efficientnet VS segmentation_models

Examples of lightweight models include MobileNet, a computer vision model designed for mobile and embedded vision applications, EfficientDet, an object detection model, and EfficientNet, a CNN that uses compound scaling to enable better performance. All these are lightweight models from Google.

There was a recent paper by Facebook (2022), where they modernise a vanilla ConvNet by using the latest empirical design choices and manage to achieve state-of-the-art performance with it. This was also done before, with EffecientNet in 2019.

not true at all, plenty of sota models combines batchnorm and dropout 1. efficientnet 2. resnet rs 3. timm resnet50 (appendix)

Right - I think these days they do more of a balanced tradeoff between width and depth. One more recent CNN, Efficientnet, carefully chooses the width-to-depth ratio to have the best performance for a given compute budget.

EfficientNet a lightweight convolutional neural network architecture achieving the state-of-the-art accuracy with an order of magnitude fewer parameters and FLOPS, on both ImageNet and five other commonly used transfer learning datasets.

Additionally, if you want to custom change the number of filters in the EfficientNet I would suggest using the detailed Keras implementation of the EfficientNet in this repository.

Also, I’d suggest considering switching to the segmentation-models library - it provides U-Net models with a variety of pretrained backbones of as encoders. The author also put out a PyTorch version. https://github.com/qubvel/segmentation_models.pytorch https://github.com/qubvel/segmentation_models

I used segmentation-models (https://github.com/qubvel/segmentation_models) to create a deeplabv3+ model. I havent used it in the last 2 months and now i comeback to the same code and cant use it. Getting ModuleNotFoundError: No module named 'segmentation_models_pytorch.deeplabv3'

I'm building some semantic segmentation models off of low-moderate volumes of biomedical images (~500 - 1k images). So far I've done some hyperparameter sweeping (learning rate, transfer learning, architectures, dropout layers) using the Segmentation Models package from qubvel https://github.com/qubvel/segmentation_models but I'm only seeing moderate performance and minimal differences between tested parameters.

So, image segmentation (especially for satellite images) is a known problem. Search for semantic segmentation and unet (a model used for semantic segmentation). Also, if you use tensorflow there is this library which I found useful segmentation models.

I am doing an image segmentation project using https://github.com/qubvel/segmentation_models as the baseline. I was wondering if any of you have tried the latest segmentation models from cvpr papers. If yes, which ones you found to be interesting or actually improve miou. And how difficult/easy it is to implement those?

That is just U-net from this repo, anything aside is slicing images to fit into window and predict call. I measure time of predict() and it is the same as profiler numbers, so definitely my other operations are beyond profiler. C API code is just creating tensors and calling TF_SessionRun plus slice operations with opencv. Can't post code, sorry.

I'm trying to train image segmentation model with transfer learning using https://github.com/qubvel/segmentation_models/.