torchsig.models.spectrogram_models.detr.detr.DETR

torchsig.models.spectrogram_models.detr.detr.DETR(version: str, num_classes: int = 1, drop_rate_backbone: float = 0.2, drop_path_rate_backbone: float = 0.2, drop_path_rate_transformer: float = 0.1)

Constructs a DETR architecture with an EfficientNet-B0 backbone and an XCiT-Nano transformer. DETR from “End-to-End Object Detection with Transformers”. EfficientNet from “EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks”. XCiT from “XCiT: Cross-Covariance Image Transformers”.

Parameters:

• version (str) – Which DETR model to load.

• num_classes (int) – Number of output classes; if loading checkpoint and number does not equal 1, final layer will not be loaded from checkpoint

• drop_path_rate_backbone (float) – Backbone drop path rate for training

• drop_rate_backbone (float) – Backbone dropout rate for training

• drop_path_rate_transformer (float) – Transformer drop path rate for training