Accuracy on ImageNet of Image Classification Models
Examining the Relationship between the number of trainable parameters and the accuracy of Image classification models trained on ImageNet benchmarking dataset.
Created on November 17|Last edited on February 1
Comment
Models
the models evaluated here using WandB on ImageNet test set are taken from Ross Weightmas fantastic image Repository:
you can install his model library with
pip install timm
This repository makes available (at the time of writing) more than 600 models and their weights that have been pre-trained on image net 1k.
The repository also makes available data on a given model's accuracy. Models in this repository are both history and current state-of-the-art image classification datasets.
We built and computed the number of parameters with this short google colab and logged this to a W&B table.
Dataset
Image Net consists of 1k classees and has various subsets and is widely used as a benchmarking dataset for image classification, access to this dataset and further details can be found here.
ImageNet Accuracy
The bottom chart is a scatter chart and 'mousing over' individual points will show the model. We have displayed this here as a means of examining the relationship between the scale/size of a model in a number of trainable parameters and accuracy.
And also to show what might be outliers, (look for example at VVG Nets).
Run: helpful-salad-47
29
Add a comment