Load_graph does not work for mobilenet_v2 for Object Detection on TAO (Detectnet_V2)

Please provide the following information when requesting support.
detectnet_v2_train_peoplenet_r34_3C.txt (5.3 KB)

• Hardware (T4/V100/Xavier/Nano/etc) - XavierNX
• Network Type (Detectnet_v2/Faster_rcnn/Yolo_v4/LPRnet/Mask_rcnn/Classification/etc) - Detectnet_v2
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here) - 3.22.02

Hello everyone,
I was training an object detection model using mobilenet_v2 as the architecture. I would like to know why the load_graph option does not work when training the network. In the 1st run of training if I set the load_graph = True, then I get the error saying - layer output_cov not defined (Not sure whether this is the name but the error is very similar to this)

If I set the load_graph = False then I get no error when I start the training. But on the other hand, I suspect that the pre-trained weights are not loaded. However, when I set the first_validation _epoch = 1, then I get mAP as 0 for all the classes.

Following is the spec file I am using for training mobilenet_v2:
detectnet_v2_train_peoplenet_mv2_3C.txt (5.2 KB)

If I use resnet34 as the architecture and set load_graph = True, the pre-trained weights are loaded in the 1st run of training. I do not have any error and the result for the first_validation _epoch = 1 is mAP value of 28.3.

Following is the spec file I am using for training resnet34:
detectnet_v2_train_peoplenet_r34_3C.txt (5.3 KB)

I would like to know this difference in behavior when using mobilenet_v2 and resnet34 and the possible solution.

Looking forward to your replies !!

The load_graph is a flag to determine whether or not to load the graph from the pretrained model file, or just the weights. For a pruned model, set this parameter to True. Pruning modifies the original graph, so the pruned model graph and the weights need to be imported.

In your mobilenet_v2 spec, the pretrained model is
pretrained_model_file: “/workspace/tao-experiments/detectnet_v2/pretrained_mobilenetv2/pretrained_detectnet_v2_vmobilenet_v2-1/mobilenet_v2.hdf5”

In resnet34 spec, it is another model.
pretrained_model_file: “/workspace/tao-experiments/detectnet_v2/tlt_peoplenet_vunpruned_v2.1/resnet34_peoplenet.tlt”

So, that is different.

Hello @Morganh,
Thank you for your reply. So does it mean that with the hdf5 model we cannot use the load_graph option?


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