Model - Overview

Netspresso Trainer provides a variety of backbones and heads, allowing flexible combinations. Users can choose appropriate backbones and heads based on their dataset and task requirements. The models can be optimized for on-device environments using Netspresso's compression and converting services.

We provide a configuration format which can easily construct backbones and heads to meet user requirements. As composed in the example of the ResNet50 model below, backbones and heads are structured as separate fields, and then connected. Also, Users can freely choose suitable loss modules suitable for the head and task. The range of supported models and the detailed configuration definitions for each model are extensively described in the separated Models page.

model:
  task: classification
  name: resnet50
  checkpoint:
    use_pretrained: True
    load_head: False
    path: ~
    fx_model_path: ~
    optimizer_path: ~
  freeze_backbone: False
  architecture:
    full: ...
    backbone: ...
    neck: ...
    head: ...
  losses:
    - criterion: cross_entropy
      label_smoothing: 0.1
      weight: ~

Retraining the model from NetsPresso

If you have compressed model from NetsPresso, then it's time to retrain your model to get the best performance. Netspresso Trainer uses the same configuration format for retraining torch.fx GraphModule. This can be executed by specifying the path to the torch.fx model in the fx_model_checkpoint field. Since the torch.fx model file contains the complete model definition, fields like architecture become unnecessary, can be ignored.

model:
  task: classification
  name: resnet50
  checkpoint:
    use_pretrained: # This field will be ignored since fx_model_path is activated
    load_head: # This field will be ignored since fx_model_path is activated
    path: # This field will be ignored since fx_model_path is activated
    fx_model_path: ./path_to_your_fx_model.pt
    optimizer_path: # This field will be ignored since fx_model_path is activated
  freeze_backbone: False
  architecture: # This field will be ignored since fx_model_path is activated
  losses:
    - criterion: cross_entropy
      label_smoothing: 0.1
      weight: ~

Field list

Field	Description
model.task	(str) We support "classification", "segmentation", and "detection" now.
model.name	(str) A nickname to identify the model.
model.checkpoint.use_pretrained	(bool) Whether to use the pretrained checkpoint. At first time, you will download and save the pretrained checkpoint.
model.checkpoint.load_head	(bool) Whether to use the pretrained checkpoint for head module.
model.checkpoint.path	(str) Checkpoint path to resume training. If None and use_pretrained is False, you can train you model from scratch.
model.checkpoint.optimizer_path	(str) Optimizer checkpoint path for resuming training.
model.checkpoint.fx_model_path	(str) Model path for fx model retraining. If you have to train the model from NP Compressor, you have to fill your compressed model path at this field. If fx_model_path is filled, use_pretrained, load_head, path, optimizer_path, and freeze_backbone are ignored.
model.freeze_backbone	(bool) Whether to freeze backbone in training.
model.architecture	(dict) Detailed configuration of the model architecture. Please see Model page to find NetsPresso supporting models.
model.losses	(list) List of losses that model to learn. Please see Losses page to find NetsPresso supporting loss modules.