Learn to Build, Train, and Export RetinaNet (ResNet-50) on a Custom Dataset

You’ll fine-tune RetinaNet (ResNet-50 + FPN) from TF Model Garden on the BCCD dataset. Steps: convert COCO JSON to TFRecords via the official CLI; load the retinanet_resnetfpn_coco experiment and adjust input size, classes, and training params; set a distribution strategy; build a Task and run train_and_eval with mAP metrics; monitor in TensorBoard; then export a SavedModel that accepts uint8 images and perform inference/visualization with visualization_utils. The notebook covers the full loop: data → config → training → evaluation → export → predictions.