Skip to content

Latest commit

 

History

History
60 lines (49 loc) · 5.28 KB

models.md

File metadata and controls

60 lines (49 loc) · 5.28 KB

Models

Each model in the SparseZoo has a specific stub that identifies it. The stubs are made up of the following structure:

DOMAIN/SUB_DOMAIN/ARCHITECTURE{-SUB_ARCHITECTURE}/FRAMEWORK/REPO/DATASET{-TRAINING_SCHEME}/SPARSE_NAME-SPARSE_CATEGORY-{SPARSE_TARGET}

The properties within each model stub are defined as the following:

Model Property Definition Examples
DOMAIN The type of solution the model is architected and trained for cv, nlp
SUB_DOMAIN The sub type of solution the model is architected and trained for classification, segmentation
ARCHITECTURE The name of the guiding setup for the network's graph resnet_v1, mobilenet_v1
SUB_ARCHITECTURE (optional) The scaled version of the architecture such as width or depth 50, 101, 152
FRAMEWORK The machine learning framework the model was defined and trained in pytorch, tensorflow_v1
REPO The model repository the model and baseline weights originated from sparseml, torchvision
DATASET The dataset the model was trained on imagenet, cifar10
TRAINING_SCHEME (optional) A description on how the model was trained augmented, lower_lr
SPARSE_NAME An overview of what was done to sparsify the model base, pruned, quant (quantized), pruned_quant, arch (architecture modified)
SPARSE_CATEGORY Descriptor on the degree to which the model is sparsified as compared with the baseline metric none, conservative (100% baseline), moderate (>= 99% baseline), aggressive (< 99%)
SPARSE_TARGET (optional) Descriptor for the target environment the model was sparsified for disk, edge, deepsparse, gpu

The contents of each model are made up of the following:

  • training: The directory containing checkpoints. Every checkpoint contains a set of files required to load a model in the specific state (e.g. directly after pruning). The checkpoint stores the trained model in the native ML framework in which it was originally trained such as PyTorch, Keras, Tensorflow V1.
  • deployment: The directory containing all the files necessary for the deployment of the model within an inference pipeline.
  • logs: The directory containing the artifacts generated during the training flow that helps to track the reproducibility and audibility. Optional directory.
  • model.onnx: The ONNX representation of the model's graph.
  • onnx: The directory to store different opset representations of the model.onnx. Optional directory.
  • model.md: The model card containing metadata, descriptions, and information for the model.
  • benchmarks.yaml: The information about the performance of the model.onnx on given hardware systems. Optional file.
  • metrics.yaml: Reporting metrics such as accuracy for the model on the given datasets such as validation and training. Optional file.
  • recipes: The directory containing the recipes - the original sparsification recipe (recipe_original.md) or others(e.g. transfer learning recipe).
  • sample_originals: The original sample data without any pre-processing for use with the model.
  • sample_inputs: The sample data after pre-processing for use with the model.
  • sample_outputs: The outputs after running the sample inputs through the model.
  • sample_labels: The labels that classify the sample inputs.
  • sample_originals: The unedited data that can be used as inputs to a training pipeline (images, text files, numpy arrays, etc).