Arulkumar03/Fox_Sheep_Detector_Computer_Vision_model

1# Rethinking "Batch" in BatchNorm2 3We provide configs that reproduce detection experiments in the paper [Rethinking "Batch" in BatchNorm](https://arxiv.org/abs/2105.07576).4 5All configs can be trained with:6 7```8../../tools/lazyconfig_train_net.py --config-file configs/X.py --num-gpus 89```10 11## Mask R-CNN12 13* `mask_rcnn_BNhead.py`, `mask_rcnn_BNhead_batch_stats.py`:14  Mask R-CNN with BatchNorm in the head. See Table 3 in the paper.15 16* `mask_rcnn_BNhead_shuffle.py`: Mask R-CNN with cross-GPU shuffling of head inputs.17  See Figure 9 and Table 6 in the paper.18 19* `mask_rcnn_SyncBNhead.py`: Mask R-CNN with cross-GPU SyncBatchNorm in the head.20  It matches Table 6 in the paper.21 22## RetinaNet23 24* `retinanet_SyncBNhead.py`: RetinaNet with SyncBN in head, a straightforward implementation25  which matches row 3 of Table 5.26 27* `retinanet_SyncBNhead_SharedTraining.py`: RetinaNet with SyncBN in head, normalizing28  all 5 feature levels together. Match row 1 of Table 5.29 30The script `retinanet-eval-domain-specific.py` evaluates a checkpoint after recomputing31domain-specific statistics. Running it with32```33./retinanet-eval-domain-specific.py checkpoint.pth34```35on a model produced by the above two configs, can produce results that match row 4 and36row 2 of Table 5.37