Tensorflow Keras implementation of : Image classification with ConvMixer
The full credit goes to: Sayak Paul
Short description:
ConvMixer is a simple model based on the ideas of representing an image as patches( used in ViT) and separating the mixing of Spatial and channel dimensions (used in MLP-Mixer). Unlike ViT and MLP-Mixer, they use only standard Convolution operations. The full paper is a submission to ICLR 22 and can be found here
Model and Dataset used
The Dataset used here is CIFAR-10. The model is called ConvMixer-256/8 where 256 is the hidden dimension (the dimension of patches) and 8 is the depth(number of repetitions of ConvMix layers)
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
| Hyperparameters | Value |
|---|---|
| name | AdamW |
| learning_rate | 0.0010000000474974513 |
| decay | 0.0 |
| beta_1 | 0.8999999761581421 |
| beta_2 | 0.9990000128746033 |
| epsilon | 1e-07 |
| amsgrad | False |
| weight_decay | 9.999999747378752e-05 |
| exclude_from_weight_decay | None |
| training_precision | float32 |
Training Metrics
After 10 Epocs, the test accuracy of the model is 83.57%
Model Plot
View Model Plot
- Downloads last month
- 12
Inference API (serverless) does not yet support tf-keras models for this pipeline type.