Tao Training Detectnet_v2 custom dataset : Average precision value 0.0000%

[steevep]

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc)
• Network Type (Detectnet_v2)
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here)
nvidia/tao/tao-toolkit:
5.0.0-tf2.11.0:
• Training spec file(If have, please share here)
detectnet_v2_train_resnet18_kitti.txt (5.9 KB)

My Tfrecord log
tfrecord_log.txt (12.6 KB)

my evaluation part
Validation cost: 0.000095
Mean average_precision (in %): 0.0000

±-----------------±-------------------------+
| class name | average precision (in %) |
±-----------------±-------------------------+
| animal | 0.0 |
| autorickshaw | 0.0 |
| bicycle | 0.0 |
| bus | 0.0 |
| car | 0.0 |
| caravan | 0.0 |
| motorcycle | 0.0 |
| person | 0.0 |
| rider | 0.0 |
| traffic_light | 0.0 |
| traffic_sign | 0.0 |
| trailer | 0.0 |
| train | 0.0 |
| truck | 0.0 |
| vehicle_fallback | 0.0 |
±-----------------±-------------------------+

Median Inference Time: 0.017611
2024-05-24 08:57:55,294 [TAO Toolkit] [INFO] main 218: Evaluation complete.
Time taken to run main:main: 0:01:34.650363.
Execution status: PASS

and my input image ratio is width : 1920 and height is 1080 in spec file the output image size is different will it make a issue
and my image height is 1080 it not a multiple of 16 will it make a issue should i want to resize my image manually or putting 1088 is it fine?

how i can fix this please help me

[Morganh]

@steevep

and my image height is 1080 it not a multiple of 16 will it make a issue should i want to resize my image manually or putting 1088 is it fine?

It is not needed since you already set enable_auto_resize to true.

@steevep

detectnet_v2_train_resnet18_kitti.txt (5.9 KB)

target_class_mapping {
key: “standard”
value: “standard”
}
target_class_mapping {
key: “cracked”
value: “cracked”
}
target_class_mapping {
key: “offset”
value: “offset”
}

Above 3 classes are not in your dataset. Can you double check if the spec is correct?

[steevep]

in target class mapping i given all my dataset labels name in it
what is standard , cracked and offset what does it mean

[Morganh]

I get “standard , cracked and offset” from your spec file detectnet_v2_train_resnet18_kitti.txt.

[yingliu]

There is no update from you for a period, assuming this is not an issue anymore. Hence we are closing this topic. If need further support, please open a new one. Thanks