v0.3.0

simplify vortex implementation and use Pytorch Lightning for trainer.

Added

• Pytorch Lightning training pipeline
• registry class
• model base class
• add 'config.trainer.validation_interval' field with same functionality as 'config.validator.val_epoch'
• add 'config.trainer.logger' field with same functionality as 'config.logging'
• add 'config.dataloader.train' and 'config.dataloader.eval' field to support different args for train and val dataloader
• support for bool value for logger config, with default logger use tensorboard logger
• backbone base class
• backbone class for main models

Changed

• model_components is removed, and changed with model base class
• deprecating 'config.seed', and move to 'config.trainer.seed'
• deprecating stage argument in create_model
• all defined backbones use backbone base class
• removed various old features affected by API redesign: cli, predictor, pipelines

v0.2.1

This is the last version that will use our own training engine. In the next release (v0.3) we'll be using Pytorch Lightning for training engine and model creation API.

Added

• learning rate scheduler value plan visualizer script in scripts/visualize_learning_rate.py

• save_best_metrics config to save model checkpoint on best metrics

• always save checkpoint of last epoch model

• save best_metrics value in model checkpoint

• support for experimental DETR model, this model is still unable to be exported caused by the limitation in current exporter design

• support for per parameter options (see PyTorch Optimizer) using 'param_groups' key in model components.

• jit and export global context, see vortex/development/networks/modules/utils/config.py

• support for per parameter options (see PyTorch Optimizer)
  using 'param_groups' key in model components.

• support for None additional_input shape in model for scalar tensor input.

• support for image size type of list (w,h), or non-square image, in predict and export.

• validation loss calculation to experiment logger

• support for changing batch norm layer in all backbone with norm_layer argument.

• new backbone models: CSPNet, ResNest, RexNet, RegNet, TResNet. see more details in spreadsheet or csv summary

• darknet weight converter in scripts/convert_darknet_weight.py

• yolov3 converted darknet weight and config file

• get model path from config.checkpoint if found and valid for predict and export

• yolov3 support for changing backbone stage to output with backbone_stages argument.

• BasePredictionPipeline argument to plot visualized result.

• new predict and ir_runtime_predict argument to not visualise and not save prediction result.

• new list command in main cli to list available components

• shape inference for onnx export

Changed

• model checkpoint not save on hyperparameter optimization
• save_epoch config is not required
• Fix bug on scheduler.step() placement on trainer
• Update docs on supported Pytorch scheduler
• changed StepLRWithBurnIn to StepLRWithWarmUp scheduler
• changed CosineLRScheduler to CosineLRWithWarmUp scheduler
• changed TanhLRScheduler to TanhLRWithWarmUp scheduler
• BaseTrainer.create_optimizer only accept model parameters dict (e.g. from model.parameters()) instead of the model itself
• make class_label for detection not required
• update backbone components
• use mobilenetv3 definition and pretrained from rwightman
• improved cli help message
• improved progress bar, on train also display some metrics when training

Fixed

• Fix error when using ir_runtime_validate with uneven batch splitting caused by different batch size on the last batch sample
• Fix error when save_best_metrics not present in experiment file
• Fix ONNX graph ops to convert iou_threshold as input
• lr scheduler bug when arguments changed and resumed
• yolov3 model definition

Release v0.2.0

Added

• Added new classification metrics for validation :

  • roc_curve
  • precision
  • recall
  • f1-score

• new checkpoint (saved) model format
  checkpoint model is loaded as a dictionary with:

  • default (required) member in checkpoint:

    • epoch --> number of epochs the model have been trained
    • state_dict -> trained model's state dict, this is the same as the entire data as the old format
    • optimizer_state --> optimizer state in training
    • config --> configuration used to train the model

  • additional (optional) member (if any):

    • class_names --> model's output class names map
    • metrics --> per epoch training metrics
    • scheduler_state -> state_dict for trainer's lr scheduler

• support for resume training in training pipeline using checkpoint field in config file and --resume flag in train command.

• checkpoint field in configuration file to point to the checkpoint model used to resume training if --resume flag is given. init_state_dict is still possible to be used but will automatically load state_dict to model even if --resume flag is not defined, but removed from docs.

• model checkpoint update script in script/update_model.py

• Now possible to access class_names and input_specs attributes from both PytorchPredictionPipeline.model and IRPredictionPipeline.model

• config deprecation checks for new config format

• Support for adding external model from user space

• Added DALIDataLoader and nvidia_dali augmentation module

  • nvidia_dali augmentation module must be used with DALIDataLoader
  • DALIDataLoader can use other augmentation module, however if nvidia_dali augmentation module if specified in the experiment file, it must be in the first order
  • DALIDataLoader utilize ray library to paralelize external module augmentation for each batch sample

Changed

• reading class_names attribute from checkpoint in prediction and export pipeline
• class_names in dataset, prediction, and export is optional. If not specified or None, will create a numbered class label [class_0, class_1, ..., class_n]
• class_names moved from PytorchPredictionPipeline.class_names to PytorchPredictionPipeline.model.class_names ( also applied to IRPredictionPipeline)
• None in string to normal python None
• new config format, as described in #7.
  • change dataset name field dataset to name in config.dataset.train and config.dataset.eval.
  • move device from config.trainer to main config.
  • move dataloader field from config.dataset to main config.
  • change dataloader module from dataloader to module in config.dataloader.
  • change default dataloader module name from DataLoader to PytorchDataLoader in config.dataloader.module.
  • change scheduler to lr_scheduler in config.trainer.
  • change field validation to validator and move it from config.trainer to main config.
• Refactor DatasetWrapper into BasicDatasetWrapper and DefaultDatasetWrapper
• Change output type of BaseRuntime __call__ method to list of orderedDict
• It is now possible to override image auto padding in dataset object by adding and set self.disable_image_auto_pad = True attribute on the collate_fn object provided by the model_components
• Refactor package into 2 namespace packages : vortex.runtime and vortex.development. In which case vortex.development is depend on vortex.runtime, but vortex.runtime can be installed as a standalone package for minimal requirement inferencing library

Initial Release

Added

• Python package installation
• Command line interface
• Initial documentation
• Base development pipeline : training, validation, export, predict, ir-predict, ir-validation, hypopt
• Modular design
• Integration with image augmentation library ( Albumentations )
• Integration with hyperparameter optimization library ( Optuna )
• Integration with 3rd party experiment logger ( Comet.ml )
• Graph export to Torchscript and ONNX
• Visual report of model's performance and resource usage, see this example
• Various architecture support
  • 50 + infamous backbone networks
  • Classification and Detection architecture support