|
--- |
|
datasets: |
|
- MNIST |
|
- SVHN |
|
library_name: tf-keras |
|
tags: |
|
- semi-supervised |
|
- image classification |
|
- domain adaption |
|
--- |
|
|
|
## Model description |
|
|
|
This is an image classification model based on a [WideResNet-2-28](https://arxiv.org/abs/1605.07146v4), trained using the [AdaMatch](https://arxiv.org/abs/2106.04732) method by Berthelot et al. |
|
|
|
The training was based on the example [Semi-supervision and domain adaptation with AdaMatch]('https://keras.io/examples/vision/adamatch/') on keras.io by [Sayak Paul](https://twitter.com/RisingSayak). |
|
|
|
The main difference to the training in the keras.io example is that here I increased the number of Epochs to 30, for a better target dataset performance. |
|
|
|
|
|
## Intended uses & limitations |
|
|
|
AdaMatch attempts to combine *semi-supervised learning*, i.e. learning with a partially labelled dataset and *unsupersived domain adaption*, i.e. adapting a model to a different domain dataset without any labels. |
|
|
|
So it actually performs **semi-supervised domain adaptation (SSDA)**. |
|
|
|
The model is inteded to show that AdaMatch is able to carry out SSDA, with a accuracy on the target domain (SVHN) that is exceeding or competitive with other methods. |
|
|
|
### Limitations |
|
The model was trained on MNIST as source and SVHN as target dataset. Thus, the classification performance on MNIST is very good (98.46%), while the accuracy on SVHN is "only" at 26.51%. Compared to the training of the same architecture without AdaMatch, this still is about 17% better |
|
|
|
## Training and evaluation data |
|
|
|
### Training Data |
|
The model was trained using the [MNIST](https://huggingface.co/datasets/mnist) (as source domain) and [SVHN cropped](http://ufldl.stanford.edu/housenumbers/) (as target domain) datasets. For training the images were used at a resolution of (32,32,3). |
|
|
|
Augmented versions of the source and target data were created in two versions - weakly and strongly augmented, as written in the original paper. |
|
|
|
### Training Procedure |
|
This image from the original paper shows the workflow of AdaMatch: |
|
 |
|
For more information, refer to the [paper](https://arxiv.org/abs/2106.04732) or the original example at [keras.io]('https://keras.io/examples/vision/adamatch/'). |
|
|
|
### Hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
|
|
- Epochs: 30 |
|
- Source Batch Size: 64 |
|
- Target Batch Size: 3 * 64 |
|
- Learning Rate: 0.03 |
|
- Weight Decay: 0.0005 |
|
- Network Depth: 28 |
|
- Network Width Multiplier = 2 |
|
|
|
## Evaluation |
|
|
|
Accuracy on **source** test set: **98.46%** |
|
|
|
Accuracy on **target** test set: **26.51%** |
|
|
|
|
