Skip to content
This repository has been archived by the owner on Jul 10, 2024. It is now read-only.

Latest commit

 

History

History
48 lines (36 loc) · 3.14 KB

README.md

File metadata and controls

48 lines (36 loc) · 3.14 KB
Raw

Setting parameters and configurations

Please set the following in the finetuning.yaml file:

  • num_workers: number of sub-processes or threads to use for data loading. Setting the argument num_workers as a positive integer will turn on multi-process data loading. (Default=32)

  • precision: precision of data type in which model to be fine-tuned. Choices are [float32, bfloat16]

  • fine_tune: set 'True' to run SimSiam or CutPaste self-supervised learning using Intel Transfer Learning Tool APIs. Set 'False' to run a pre-trained backbone by providing a model path under 'model_path' category

  • output_path: path to save the checkpoints or final model

  • tlt_wf_path: set by default to point to the workflow in the Intel Transfer Learning Tool

  • dataset:

    • root_dir: path to the root directory of MVTEC dataset
    • category_type: category type within MVTEC dataset, e.g.: hazelnut or all (for running all categories in MVTEC)
    • batch_size: batch size for inference (Default=32)
    • image_size: each image resized to this size (Default=224x224)

  • model: Options to select when running with a pre-trained backbone, no fine-tuning on custom dataset

    • name: pretrained backbone model E.g.: resnet50, resnet18
    • layer: intermediate layer from which features will be extracted
    • pool: pooling kernel size for average pooling
    • feature_extractor: select the type of modelling and subsequent feature extractor. Options are:
      • pretrained - No fine-tuning on custom dataset, features will be extracted from pretrained model which is set in model/name
      • simsiam - SimSiam self-supervised training on custom dataset
      • cutpaste - CutPaste self-supervised training on custom dataset

  • simsiam: Set when 'feature_extractor' is set to simsiam. For details about simsiam method, please refer to https://arxiv.org/abs/2011.10566

    • batch_size: batch size for fine-tuning (Default=64)
    • epochs: number of epochs to fine-tune the model
    • optim: optimization algorithm E.g.: sgd, adam
    • model_path: path to save the checkpoints or final model
    • ckpt: flag to specify whether intermediate checkpoints should be saved or not

  • cutpaste: Set when 'feature_extractor' is set to cutpaste. For details about cutpaste method, please refer to https://arxiv.org/abs/2104.04015

    • cutpaste_type: type of image augmentation for cutpaste fine-tuning, choices are ['normal', 'scar', '3way', 'union'].
    • head_layer: number of fully-connected layers on top of average pooling layer followed by the last linear layer of backbone network
    • freeze_resnet: number of epochs till only head layers will be trained. After this, complete network will be trained.
    • batch_size: batch size for fine-tuning (Default=64)
    • epochs: number of epochs to fine-tune the model
    • optim: optimization algorithm E.g.: sgd, adam
    • model_path: path to save the checkpoints or final model
    • ckpt: flag to specify whether intermediate checkpoints should be saved or not

  • pca_thresholds: percentage of variance ratio to be retained. Number of PCA components are selected according to it