Many existing NAS methods can only get one network structure during the search iteration, and cannot take into account various hardware constraints, like power, performance, latency etc. In addition, although the evolutionary algorithm-based NAS method achieves good performance, each generation of samples needs to be retrained for evaluation, which greatly affects the search efficiency. Considering the insufficiency of the existing NAS methods, we propose a continuous evolution approach called Efficient Neural Architecture Search (CARS). CARS maintains an optimal model solution set, and updates the parameters of the supernet through the model in the solution set. In each iteration of evolutionary algorithm, current samples can directly inherit the parameters from the supernet and parent samples, which effectively improves the efficiency of evolution. CARS can obtain a series of models with different sizes and precisions in one search. Users can select the corresponding model based on the resource constraints in the actual application. For details about the CARS algorithm, see HYPERLINK "https://arxiv.org/abs/1909.04977" https://arxiv.org/abs/1909.04977.
The following figure shows the principle of the CARS algorithm. The CARS algorithm maintains a supernet, and each subnet is a sample of the supernet. In the process of evolutionary algorithm, the weight of all the subnets is shared with the supernet, so that the samples of the sub-generation can inherit the weight of the samples of the parent generation directly, thus realizing continuous evolution. CARS algorithm uses pNSGA - III algorithm to protect large models in the searching process and increase the coverage of the searching model. The CARS algorithm uses the subnets in the Pareto frontier to update the weight of supernet. The weight update of supernet and the evolutionary update of subnets are performed alternately.
The CARS in the pipeline integrates the search space of DARTS. The overall structure is as follows:
For a detailed description of the DARTS search space, see the corresponding HYPERLINK "https://arxiv.org/abs/1806.09055" ICLR '19 article.
The search space configuration is as follows:
search_space:
type: SearchSpace
modules: ['super_network']
super_network:
name: CARSDartsNetwork
network: ['PreOneStem',
'normal', 'normal', 'reduce',
'normal', 'normal', 'reduce',
'normal', 'normal',
]
input_size: 32
init_channels: 16
num_classes: 10
auxiliary: False
search: True
normal:
type: 'block'
name: 'Cell'
steps: 4
reduction: False
genotype:
[
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 2, 0 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 2, 1 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 3, 0 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 3, 1 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 3, 2 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 4, 0 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 4, 1 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 4, 2 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 4, 3 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 5, 0 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 5, 1 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 5, 2 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 5, 3 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 5, 4 ],
]
concat: [2, 3, 4, 5]
reduce:
type: 'block'
name: 'Cell'
steps: 4
reduction: True
genotype:
[
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 2, 0 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 2, 1 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 3, 0 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 3, 1 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 3, 2 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 4, 0 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 4, 1 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 4, 2 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 4, 3 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 5, 0 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 5, 1 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 5, 2 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 5, 3 ],
[ ['none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'], 5, 4 ],
]
concat: [2, 3, 4, 5]
preprocess:
name: 'darts_stem1'
linear:
name: 'linear'By default, the CARS uses the CIFAR-10 data set. You can also use customized data sets. To use datasets in user-defined format, you need to implement dataset interface that meet Vega requirements. For details, see the development manual.
The configuration information of the CIFAR-10 database is as follows:
dataset:
type: Cifar10
common:
data_path: /cache/datasets/cifar10/
num_workers: 8
train_portion: 0.5
drop_last: False
train:
shuffle: True
batch_size: 128
val:
batch_size: 3500For details about how to search and train a model, see the following configuration file for parameter setting:
- vega/examples/nas/cars/cars.yml
The configuration file is directly transferred to the pipeline through main.py. The two processes are performed in sequence. During the search process, a series of models at the Pareto Front are found. During the training process, the selected models are fully trained to obtain the final performance.
The configuration of the search process is as follows:
nas:
pipe_step:
type: NasPipeStep
search_algorithm:
type: CARSAlgorithm
policy:
num_individual: 128
start_ga_epoch: 50
ga_interval: 10
select_method: uniform #pareto
warmup: 50
trainer:
type: Trainer
darts_template_file: "{default_darts_cifar10_template}"
callbacks: CARSTrainerCallback
model_statistics: False
epochs: 500
optim:
type: SGD
lr: 0.025
momentum: 0.9
weight_decay: !!float 3e-4
lr_scheduler:
type: CosineAnnealingLR
T_max: 500
eta_min: 0.001
loss:
type: CrossEntropyLoss
metric:
type: accuracy
grad_clip: 5.0
seed: 10
unrolled: TrueConfiguration of the fully train phase:
pipe_step:
type: FullyTrainPipeStep
models_folder: "{local_base_path}/output/nas/"
trainer:
ref: nas.trainer
callbacks: DartsFullTrainerCallback
epochs: 600
lr_scheduler:
type: CosineAnnealingLR
T_max: 600.0
eta_min: 0
loss:
type: MixAuxiliaryLoss
loss_base:
type: torch.nn.CrossEntropyLoss
aux_weight: 0.4
seed: 100
drop_path_prob: 0.2
dataset:
type: Cifar10
common:
data_path: /cache/datasets/cifar10/
num_workers: 8
drop_last: False
batch_size: 96
train:
shuffle: True
transforms:
- type: RandomCrop
size: 32
padding: 4
- type: RandomHorizontalFlip
- type: ToTensor
- type: Normalize
mean:
- 0.49139968
- 0.48215827
- 0.44653124
std:
- 0.24703233
- 0.24348505
- 0.26158768
- type: Cutout
length: 8
test:
shuffle: FalseThe models_folder and model_desc_n.json parameters in the fully train are described as follows:
- models_folder: specifies a directory where one or more model_desc_n.json files are stored. n indicates the sequence number. This parameter is mutually exclusive with model_desc_file. If both of them are set, model_desc_file are preferred.
- model_desc_file: specifies a model description file as the input of the trainer.
The output:
- nas directory: outputs multiple optimal model description files.
- fully_train: outputs the trained model weight file.
Benchmark configuration: cars.yml