automl VS FLAML

I'm using Google Brain's EfficientDet repo to train an object detector. What hyperparameters should I choose to increase accuracy for small objects.

EfficientDet-D0 has comparable accuracy as YOLOv3.

Code for https://arxiv.org/abs/2104.00298 found: https://github.com/google/automl/efficientnetv2

7 Min Read | Paper (CoAtNet) | Paper (EfficientNetV2) | Google blog | Code

Abstract: This paper introduces EfficientNetV2, a new family of convolutional networks that have faster training speed and better parameter efficiency than previous models. To develop this family of models, we use a combination of training-aware neural architecture search and scaling, to jointly optimize training speed and parameter efficiency. The models were searched from the search space enriched with new ops such as Fused-MBConv. Our experiments show that EfficientNetV2 models train much faster than state-of- the-art models while being up to 6.8x smaller. > Our training can be further sped up by progressively increasing the image size during training, but it often causes a drop in accuracy. To compensate for this accuracy drop, we propose to adaptively adjust regularization (e.g., dropout and data augmentation) as well, such that we can achieve both fast training and good accuracy. > With progressive learning, our EfficientNetV2 significantly outperforms previous models on ImageNet and CIFAR/Cars/Flowers datasets. By pretraining on the same ImageNet21k, our EfficientNetV2 achieves 87.3% top-1 accuracy on ImageNet ILSVRC2012, outperforming the recent ViT by 2.0% accuracy while training 5x-11x faster using the same computing resources. Code will be available at this https URL.

I really like the simplicity of this framework, and they hit on a lot of common problems found in other agent-based frameworks. Most intrigued by the RAG improvements.

Seems like Microsoft was frustrated with the pace of movement in this space and the shitty results of agents (which admittedly kept my interest turned away from agents for the last few months). I'm interested again because it makes practical sense, and from looking at the example notebooks, seems fairly easy to integrate into existing applications.

Maybe this is the 'low code' approach that might actually work, and bridge together engineering and non-engineering resources.

This example was what caught my eye: https://github.com/microsoft/FLAML/blob/main/notebook/autoge...

4. FLAML

Improved AutoML - Again, with collaboration from Microsoft Research, we used FLAML to update our existing AutoML solutions. What does this mean for you? You're using the latest techniques but all you need is a problem to solve and some data to get started.

For a complete tutorial, navigate to this Jupyter Notebook: https://github.com/microsoft/FLAML/blob/main/notebook/flaml_automl.ipynb

Microsoft Researchers have developed FLAML (Fast Lightweight AutoML) which can now utilize Ray Tune for distributed hyperparameter tuning to scale up FLAML’s resource-efficient & seamlessly parallelizable algorithms across a cluster.

Looks nice but I wonder if this is practical for non-tiny problems. The papers are a bit hard to follow but it looks like training is restarted with every new architecture choice. As for the library itself, the only large neural net example is a finetune of an NLP model that only searches over ADAM's optimizer params - which could be useful but it's a stretch to call that AutoML.