File size: 2,572 Bytes
ad64a41 4e8e080 ad64a41 7557692 ad64a41 7557692 ad64a41 7557692 ad64a41 7557692 ad64a41 7557692 ad64a41 7557692 ad64a41 7557692 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 |
---
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%**
|