Skip to content

Latest commit

 

History

History
282 lines (210 loc) · 16.8 KB

examples.md

File metadata and controls

282 lines (210 loc) · 16.8 KB
Raw

Examples Reference

Vega provides guidance on how to use algorithms and tasks, and provides guidance on algorithm development for developers, such as expanding the search space and search algorithm and building datasets applicable to Vega.

1. Example List

In the following example, the following directory is generated after the package is decompressed:

Contents Description
compression compression algorithm usage example, including Quant-EA、 Prune-EA.
data augmentation Example of using the data augmentation algorithm, including the PBA, CycleSR
hpo hyperparameter optimization algorithm use examples, including ASHA, BO, TPE, BOHB, BOSS, Random, Random Pareto.
nas Examples of network architecture search, including CARS, SP-NAS, auto-lane, SR-EA, ESR-EA, Adelaide-EA, NAGO, BackboneNas, DartsCNN, FIS, GDAS, MFKD, SegmentationEA, SGAS, ModuleNas, DNet-NAS
fully train Samples related to fully train, including the EfficientNet B0/B4 model training example and FMD operator example.
classification An example of using NAS + HPO + FullyTrain to complete an image classification task
features Cluster, custom dataset, model evaluation, quota samples.

2. Run Example

2.1 Run PyTorch Example

Generally, an algorithm example contains a configuration file, some algorithms, and some matching code.

You can execute the following commands in the examples directory:

vega <algorithm config file>

For example, the command of CARS example is as follows:

vega ./nas/cars/cars.yml

All information is stored in configuration files. Configuration items are classified into public configuration items and algorithm configuration items. For details about public configuration items, see the configuration reference. For details about algorithm configuration, see the reference documents of each algorithm.Error! Hyperlink reference not valid.

Before running an example, you need to configure the directory where the dataset is located in the algorithm configuration file. The root directory of the default dataset is /cache/datasets/, for example, the directory of the Cifar10 is /cache/datasets/cifar10/.

Before running the example, you need to download the dataset to the default data configuration directory. Before running the example, you need to create the directory /cache/datasets/, then download each dataset to the directory and unzip it. The default directory configuration of each dataset is as follows:

example Pre-trained Model Default Path Model Source
adelaide_ea mobilenet_v2-b0353104.pth /cache/models/mobilenet_v2-b0353104.pth download
BackboneNas (mindspore) resnet50-19c8e357.pth /cache/models/resnet50-19c8e357.pth download
BackboneNas (tensorflow), classification, prune_ea(tensorflow) resnet_imagenet_v1_fp32_20181001 /cache/models/resnet_imagenet_v1_fp32_20181001/
keep only these files: checkpoint, graph.pbtxt, model.ckpt-225207.data-00000-of-00002, model.ckpt-225207.data-00001-of-00002, model.ckpt-225207.index, model.ckpt-225207.meta		 download
dnet_nas 031-_64_12-1111-11211112-2.pth /cache/models/031-_64_12-1111-11211112-2.pth download
prune_ea(pytorch) resnet20.pth /cache/models/resnet20.pth download
prune_ea(mindspore) resnet20.ckpt /cache/models/resnet20.ckpt download
sp_nas fasterrcnn_resnet50_fpn_coco-258fb6c6.pth /cache/models/fasterrcnn_resnet50_fpn_coco-258fb6c6.pth download

In the configuration file of each example, the platform description (PyTorch, TensorFlow, and MindSpore) applicable to the example is provided in general/backend.

For example, the following configuration indicates that the sample can run on three platforms:

general:
    backend: pytorch # pytorch | tensorflow | mindspore

The following configurations can run only in TensorFlow:

general:
    backend: tensorflow

2.2 Run TensorFlow Example

  1. Command line (GPU):

    vega <algorithm config file> -b t

    for example:

    vega ./nas/backbone_nas/backbone_nas.yml -b t

  2. Command line (Atlas 900):

    vega <algorithm config file> -b t -d NPU

    for example:

    vega ./nas/backbone_nas/backbone_nas.yml -b t -d NPU

2.3 Run MindSpore Example

Command line (Atlas 900):

vega <algorithm config file> -b m -d NPU

for example:

vega ./nas/backbone_nas/backbone_nas.yml -b m -d NPU

3. Examples' Input and Output

3.1 Model Compression

  1. Prune-EA

    Stage Option Content
    nas Input Config File: compression/prune-ea/prune.yml
    Pre-Trained Model: /cache/models/resnet20.pth
    Dataset: /cache/datasets/cifar10
    nas Output Network Description File: tasks/<task id>/output/nas/model_desc_<id>.json
    nas approximate running time (random_samples + num_generation * num_individual) * epochs / Number of GPUs * Training time per epoch
    fully train Input Config File: compression/prune-ea/prune.yml
    Network Description File: tasks/<task id>/output/nas/model_desc_<id>.json
    Dataset: /cache/datasets/cifar10
    fully train Output Model: tasks/<task id>/output/fully_train/model_<id>.pth
    fully train approximate running time epochs * Training time per epoch

  2. Quant-EA

    Stage Option Content
    nas Input Config File: compression/quant-ea/quant.yml
    Dataset: /cache/datasets/cifar10
    nas Output Network Description File: tasks/<task id>/output/nas/model_desc_<id>.json
    nas approximate running time (random_samples + num_generation * num_individual) * epochs / Number of GPUs * Training time per epoch
    fully train Input Config File: compression/quant-ea/quant.yml
    Network Description File: tasks/<task id>/output/nas/model_desc_<id>.json
    Dataset: /cache/datasets/cifar10
    fully train Output Model: tasks/<task id>/output/fully_train/model_<id>.pth
    fully train approximate running time epochs * Training time per epoch

3.2 NAS

  1. CARS

    Stage Option Content
    nas Input Config File: nas/cars/cars.yml
    Dataset: /cache/datasets/cifar10
    nas Output Network Description File: tasks/<task id>/output/nas/model_desc_<id>.json
    nas approximate running time epochs * Training time per epoch (The training time is affected by num_individual)
    fully train Input Config File: nas/cars/cars.yml
    Network Description File: tasks/<task id>/output/nas/model_desc_<id>.json
    Dataset: /cache/datasets/cifar10
    fully train Output Model: tasks/<task id>/output/fully_train/model_<id>.pth
    fully train approximate running time epochs * Training time per epoch

  2. Adelaide-EA

    Stage Option Content
    random Input Config File: nas/adelaide_ea/adelaide_ea.yml
    Dataset: /cache/datasets/cityscapes
    random Output Network Description File: tasks/<task id>/output/random/model_desc_<id>.json
    random approximate running time max_sample * epochs / Number of GPUs * Training time per epoch
    mutate Input Config File: nas/adelaide_ea/adelaide_ea.yml
    Dataset: /cache/datasets/cityscapes
    Network Description File: tasks/<task id>/output/random/model_desc_<id>.json
    mutate Output Network Description File: tasks/<task id>/output/mutate/model_desc_<id>.json
    mutate approximate running time max_sample * epochs / Number of GPUs * Training time per epoch
    fully train Input Config File: nas/adelaide_ea/adelaide_ea.yml
    Network Description File: tasks/<task id>/output/mutate/model_desc_<id>.json
    Dataset: /cache/datasets/cityscapes
    fully train Output Model: tasks/<task id>/output/fully_train/model_<id>.pth
    fully train approximate running time epochs * Training time per epoch

  3. ESR-EA

    Stage Option Content
    nas Input Config File: nas/esr_ea/esr_ea.yml
    Dataset: /cache/datasets/DIV2K
    nas Output Network Description File: tasks/<task id>/output/nas/selected_arch.npy
    nas approximate running time num_generation * num_individual * epochs / Number of GPUs * Training time per epoch
    fully train Input Config File: nas/esr_ea/esr_ea.yml
    Network Description File: tasks/<task id>/output/nas/selected_arch.npy
    Dataset: /cache/datasets/DIV2K
    fully train Output Model: tasks/<task id>/output/fully_train/model_<id>.pth
    fully train approximate running time epochs * Training time per epoch

  4. SR-EA

    Stage Option Content
    random Input Config File: nas/sr_ea/sr_ea.yml
    Dataset: /cache/datasets/DIV2K
    random Output Network Description File: tasks/<task id>/output/random/model_desc_<id>.json
    random approximate running time num_sample * epochs / Number of GPUs * Training time per epoch
    mutate Input Config File: nas/sr_ea/sr_ea.yml
    Dataset: /cache/datasets/DIV2K
    Network Description File: tasks/<task id>/output/random/model_desc_<id>.json
    mutate Output Network Description File: tasks/<task id>/output/mutate/model_desc_<id>.json
    mutate approximate running time num_sample * epochs / Number of GPUs * Training time per epoch
    fully train Input Config File: nas/sr_ea/sr_ea.yml
    Network Description File: tasks/<task id>/output/mutate/model_desc_<id>.json
    Dataset: /cache/datasets/DIV2K
    fully train Output Model: tasks/<task id>/output/fully_train/model_<id>.pth
    fully train approximate running time epochs * Training time per epoch

  5. SP-NAS

    Stage Option Content
    nas1 Input Config File: nas/sp_nas/spnas.yml
    Dataset: /cache/datasets/COCO2017
    Pre-Trained Model: /cache/models/resnet50-19c8e357.pth
    Config File: nas/sp_nas/faster_rcnn_r50_fpn_1x.py
    nas1 Output Network Description File: tasks/<task id>/output/nas1/model_desc_<id>.json
    Model List: tasks/<task id>/output/total_list_p.csv
    nas1 approximate running time max_sample * epochs / Number of GPUs * Training time per epoch
    nas2 Input Config File: nas/sp_nas/spnas.yml
    Dataset: /cache/datasets/COCO2017
    Network Description File: tasks/<task id>/output/nas1/model_desc_<id>.json
    Model List: tasks/<task id>/output/total_list_p.csv
    Config File: nas/sp_nas/faster_rcnn_r50_fpn_1x.py
    nas2 Output Network Description File: tasks/<task id>/output/nas2/model_desc_<id>.json
    Model List: tasks/<task id>/output/total_list_s.csv
    nas2 approximate running time max_sample * epochs / Number of GPUs * Training time per epoch
    fully train Input Config File: nas/sp_nas/spnas.yml
    Dataset: /cache/datasets/COCO2017
    Network Description File: tasks/<task id>/output/nas2/model_desc_<id>.json
    Model List: tasks/<task id>/output/total_list_s.csv
    Config File: nas/sp_nas/faster_rcnn_r50_fpn_1x.py
    fully train Output Model: tasks/<task id>/output/fullytrain/model_<id>.pth
    fully train approximate running time epochs * Training time per epoch

  6. Auto_Lane

    Stage Option Content
    nas Input Config File：nas/auto_lane/auto_lane.yml
    Dataset：/cache/datasets/CULane OR /cache/datasets/CurveLane
    nas Output Network Description File：tasks/<task id>/output/nas/model_desc_<id>.json
    nas approximate running time max_sample * epoch / Numbers of GPUs * Training time per epoch
    fully train Input Config File：nas/sp_nas/auto_lane.yml
    Dataset：/cache/datasets/CULane OR /cache/datasets/CurveLane
    Network Description File：tasks/<task id>/output/nas/model_desc_<id>.json
    fully train Output Model：tasks/<task id>/output/fullytrain/model_<id>.pth
    fully train approximate running time epochs * Training time per epoch

  7. AutoGroup

    Stage Option Content
    fully train Input Config File nas/fis/autogroup.yml
    Dataset: /cache/datasets/avazu
    fully train Output Network Description File: tasks/<task id>/output/mutate/model_desc_<id>.json
    Model: tasks/<task id>/output/fully_train/model_0.pth
    fully train approximate running time epochs * Training time per epoch

  8. AutoFis

    Stage Option Content
    search Input Config File nas/fis/autogate_grda.yml
    Dataset: /cache/datasets/avazu
    search Output Model: tasks/<task id>/output/search/0/model.pth
    search approximate running time epochs * Training time per epoch
    retrain Input Config File nas/fis/autogate_grda.yml
    Dataset: /cache/datasets/avazu
    retrain Output Model: tasks/<task id>/output/retrain/0/model.pth
    retrain approximate running time epochs * Training time per epoch

3.3 Data Augmentation

  1. PBA

    Stage Option Content
    pba Input Config File: data_augmentation/pba/pba.yml
    Dataset: /cache/datasets/cifar10
    pba Output Transformer List: tasks/<task id>/output/pba/best_hps.json
    pba approximate running time total_rungs * each_epochs * config_count / Number of GPUs * Training time per epoch

  2. CycleSR

    Stage Option Content
    fully train Input Config File: data_augmentation/cyclesr/cyclesr.yml
    Dataset: /cache/datasets/DIV2K_unpair
    fully train Output Model: tasks/<task id>/output/fully_train/model_0.pth
    fully train approximate running time n_epoch * Training time per epoch

3.4 HPO

  1. ASHA、BOHB、BOSS

    Stage Option Content
    hpo Input Config File: hpo/asha|bohb|boss/hpo/asha|bohb|boss.yml
    Dataset: /cache/datasets/cifar10
    hpo Output Hyperparameter file: tasks/<task id>/output/hpo/best_hps.json

3.5 Fully Train

  1. EfficientNet

    Stage Option Content
    fully train Input Config File: fully_train/efficientnet/efficientnet_b0.yml
    Dataset: /cache/datasets/ILSVRC
    fully train Output Network Description File: tasks/<task id>/output/mutate/model_desc_<id>.json
    Model: tasks/<task id>/output/fully_train/model_<id>.pth
    fully train approximate running time epochs * Training time per epoch

  2. FMD

    Stage Option Content
    fully train Input Config File: fully_train/fmd/fmd.yml
    Dataset: /cache/datasets/cifar10
    fully train Output Network Description File: tasks/<task id>/output/mutate/model_desc_<id>.json
    Model: tasks/<task id>/output/fully_train/model_0.pth
    fully train approximate running time epochs * Training time per epoch

  3. ResNet

    Stage Option Content
    fully train Input Config File fully_train/trainer/resnet.yml
    Dataset: /cache/datasets/ILSVRC
    fully train Output Network Description File: tasks/<task id>/output/mutate/model_desc_<id>.json
    Model: tasks/<task id>/output/fully_train/model_0.pth
    fully train approximate running time epochs * Training time per epoch