Awesome

Unsupervised Data Augmentation experiments in PyTorch

Experiments with "Unsupervised Data Augmentation" method on Cifar10 dataset.

Based on "Unsupervised Data Augmentation"

Unsupervised Data Augmentation in nutshell

UDA

Requirements

All experiments are run using mlflow, please install the latest version of this library

pip install --upgrade mlflow

Experiments

Start MLFlow UI server

Please create output folder (e.g. $PWD/output) and setup mlflow server:

export OUTPUT_PATH=/path/to/output

and

mlflow server --backend-store-uri $OUTPUT_PATH/mlruns --default-artifact-root $OUTPUT_PATH/mlruns -p 5566 -h 0.0.0.0

MLflow dashboard is available in the browser at 0.0.0.0:5566

CIFAR10 dataset

Create once "CIFAR10" experiment

export MLFLOW_TRACKING_URI=$OUTPUT_PATH/mlruns
mlflow experiments create -n CIFAR10

Implementation details:

Models
- FastResnet inspired from cifar10-fast repository
- Wide-ResNet 28-2 from Wide-ResNet repository
Consistency loss: KL
Data augs: AutoAugment + Cutout
Cosine LR decay
Training Signal Annealing
Updated UDA version: see main_uda2.py
- training 4k batchs are also passed into unsupervised learning part

Fast ResNet

Start a single run

export MLFLOW_TRACKING_URI=$OUTPUT_PATH/mlruns

mlflow run experiments/ --experiment-name=CIFAR10 -P dataset=CIFAR10 -P network=fastresnet -P params="data_path=../input/cifar10;num_epochs=100;learning_rate=0.08;batch_size=512;TSA_proba_min=0.5;unlabelled_batch_size=1024"

Wide ResNet

Start a single run

export MLFLOW_TRACKING_URI=$OUTPUT_PATH/mlruns

mlflow run experiments/ --experiment-name=CIFAR10 -P dataset=CIFAR10 -P network=wideresnet -P params="data_path=../input/cifar10;num_epochs=100;learning_rate=0.1;batch_size=512;TSA_proba_min=0.1;unlabelled_batch_size=1024"

Paper's configuration

export MLFLOW_TRACKING_URI=$OUTPUT_PATH/mlruns

mlflow run experiments/ --experiment-name=CIFAR10 -P dataset=CIFAR10 -P network=wideresnet -P params="data_path=../input/cifar10;num_epochs=6250;learning_rate=0.03;batch_size=64;TSA_proba_min=0.1;unlabelled_batch_size=320;num_warmup_steps=20000"

Unfortunately, I can not reproduce paper's result with 5.3 test error.

Updated version of UDA

export MLFLOW_TRACKING_URI=$OUTPUT_PATH/mlruns

mlflow run experiments/ -e main_uda2 --experiment-name=CIFAR10 -P dataset=CIFAR10 -P network=fastresnet -P params="data_path=../input/cifar10;num_epochs=100;learning_rate=0.08;batch_size=512;unlabelled_batch_size=512"

Some results

fastresnet_uda_vs_uda2

Tensorboard

All experiments are also logged to the Tensorboard. To visualize the experiments, please install tensorboard and run :

# tensorboard --logdir=$OUTPUT_PATH/mlruns/<experiment_id>
tensorboard --logdir=$OUTPUT_PATH/mlruns/1

Acknowledgements

In this repository we are using the code from

Thanks to the authors for sharing their code!