TV-detection Retinanet

[adam-peaston-SC]

Source / repo

https://github.com/pytorch/vision/tree/main/references/detection

Model description

Retinanet with ResNet101 FPN backbone

Dataset

COCO

Literature benchmark source

ResNet101 https://arxiv.org/abs/1708.02002

ResNet50 https://pytorch.org/vision/main/models/generated/torchvision.models.detection.retinanet_resnet50_fpn.html

Literature benchmark performance

ResNet101 BBox [AP, AP50, AP75, AP-S, AP-M, AP-L] BBox [39.1, 59.1, 42.3, 21.8, 42.7, 50.2]

ResNet50 box_map (on COCO-val2017) [36.4]

Strong Compute result achieved

ResNet101 BBox [26.8, 45.9, 25.9, 16.3, 33.0, 39.8]

ResNet50 [NA]

Basic training config (as applicable)

Nodes: [N] Epochs: [N] Effective batch size: [N] Learning rate: [L] Optimizer: [OPT]

Logs gist

[URL]

[adam-peaston-SC]

Source repo used ResNet50 FPN backbone which successfully completes training. Best reference benchmarks from academic literature were for ResNet101 FPN backbone so attempted to train based on ResNet101 FPN backbone for benchmarking. Numeric instability continues to frustrate this effort. Investigation required to determine the cause of numeric instability (potentially due to focal loss calculation) and harden.