Retraining a pruned model has low AP, precision, recall and RPN recall

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc) A30
• Network Type : FasterRCNN

Retraing a pruned model and the results are very low in AP, precision, recall and RPN recall.
The FasterRCNN model has very high AP, precision, recall and RPN recall. The model is pruned using the following command.

faster_rcnn prune -m /workspace/Nyan/tao_source_codes_v5.0.0/notebooks/tao_launcher_starter_kit/faster_rcnn/results/ver_2/unpruned/frcnn_resnet_50.epoch_50.hdf5 \
                  -o /workspace/Nyan/tao_source_codes_v5.0.0/notebooks/tao_launcher_starter_kit/faster_rcnn/results/ver_2/pruned/frcnn_resnet_50_pruned.hdf5 \
                  -k nvidia_tao \
                  -eq union \
                  -pth 0.7

After that the pruned model is retrained with the following command.

faster_rcnn train -e /workspace/Nyan/tao_source_codes_v5.0.0/notebooks/tao_launcher_starter_kit/faster_rcnn/specs/default_spec_resnet50_retrain.txt \
                  -r /workspace/Nyan/tao_source_codes_v5.0.0/notebooks/tao_launcher_starter_kit/faster_rcnn/results/ver_2/retrained \
                  -k nvidia_tao \
                  --gpus 1

But I have very low accuracy in validation.

==========================================================================================
Class               AP                  precision           recall              RPN_recall
------------------------------------------------------------------------------------------
nohelmet            0.0000              0.0000              0.0000              0.0008
------------------------------------------------------------------------------------------
withhelmet          0.0000              0.0002              0.0333              0.0750
------------------------------------------------------------------------------------------

The following spec file is used for retraining.

default_spec_resnet50_retrain.txt (3.7 KB)

What did I do wrong?

Higher pth will give you smaller model (and thus higher inference speed) but worse accuracy. If the retrain accuracy is good, you can increase this value to get smaller models. Otherwise, lower this value to get better accuracy.
You can set to lower pth and retry, for example, 0.2.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.