[go: nahoru, domu]
Skip to content

Latest commit

 

History

History
69 lines (42 loc) · 2.95 KB

Tutorial.md

File metadata and controls

69 lines (42 loc) · 2.95 KB
Raw

Tutorial

This page provides basic tutorials about the usage of iNAS. Please run the commands in the root path of iNAS.
In general, both the training and testing include the following steps:

  1. Prepare datasets.
  2. Modify config files. The config files are under the options folder. For more specific configuration information, please refer to Config.md.
  3. You may need to download pre-trained models. Please see ModelZoo
  4. Run commands. Use Training Commands, Searching Commands, Converting Commands and Testing Commands accordingly.

Contents

  1. Training Commands
    1. Single GPU Training
    2. Distributed (Multi-GPUs) Training
  2. Searching Commands 3. Benchmark Latency Lookup Table on Devices 4. Single GPU Searching
  3. Converting Commands
  4. Testing Commands

Training Commands

Single GPU Training

CUDA_VISIBLE_DEVICES=0 python iNAS/train.py -opt [config file]

Distributed Training

4 GPUs (default)

CUDA_VISIBLE_DEVICES=0,1,2,3 python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 iNAS/train.py -opt options/train/SOD/train_supernet_sod_4gpu_b10_e100_noaug.yml --launcher pytorch

Searching Commands

Benchmark Latency Lookup Table on Devices

Here, we provide a sample latency lookup table (Baidu Drive (f3au) | Google Drive) tested on Intel Core CPU. You can generate it yourself by following command:

Generate the latency lookup table template (lut_template.txt) based on our search space.

python scripts/build_latency_table/build_latency_lut_template.py

Use the template file, we can benchmark the component latency on different devices by running the following commands:

python scripts/build_latency_table/compute_lut_on_devices.py

For mobile phone, this command only generate jit models. We need to benchmark it on mobile phones by Pytorch Mobile. Key codes can be found in scripts/build_latency_table/mobile_scripts.

Single GPU Searching

CUDA_VISIBLE_DEVICES=0 \
python iNAS/search.py -opt options/search/SOD/search_iNAS_1gpu_iter10.yml

Converting Commands

We can convert the supernet weight for each stand-alone models (initialized by json configuration) using the following command:

CUDA_VISIBLE_DEVICES=0 \
python iNAS/search.py -opt options/convert/SOD/convert_sod.yml

Testing Commands

We benchmark the stand-alone model accuracy on different SOD test set by the following command:

CUDA_VISIBLE_DEVICES=0 \
python iNAS/test.py -opt options/test/SOD/test_standalone_sod.yml