Building and productionizing Machine Learning (ML) models is a process of interdependent steps of iterative code updates, including exploratory model design, hyperparameter tuning, ablation experiments, and model training. Industrial-strength ML involves doing this at scale, using many compute resources, and this requires rewriting the training code to account for distribution. The result is that moving from a single host program to a cluster hinders iterative development of the software, as iterative development would require multiple versions of the software to be maintained and kept consistent. In this paper, we introduce the distribution oblivious training function as an abstraction for ML development in Python, whereby developers can reuse the same training function when running a notebook on a laptop or performing scale-out hyperparameter search and distributed training on clusters. Programs written in our framework look like industry-standard ML programs as we factor out dependencies using best-practice programming idioms (such as functions to generate models and data batches). We believe that our approach takes a step towards unifying single-host and distributed ML development.

Running extensive experiments is essential for building Machine Learning (ML) models. Such experiments usually require iterative execution of many trials with varying run times. In recent years, Apache Spark has become the de-facto standard for parallel data processing in the industry, in which iterative processes are implemented within the bulk-synchronous parallel (BSP) execution model. The BSP approach is also being used to parallelize ML trials in Spark. However, the BSP task synchronization barriers prevent asynchronous execution of trials, which leads to a reduced number of trials that can be run on a given computational budget. In this paper, we introduce Maggy, an open-source framework based on Spark, to execute ML trials asynchronously in parallel, with the ability to early stop poorly performing trials. In the experiments, we compare Maggy with the BSP execution of parallel trials in Spark and show that on random hyperparameter search on a convolutional neural network for the Fashion-MNIST dataset Maggy reduces the required time to execute a fixed number of trials by 33% to 58%, without any loss in the final model accuracy.

Ablation studies provide insights into the relative contribution of different architectural and regularization components to machine learning models' performance. In this paper, we introduce AutoAblation, a new framework for the design and parallel execution of ablation experiments. AutoAblation provides a declarative approach to defining ablation experiments on model architectures and training datasets, and enables the parallel execution of ablation trials. This reduces the execution time and allows more comprehensive experiments by exploiting larger amounts of computational resources. We show that AutoAblation can provide near-linear scalability by performing an ablation study on the modules of the Inception-v3 network trained on the TenGeoPSAR dataset.  

Deep neural networks used for computer vision tasks are typically trained on datasets consisting of thousands of images, called examples. Recent studies have shown that examples in a dataset are not of equal importance for model training and can be categorized based on quantifiable measures reflecting a notion of “hardness” or “importance”. In this work, we conduct an empirical study of the impact of importance-aware partitioning of the dataset examples across workers on the performance of data-parallel training of deep neural networks. Our experiments with CIFAR-10 and CIFAR-100 image datasets show that data-parallel training with importance-aware partitioning can perform better than vanilla data-parallel training, which is oblivious to the importance of examples. More specifically, the proper choice of the importance measure, partitioning heuristic, and the number of intervals for dataset repartitioning can improve the best accuracy of the model trained for a fixed number of epochs. We conclude that the parameters related to importance-aware data-parallel training, including the importance measure, number of warmup training epochs, and others defined in the paper, may be considered as hyperparameters of data-parallel model training.

Training of deep neural networks from scratch requires initialization of the neural network weights as a first step. Over the years, many policies and techniques for weight initialization have been proposed and widely used, including Kaiming initialization and different variants of random initialization. On the other hand, another requirement for starting the training stage is to choose and set suitable hyperparameter values, which are usually obtained by performing several hyperparameter tuning trials. In this paper, we study the suitability of weight initialization using weights obtained from different epochs of hyperparameter tuning trials and compare it to Kaiming uniform (random) weight initialization for image classification tasks. Based on an experimental evaluation using ResNet-18, ResNet-152, and InceptionV3 models, and CIFAR-10, CIFAR-100, Tiny ImageNet, and Food-101 datasets, we show that weight initialization from hyperparameter tuning trials can speed up the training of deep neural networks by up to 2x while maintaining or improving the best test accuracy of the trained models, when compared to random initialization.

In Machine Learning (ML) and Deep Learning (DL) research, ablation studies are typically performed to provide insights into the individual contribution of different building blocks and components of an ML/DL system (e.g., a deep neural network), as well as to justify that certain additions or modifications to an existing ML/DL system can result in the proposed improved performance. Although dedicated frameworks for performing ablation studies have been introduced in recent years, conducting such experiments is still associated with requiring tedious, redundant work, typically involving maintaining redundant and nearly identical versions of code that correspond to different ablation trials. Inspired by the recent promising performance of Large Language Models (LLMs) in the generation and analysis of ML/DL code, in this paper we discuss the potential of LLMs as facilitators of ablation study experiments for scientific research projects that involve or deal with ML and DL models. We first discuss the different ways in which LLMs can be utilized for ablation studies and then present the prototype of a tool called AblationMage, that leverages LLMs to semi-automate the overall process of conducting ablation study experiments. We showcase the usability of AblationMage as a tool through three experiments, including one in which we reproduce the ablation studies from a recently published applied DL paper.