/deep-n-cheap

AutoML framework balancing performance and complexity

Primary LanguagePythonMIT LicenseMIT

deep-n-cheap DnC logo

Welcome

This repository implements Deep-n-Cheap – an AutoML framework to search for deep learning models. Features include:

  • Complexity oriented: Get models with good performance and low training time or parameter count
  • Cuttomizable search space for both architecture and training hyperparameters
  • Supports CNNs and MLPs

Highlight: State-of-the-art performance on benchmark and custom datasets with training time orders of magnitude lower than competing frameworks and NAS efforts.

Research paper available on arXiv. Please consider citing it if you use or benefit from this work.

How to run?

$ pip install sobol_seq tqdm
$ git clone https://github.com/souryadey/deep-n-cheap.git
$ cd deep-n-cheap
$ python main.py

For help:

$ python main.py -h

Complexity customization

Set wc to high values to penalize complexity at the cost of performance:

  • --wc=0: Performance oriented search
  • --wc=0.1: Balance performance and complexity
  • --wc=10: Complexity oriented search
  • Any non-negative value of wc is supported!

Datasets (including custom)

Set dataset to either:

  • --dataset=torchvision.datasets.<dataset>. Currently supported values of <dataset> = MNIST, FashionMNIST, CIFAR10, CIFAR100
  • --dataset='<dataset>.npz', where <dataset> is a .npz file with 4 keys:
    • xtr: numpy array of shape (num_train_samples, num_features...), example (50000,3,32,32) or (60000,784). Image data should be in channels_first format.
    • ytr: numpy array of shape (num_train_samples,)
    • xte: numpy array of shape (num_test_samples, num_features...)
    • yte: numpy array of shape (num_test_samples,)
  • Some datasets can be downloaded from the links in dataset_links.txt. Alternatively, define your own custom datasets.

Examples

  1. Search for CNNs between 4-16 layers on CIFAR-10, train each for 100 epochs, run Bayesian optimization for 15 prior points and 15 steps. Optimize for performance only. Estimated search cost: 30 GPU hours
python main.py --network 'cnn' --dataset torchvision.datasets.CIFAR10 --input_size 3 32 32 --output_size 10 --wc 0 --numepochs 100 --bo_prior_states 15 --bo_steps 15 --num_conv_layers 4 16
  1. Search for CNNs between 5-10 layers on Fashion MNIST without augmentation, max channels in any conv layer is 256, search for batch sizes from 64 to 128 and dropout drop probabilities in [0.1,0.2]. Optimize for fast training by using a high wc. Download the dataset to the parent directory.
python main.py --network 'cnn' --data_folder '../' --dataset torchvision.datasets.FashionMNIST --input_size 1 28 28 --output_size 10 --augment False --wc 1 --num_conv_layers 5 10 --channels_upper 256 --batch_size 64 128 --drop_probs_cnn 0.1 0.2
  1. Search for MLPs on the custom Reuters RCV1 dataset, which has 2000 input features and 50 output classes. Search between 0-2 hidden layers, moderately penalize parameter count, search for initial learning rates from 1e-1 to 1e-4, run each model for 20 epochs. Use half data for validation.
python main.py --network 'mlp' --dataset 'rcv1_2000.npz' --val_split 0.5 --input_size 2000 --output_size 50 --wc 0.05 --penalize 'numparams' --numepochs 20 --num_hidden_layers 0 2 --lr -4 -1
  1. Search for CNNs on the custom Reuters RCV1 dataset reshaped to 5 channels of 20x20 pixels each and saved as 'rcv1_2000_reshaped.npz'. Use default CNN search settings.
python main.py --network 'cnn' --dataset 'rcv1_2000_reshaped.npz' --input_size 5 20 20 --output_size 50

Some results from the research paper

  • >91% accuracy on CIFAR-10 in 9 hrs of searching with a model taking 3 sec/epoch to train.
  • >95% accuracy on Fashion-MNIST in 17 hrs of searching with a model taking 5 sec/epoch to train.
  • >91% accuracy on Reuters RCV1 in 2 hrs of searching with a model with 1M parameters.

Contact

Deep-n-Cheap is developed and maintained by the USC HAL team