AdaMatch-PyTorch
Unofficial PyTorch Implementation of AdaMatch: A Unified Approach to Semi-Supervised Learning and Domain Adaptation, using Dassl.pytorch.
With reference to the official implementation at https://github.com/google-research/adamatch.
Results
Evaluation Results
Unsupervised Domain Adaptation (UDA) experiments are run on the DigitFive dataset. An experiment is run for each source-target domain pair (totaling up to 20 experiments).
Test results are directly taken from the final checkpoint at the end of training (the paper takes the median over the last ten checkpoints).
Comparison of this implementation's results to Table 127 in the paper (this implementation / original paper):
| mnist | mnistm | svhn | syndigit | usps | Avg | ||
|---|---|---|---|---|---|---|---|
| mnist | - | 98.4 / 99.2 | 95.7 / 96.9 | 99.6 / 99.7 | 99.4 / 97.8 | 98.2 / 98.4 | |
| mnistm | 99.7 / 99.4 | - | 94.4 / 96.9 | 99.2 / 99.7 | 99.2 / 97.8 | 98.1 / 98.5 | |
| svhn | 98.6 / 99.3 | 97.7 / 98.9 | - | 99.3 / 99.6 | 97.6 / 90.4 | 98.3 / 97.0 | |
| syndigit | 99.6 / 99.4 | 97.3 / 99.0 | 95.8 / 97.0 | - | 96.1 / 95.8 | 97.2 / 97.8 | |
| usps | 99.5 / 99.3 | 98.2 / 98.9 | 96.1 / 96.6 | 99.5 / 94.9 | - | 98.3 / 97.4 | |
| Avg | 99.4 / 99.4 | 97.9 / 99.0 | 95.5 / 96.8 | 99.4 / 98.5 | 98.1 / 95.5 | 98.0 / 97.8 |
If you spot any mistakes in my implementation, feel free to submit a PR :)
Differences from Paper
- Training hyperparameters
- Each experiment is only trained for 900 epochs, while the paper trains for slightly longer.
- Augmentations
- For strong augmentations, this implementation uses RandAugment + CutOut, while the paper uses CTAugment.
Installation
Follow the instructions to install a Dassl.pytorch conda environment here.
Download the Digit-5 dataset as per the instructions here.
Usage
The main AdaMatch training logic is implemented in the forward_backward method in trainers/adamatch.py
The main entry point for running an experiment is train.py, after specifying the configuration files dataset-config-file, config-file for hyperparameters:
python train.py \
--root data_dir \
--seed 42 \
--trainer AdaMatch \
--source-domains mnist \
--target-domains svhn \
--dataset-config-file configs/datasets/digit5.yaml \
--config-file configs/trainers/digit5.yaml \
--output-dir output_dir/mnist_svhn_42
TensorBoard logs are stored in the specified output directory output-dir.
To reproduce the UDA DigitFive results above, you can run the run_digit5_da.sh script, specifying the data and output directories:
cd scripts
bash run_digit5_da.sh data_dir output_dir