Detectnet_v2 Resnet34 pretrained model not copying weights for conv layers

Please provide the following information when requesting support.

• Hardware - multiple gpus
• Network Type - Detectnet_v2
• TAO Version - nvcr.io/nvidia/tao/tao-toolkit:5.0.0-tf1.15.5

I am training a resnet34 object detection model using a custom dataset. While training I noticed in the logs that the weights for the pretrained resnet34 model (pulled from ngc TAO Pretrained Detectnet V2) were not being set specifically in all of the conv layers (see output below). I have created a modified spec file that causes the same issue as a minimum reproduction.

detectnet_v2_train_resnet34.txt (3.0 KB)

train_log.txt (57.8 KB)

Please try to load resnet18 or resnet50 instead from TAO Pretrained DetectNet V2 | NVIDIA NGC. There are no similar behavior.
For resnet34, we will check further. Thanks.

1 Like