get_started.md

Commence usage

For initial usage, please follow the steps below sequentially.

Search for the model architecture.

• Directories configs/classification and configs/detection respectively contain search configuration files for classification and detection, as demonstrated below.

• Taking a classification task as an example, search for the CNN architecture of ResNet-50 and execute the following steps:

  sh tools/dist_search.sh configs/classification/R50_FLOPs.py
  #or 
  python tools/search.py configs/classification/R50_FLOPs.py

Export the search results.

• tools/export.py, A script for exporting the searched model architecture and its related dependencies is provided, allowing for quick verification of the demo

• For example R50_FLOPs：

  python tools/export.py save_model/R50_R224_FLOPs41e8 output_dir

  Copy the demo deployment and related code to the output_dir/R50_R224_FLOPs41e8/ directory, which should include the following contents:

  • best_structure.json：Several optimal model architectures that were found during the search.
  • demo.py：A simple script demonstrating how to use the models
  • cnnnet.py：The class definitions and utility functions used for constructing the models
  • modules： The foundational modules of the models.
  • weights/：Several optimal model weights that were found during the search (only available for one-shot NAS methods).

Using the searched architecture.

• demo.py is a basic usage example, but you can also run demo.py directly after exporting the model architecture in the previous step.

• Continuing with the ResNet-50 architecture for a classification task as an example, the core code is explained below:

  • Import dependencies

  import ast
  from cnnnet import CnnNet

  • Load the optimal structure from a file."

  with open('best_structure.json', 'r') as fin:
      content = fin.read()
      output_structures = ast.literal_eval(content)
  
  network_arch = output_structures['space_arch']
  best_structures = output_structures['best_structures']

  • Instantiate the classification backbone network.

  network_id = 0    # Index number. Multiple structures can be output during the search, as set by num_network=5 in example_cls_res50.sh.
  out_indices = (4, )    # Output stage. For classification tasks, only the output from the final stage needs to be obtained.
  
  backbone = CnnNet(
      structure_info=best_structures[network_id],
      out_indices=out_indices,
      num_classes=1000,
      classification=True,
      )
  backbone.init_weight(pretrained)

  • You can now fully utilize the backbone 😄

• For further usage methods of the CNN detection task model, please refer to tinynas/deploy/