPC environment takes months from acquiring hardware to set up the hardware with the right set of firmware, software installed and configured. Furthermore, scalability is hard to achieve in a rigid traditional lab environment. Therefore, it is slow to react to the dynamic change in the artificial intelligent industry. Watson Deep Learning as a service, a cloud-based deep learning platform that mitigates the long lead time and high upfront investment in hardware. It enables robust and scalable sharing of resources among the teams in an organization. It is designed for on-demand cloud environments. Providing a similar user experience in a multi-tenant cloud environment comes with its own unique challenges regarding fault tolerance, performance, and security. Watson Deep Learning as a service tackles these challenges and present a deep learning stack for the cloud environments in a secure, scalable and fault-tolerant manner. It supports a wide range of deep-learning frameworks such as Tensorflow, PyTorch, Caffe, Torch, Theano, and MXNet etc. These frameworks reduce the effort and skillset required to design, train, and use deep learning models. Deep Learning as a service is used at IBM by AI researchers in areas including machine translation, computer vision, and healthcare. 

Keywords: deep learning, machine learning, cognitive computing, model training

Procedia PDF Downloads 178

Authors: Evan Lowhorn, Rocio Alba-Flores

Abstract:

The purpose of this work is to develop a method for classifying hand signals and using the output in a drone control algorithm. To achieve this, methods based on Convolutional Neural Networks (CNN) were applied. CNN's are a subset of deep learning, which allows grid-like inputs to be processed and passed through a neural network to be trained for classification. This type of neural network allows for classification via imaging, which is less intrusive than previous methods using biosensors, such as EMG sensors. Classification CNN's operate purely from the pixel values in an image; therefore they can be used without additional exteroceptive sensors. A development bench was constructed using a desktop computer connected to a high-definition webcam mounted on a scissor arm. This allowed the camera to be pointed downwards at the desk to provide a constant solid background for the dataset and a clear detection area for the user. A MATLAB script was created to automate dataset image capture at the development bench and save the images to the desktop. This allowed the user to create their own dataset of 12,000 images within three hours. These images were evenly distributed among seven classes. The defined classes include forward, backward, left, right, idle, and land. The drone has a popular flip function which was also included as an additional class. To simplify control, the corresponding hand signals chosen were the numerical hand signs for one through five for movements, a fist for land, and the universal “ok” sign for the flip command. Transfer learning with PyTorch (Python) was performed using a pre-trained 18-layer residual learning network (ResNet-18) to retrain the network for custom classification. An algorithm was created to interpret the classification and send encoded messages to a Ryze Tello drone over its 2.4 GHz Wi-Fi connection. The drone’s movements were performed in half-meter distance increments at a constant speed. When combined with the drone control algorithm, the classification performed as desired with negligible latency when compared to the delay in the drone’s movement commands.

Keywords: classification, computer vision, convolutional neural networks, drone control

Procedia PDF Downloads 179

Authors: Robert Caulk

Abstract:

A generalized framework for adaptive machine learning deployments in algorithmic trading is introduced, tested, and released as open-source code. The presented software aims to test the hypothesis that recent data contains enough information to form a probabilistically favorable short-term price prediction. Further, the framework contains various adaptive machine learning techniques that are geared toward generating profit during strong trends and minimizing losses during trend changes. Results demonstrate that this adaptive machine learning approach is capable of capturing trends and generating profit. The presentation also discusses the importance of defining the parameter space associated with the dynamic training data-set and using the parameter space to identify and remove outliers from prediction data points. Meanwhile, the generalized architecture enables common users to exploit the powerful machinery while focusing on high-level feature engineering and model testing. The presentation also highlights common strengths and weaknesses associated with the presented technique and presents a broad range of well-tested starting points for feature set construction, target setting, and statistical methods for enforcing risk management and maintaining probabilistically favorable entry and exit points. The presentation also describes the end-to-end data processing tools associated with FreqAI, including automatic data fetching, data aggregation, feature engineering, safe and robust data pre-processing, outlier detection, custom machine learning and statistical tools, data post-processing, and adaptive training backtest emulation, and deployment of adaptive training in live environments. Finally, the generalized user interface is also discussed in the presentation. Feature engineering is simplified so that users can seed their feature sets with common indicator libraries (e.g. TA-lib, pandas-ta). The user also feeds data expansion parameters to fill out a large feature set for the model, which can contain as many as 10,000+ features. The presentation describes the various object-oriented programming techniques employed to make FreqAI agnostic to third-party libraries and external data sources. In other words, the back-end is constructed in such a way that users can leverage a broad range of common regression libraries (Catboost, LightGBM, Sklearn, etc) as well as common Neural Network libraries (TensorFlow, PyTorch) without worrying about the logistical complexities associated with data handling and API interactions. The presentation finishes by drawing conclusions about the most important parameters associated with a live deployment of the adaptive learning framework and provides the road map for future development in FreqAI.

Keywords: machine learning, market trend detection, open-source, adaptive learning, parameter space exploration

Procedia PDF Downloads 62

Authors: Hongmei Wang, Ziyun Xiang, Ying liu, Li Yu, Dongsheng Yue

Abstract:

Intro: The COVID-19 (Corona Virus Disease 2019) has greatly changed the global economic, political and financial ecology. The mutation of the coronavirus in the UK in December 2020 has brought new panic to the world. Deep learning was performed on Chest Computed tomography (CT) of COVID-19 and Influenza and describes their characteristics. The predominant features of COVID-19 pneumonia was ground-glass opacification, followed by consolidation. Lesion density: most lesions appear as ground-glass shadows, and some lesions coexist with solid lesions. Lesion distribution: the focus is mainly on the dorsal side of the periphery of the lung, with the lower lobe of the lungs as the focus, and it is often close to the pleura. Other features it has are grid-like shadows in ground glass lesions, thickening signs of diseased vessels, air bronchi signs and halo signs. The severe disease involves whole bilateral lungs, showing white lung signs, air bronchograms can be seen, and there can be a small amount of pleural effusion in the bilateral chest cavity. At the same time, this year's flu season could be near its peak after surging throughout the United States for months. Chest CT for Influenza infection is characterized by focal ground glass shadows in the lungs, with or without patchy consolidation, and bronchiole air bronchograms are visible in the concentration. There are patchy ground-glass shadows, consolidation, air bronchus signs, mosaic lung perfusion, etc. The lesions are mostly fused, which is prominent near the hilar and two lungs. Grid-like shadows and small patchy ground-glass shadows are visible. Deep neural networks have great potential in image analysis and diagnosis that traditional machine learning algorithms do not. Method: Aiming at the two major infectious diseases COVID-19 and influenza, which are currently circulating in the world, the chest CT of patients with two infectious diseases is classified and diagnosed using deep learning algorithms. The residual network is proposed to solve the problem of network degradation when there are too many hidden layers in a deep neural network (DNN). The proposed deep residual system (ResNet) is a milestone in the history of the Convolutional neural network (CNN) images, which solves the problem of difficult training of deep CNN models. Many visual tasks can get excellent results through fine-tuning ResNet. The pre-trained convolutional neural network ResNet is introduced as a feature extractor, eliminating the need to design complex models and time-consuming training. Fastai is based on Pytorch, packaging best practices for in-depth learning strategies, and finding the best way to handle diagnoses issues. Based on the one-cycle approach of the Fastai algorithm, the classification diagnosis of lung CT for two infectious diseases is realized, and a higher recognition rate is obtained. Results: A deep learning model was developed to efficiently identify the differences between COVID-19 and influenza using chest CT.

Keywords: COVID-19, Fastai, influenza, transfer network

Procedia PDF Downloads 112

Authors: Joseph M. Rich, Vinay A. Duddalwar, Assad A. Oberai

Abstract:

Prostate adenocarcinoma is the most common cancer in males, with osseous metastases as the commonest site of metastatic prostate carcinoma (mPC). Treatment monitoring is based on the evaluation and characterization of lesions on multiple imaging studies, including Computed Tomography (CT). Monitoring of the osseous disease burden, including follow-up of lesions and identification and characterization of new lesions, is a laborious task for radiologists. Deep learning algorithms are increasingly used to perform tasks such as identification and segmentation for osseous metastatic disease and provide accurate information regarding metastatic burden. Here, nnUNet was used to produce a model which can segment CT scan images of prostate adenocarcinoma vertebral bone metastatic lesions. nnUNet is an open-source Python package that adds optimizations to deep learning-based UNet architecture but has not been extensively combined with transfer learning techniques due to the absence of a readily available functionality of this method. The IRB-approved study data set includes imaging studies from patients with mPC who were enrolled in clinical trials at the University of Southern California (USC) Health Science Campus and Los Angeles County (LAC)/USC medical center. Manual segmentation of metastatic lesions was completed by an expert radiologist Dr. Vinay Duddalwar (20+ years in radiology and oncologic imaging), to serve as ground truths for the automated segmentation. Despite nnUNet’s success on some medical segmentation tasks, it only produced an average Dice Similarity Coefficient (DSC) of 0.31 on the USC dataset. DSC results fell in a bimodal distribution, with most scores falling either over 0.66 (reasonably accurate) or at 0 (no lesion detected). Applying more aggressive data augmentation techniques dropped the DSC to 0.15, and reducing the number of epochs reduced the DSC to below 0.1. Datasets have been identified for transfer learning, which involve balancing between size and similarity of the dataset. Identified datasets include the Pancreas data from the Medical Segmentation Decathlon, Pelvic Reference Data, and CT volumes with multiple organ segmentations (CT-ORG). Some of the challenges of producing an accurate model from the USC dataset include small dataset size (115 images), 2D data (as nnUNet generally performs better on 3D data), and the limited amount of public data capturing annotated CT images of bone lesions. Optimizations and improvements will be made by applying transfer learning and generative methods, including incorporating generative adversarial networks and diffusion models in order to augment the dataset. Performance with different libraries, including MONAI and custom architectures with Pytorch, will be compared. In the future, molecular correlations will be tracked with radiologic features for the purpose of multimodal composite biomarker identification. Once validated, these models will be incorporated into evaluation workflows to optimize radiologist evaluation. Our work demonstrates the challenges of applying automated image segmentation to small medical datasets and lays a foundation for techniques to improve performance. As machine learning models become increasingly incorporated into the workflow of radiologists, these findings will help improve the speed and accuracy of vertebral metastatic lesions detection.

Keywords: deep learning, image segmentation, medicine, nnUNet, prostate carcinoma, radiomics

Procedia PDF Downloads 64