While large amounts of data are available everywhere, most groups lack experts to guide the development of machine learning systems. Automated machine learning (AutoML) provides methods and processes to make advanced machine learning techniques available for non-machine learning experts. In this hackathon, we present an introduction to AutoML using the recently developed AutoGluon toolkit. AutoGluon is an AutoML tookit built on top of MXNet Gluon that enables easy-to-use and easy-to-extend AutoML, with a focus on deep learning and deploying models in real-world applications. AutoGluon supports core AutoML functionality such as: hyperparameter optimization algorithms and early stopping training mechanisms. With just a single call to AutoGluon’s fit() function, AutoGluon will automatically train many models under thousands of different hyperparameter configurations that affect the training process, and then return the best model within a reasonable runtime. In addition, you can easily exert greater control over the training process, for example to provide hard time limits for training, and what computational resources each training run should leverage.
In this repository, we will focus on the task of image classification to demonstrate the usage of AutoGluon’s main APIs. We will show you how to easily train models on your own image dataset, and how to control the hyperparameter-tuning process to obtain competitive results on a Kaggle image classification competition as well as achieve state-of-the-art performance on CIFAR10 (one of the most popular benchmark image classification datasets).
This repository provides the implementation of
- AutoGluon-Tabular: Robust and Accurate AutoML for Structured Data.
- Single Path One-Shot Neural Architecture Search with Uniform Sampling.
- On Network Design Spaces for Visual Recognition.
pip install mxnet autogluon
pip install git+https://github.com/ildoonet/pytorch-randaugment
pip install --upgrade git+https://github.com/sovrasov/flops-counter.pytorch.git- OneDrive: Link
Our trained Supernet weight is in $Link/Supernet/checkpoint-XX.pth.tar, which can be used by Search.
def get_args():
parser = argparse.ArgumentParser("OneShot_cifar_Experiments_Configuration")
parser.add_argument('--signal', type=str, default='different_hpo', help='describe:glboal_hpo/')
parser.add_argument('--different-hpo', action='store_true', help='blockwisely lr_scheduler')
parser.add_argument('--auto-continue', type=bool, default=False, help=' # False if num_trials > 1, after hpo')
parser.add_argument('--eval', default=False, action='store_true', help='Training mode')
parser.add_argument('--eval-resume', type=str, default='./snet_detnas.pkl', help='path for eval model')
# default
parser.add_argument('--momentum', type=float, default=0.9, help='momentum')
parser.add_argument('--label-smooth', type=float, default=0.1, help='label smoothing')
parser.add_argument('--save-interval', type=int, default=5, help='report frequency for saving trained models')
parser.add_argument('--train-dir', type=str, default='data/train', help='path to training dataset')
parser.add_argument('--val-dir', type=str, default='data/val', help='path to validation dataset')
parser.add_argument('--num-trials', type=int, default=2, help='num_trials')
# total_iters=7800, # 1560, 3120, 6240, 390, 780, 39000, 1950, 19500, 195*36=7020, 195*40=7800
parser.add_argument('--total-iters', type=int, default=1560, help='total iters')
# shuffle 256, mobile 288, cifar-fast 64
parser.add_argument('--batch-size', type=int, default=64, help='batch size')
# network_shuffle: 4 choice for 5 blocks (Shuffle3x3\Shuffle5x5\Shuffle7x7\Xception)
# network_mobile: 1/2/3 choice for 12 layers {'conv': [1, 2], 'rate': 1},
parser.add_argument('--choice', type=int, default=4, help='choice')
# shuffle block 5, mobile 12 layers, cifar_fast 3 layers
parser.add_argument('--block', type=int, default=3, help='block')
# shuffle ,1/2/3/4, mobiel, sample_path=ag.space.Int(1, 2, 3),
parser.add_argument('--sample_path', type=int, default=2, help='sample_path')
# learning_rate=ag.space.Real(0.1, 0.3, log=True), # glboal hpo, mobile
# learning_rate=ag.space.Real(0.4, 0.8, log=True), # glboal hpo, shuffle
# learning_rate=ag.space.Real(0.01, 0.2, log=True), # glboal hpo, fast
# parser.add_argument('--learning-rate', type=float, default=0.5, help='init learning rate')
parser.add_argument('--lr-range', type=str, default='0.01,0.2', help='learning-rate range. default is 0.01,0.2.')
# parser.add_argument('--weight-decay', type=float, default=4e-5, help='weight decay')
parser.add_argument('--wd-range', type=str, default='4e-5,5e-3', help='weight-decay range. default is 4e-5,5e-3.')
args = parser.parse_args()
return args
if __name__ == "__main__":
arg = get_args()
arg.local = os.path.split(os.path.realpath(__file__))[0]
# Select A Best Supernet
@ag.args(
# Configuration
signal=arg.signal,
different_hpo=arg.different_hpo,
eval=arg.eval,
auto_continue=arg.auto_continue,
# update
flops_restriction=False, # spos
activations_count_restriction=False, # pycls
RandA=False, # RA(RandAugment)
choice=arg.choice,
block=arg.block,
sample_path=arg.sample_path,
total_iters=arg.total_iters,
batch_size=arg.batch_size,
# Supernet HPO, Some tricks work for specific hyperparameters
# fake=ag.space.Real(0.4, 0.8, log=True),
# lr & wd & randaug
# learning_rate=0.4, # test_lr_range # without hpo
learning_rate=ag.space.Real(arg.lr_range.split(',')[0], arg.lr_range.split(',')[1], log=True), # glboal hpo, fast
lr_range = arg.lr_range, # different hpo
# weight_decay=ag.space.Real(4e-5, 5e-3, log=True),
weight_decay=ag.space.Real(arg.wd_range.split(',')[0], arg.wd_range.split(',')[1], log=True), # glboal hpo, fast
# randaug_n=ag.space.Int(3, 4, 5),
# randaug_m=ag.space.Int(5, 10, 15),
# default hyperparameters
randaug_n=3,
randaug_m=5,
momentum=arg.momentum,
label_smooth=arg.label_smooth,
save_interval=1,
# save='./models',
# weight_decay=4e-5,
)
def ag_train_cifar(args, reporter):
return pipeline(args, reporter)
myscheduler = ag.scheduler.FIFOScheduler(ag_train_cifar,
# resource={'num_cpus': 4, 'num_gpus': 1},
num_trials=arg.num_trials,
time_attr='epoch',
reward_attr="val_acc")
myscheduler.run()
myscheduler.join_jobs()
myscheduler.get_training_curves(filename='{}/save/{}/Supernet_curves'.format(arg.local, arg.signal), plot=True, use_legend=True)
print('The Best Supernet Configuration and Accuracy are: {}, {}'.format(myscheduler.get_best_config(),
myscheduler.get_best_reward()))Our search result is in $Link/Search/checkpoint.pth.tar, which can be used by Evaluation.
Out searched models have been trained from scratch, is can be found in $Link/Evaluation/$ARCHITECTURE.
Here is a summary:
- Cifar10:
| Architecture | FLOPs | #Params | Top-1 Acc | Top-5 Acc |
|---|---|---|---|---|
| (2, 1, 0, 1, 2) | 9M | 1.4M | 96.4 | 99.8 |
- Imagenet:
| Architecture | FLOPs | #Params | Top-1 | Top-5 |
|---|---|---|---|---|
| (2, 1, 0, 1, 2, 0, 2, 0, 2, 0, 2, 3, 0, 0, 0, 0, 3, 2, 3, 3) | 323M | 3.5M | 25.6 | 8.0 |
Download the ImageNet Dataset and move validation images to labeled subfolders. To do this, you can use the following script: https://github.com/raw/soumith/imagenetloader.torch/master/valprep.sh
Download the flops table to accelerate Flops calculation which is required in Uniform Sampling. It can be found in $Link/op_flops_dict.pkl.
We recommend to create a folder data and use it in both Supernet training and Evaluation training.
Here is a example structure of data:
data
|--- train ImageNet Training Dataset
|--- val ImageNet Validation Dataset
|--- op_flops_dict.pkl Flops Table
Train supernet with the following command:
- Cifar10: train shuffle/mobile/resnet block with different_hpo/global_hpo
In my test, only cost 2min in a trial, reached at least 94% test set accuracy, Runtime for 1 epochs is roughly 15s by 64 batch-size.
cd src/Supernet_cifar
python3 train.py --num-classes 10 --signal different_hpo --different-hpo --num-trials 16 --total-iters 7800 --batch-size 64 --block 4 --lr-range "0.01,0.2" --wd-range "4e-5,5e-3"
python3 train.py --num-classes 10 --signal glboal_hpo --num-trials 16 --total-iters 7800 --batch-size 64 --block 4 --lr-range '0.01,0.2' --wd-range '4e-5,5e-3'
python3 train.py --num-classes 10 --signal without_hpo --num-trials 1 --total-iters 7800 --batch-size 64 --block 4 --lr-range "0.2,0.201" --wd-range "4e-5,5e-3"
---
python3 train.py --num-classes 10 --signal without_hpo --num-trials 2 --total-iters 7800 --batch-size 64 --block 3 --lr-range "0.2,0.201" --wd-range "4e-5,5e-3"
# cifar100
python3 train.py --num-classes 100 --signal cifar100_without_hpo --num-trials 1 --total-iters 7800 --batch-size 64 --block 4 --lr-range "0.2,0.201" --wd-range "4e-5,5e-3"
- Imagenet:
cd src/Supernet
python3 train.py --train-dir $YOUR_TRAINDATASET_PATH --val-dir $YOUR_VALDATASET_PATHSearch in supernet with the following command:
cd src/Search
python3 search.pyIt will use ../Supernet/checkpoint-latest.pth.tar as Supernet's weight, please make sure it exists or modify the path manually.
Get searched architecture with the following command:
cd src/Evaluation
python3 eval.pyIt will generate folder in data/$YOUR_ARCHITECTURE. You can train the searched architecture from scratch in the folder.
Finally, train and evaluate the searched architecture with the following command.
Train:
cd src/Evaluation/data/$YOUR_ARCHITECTURE
python3 train.py --train-dir $YOUR_TRAINDATASET_PATH --val-dir $YOUR_VALDATASET_PATHEvaluate:
cd src/Evaluation/data/$YOUR_ARCHITECTURE
python3 train.py --eval --eval-resume $YOUR_WEIGHT_PATH --train-dir $YOUR_TRAINDATASET_PATH --val-dir $YOUR_VALDATASET_PATH