tianbaochou / CNAS

CNAS

Automatic Convolutional Neural Architecture Search for Image Classification under Different Scenes

Our work is based on DARTS

1. Requirements

• Ubuntu14.04/16.04 or Window10 (Win7 may also support.)

• Python >= 3.6

• torch >= 0.4.1

• torchvision == 0.2.1

• seaborn (optional)

• pygraphviz (optional)

NOTE: pytorch 0.4.x also work ok, but will meet ConnectionResetError When use the second approximation.

Todo:

• Merge into the CNASV prototype
• Replace the backbone network by automatically search outer network
• Give up the cell-based architecture for designing more flexible search space
• Merge this work to CNASV ( Search-Train Prototype for Computer Vision CNASV)

2. Usage

pip3 install requirements.txt

Notice

if you use win10, and want to show the cell architecture with graph, you need install the pygraphviz and add $workdir$\\3rd_tools\\graphviz-2.38\\bin into environment path. Here $workdir$ is the custom work directory. such as E:\\workspace\\cnas

How to search:

python train_search.py --num-meta-node 4 --cutout --data '../data/cifar10'

How to evaluate the CNN architecture we searched

• run on multi gpus (gpu1, gpu2)

CUDA_VISIBLE_DEVICES=1,2  train_cnn.py --cutout  --multi-gpus

• run on single gpus we will find a max free memory gpus and run on it.

python train_cnn.py --cutout  --multi-gpus

Configure details

Notice: num-meta-node, use-sparse and train-portion invalid in evaluate stage and multi-gpus, auxiliary and auxuliary-weight invalid in search stage

• train-dataset ：the train dataset for search or train ('cifar10','cifar100', 'tiny-imagenet200')
• data : the train dataset path ('../data/cifar10')
• arch : the search arch name
• start-epoch: manual epoch number (0)
• batch-size: the batch size (64)
• num-meta-node: the number of intermediate nodes (4)
• learning-rate: init learning rate (0.025)
• learning-rate-min: minimize learning rate (0.003)
• arch-learning-rate: learning rate for arch encoding
• arch-weight-decay: weight decay for arch encoding
• momentum: momentum for sgd (0.9)
• weight-decay: weight decay (3e-4)
• epochs: num of training epochs (60)
• init-channels: num of init channels (16)
• image-channels: num of image channels (3)
• layers: total number of layers(cells) stacked for architecture in search stage
• model-path: path to save the model (use for restart)
• cutout: use cutout (False)
• cutout-length: cutout length (16)
• save: experiment name
• resume: path to latest checkpoint (default: none)
• grad-clip: gradient clipping (5)
• train-portion: portion of training data (0.5)
• sec-approx: use 2 order approximate validation loss (False)
• use-sparse: use sparse framework (False)
• multi-gpus: train network use multi-gpus
• auxiliary: use auxiliary tower (True)
• auxuliary-weight: weight for auxiliary loss (0.4)
• opt: optimizer (sgd)

3. Our works

• We add the cweight operation (squeeze-and-excitation) and channel shuffle operation and remove 5x5 operation.

• We allow to keep the none operation when derive the architecture and decode as sparse architecture.

• We found the weights of our operations introduced is much higher than any others! (show in heat map below!)