Semantic Segmentation

Implementation of Deeplabv3+ with W&B logging

Several backbones available:

• ResNet (resnet50, resnet101)
• Swin-Transformer (swinT, swinS, swinB, swinL)
• ConvNeXt (convnextT, convnextS, convnextB, convnextL, convnextXL)

Note that swinB is the base model from Swin Transformers where instead swinT, swinS and swinL have a dimension and complexity of about 0.25x, 0.5x and 2x of swinB. Note that swinT is comparable to resnet50 and swinS to resnet101 complexity-wise. ConvNeXt if from here.

Pretrained models on ImageNet for the swin transformers backbones can be downloaded here and placed inside backbone_checkpoints in the main folder.

Several learning rate schedulers available:

• step (stepLR)
• cosine (CosineAnnealingLR)
• poly (PolyLR)