Overview

Eztorch allows you to instantiate several kind of models as well as different parts of models. We will detail each of the different kind of customization you can have below.

Trunks

• Trunks
  • Image
    • ResNet and ResNext
    • Timm
  • Video
    • Pytorchvideo
    • Video model and Head wrapper
    • ResNet 3D with basic blocks
    • R2+1D
    • S3D
    • X3D

Trunks, also called encoders, are the main parts of your model to learn representations. Eztorch allows you to instantiate trunks for Image or Video representation.

Heads

• Heads
  • Linear
    • LinearHead
  • Linear 3D
    • Linear3DHead
    • create_linear3d_head()
  • ResNet 3D head
    • VideoResNetHead
    • create_video_resnet_head()
  • MLP
    • MLPHead

Heads are fixed on top of trunks to specialize to a specific task. Among different heads we can cite:

• Linear heads to perform classification.

• MLP heads to make projectors in Self-Supervised Learning.

Siamese models

• Siamese
  • Base classes
    • SiameseBaseModel
    • MomentumSiameseBaseModel
    • ShuffleMomentumSiameseBaseModel
    • ShuffleMomentumQueueBaseModel
  • Self-Supervised Models
    • SimCLR
    • MoCo
    • MoCo v3
    • ReSSL
    • SCE
    • SCE Tokens

Siamese models are models based on a siamese architecture. It is generally used for Self-Supervised or/and Knowledge-distillation training. It means that a model is used twice in a pipeline. Sometimes the model is duplicated to have a self-distillation paradigm through a momentum update of the duplication of the model.

Supervised model

• Supervised
  • SupervisedModel

The supervised model works based on a trunk and a head.

Finetuning model

• Finetuning
  • FinetuningModel
  • SoccerNetSpottingModel
  • SpottingModel

The finetuning model works based on a pretrained trunk and a head and has several parameters trick implemented for finetuning.

Dummy model

• Dummy

The dummy model is a model implemented just to make some quick pipeline tests.