Retraining peoplenet model for detecting face and person only

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Hello i am trying to retrain peoplenet model (resnet34_peoplenet.tlt) but i am getting assertion error during the training .

i am following this blog https://developer.nvidia.com/blog/training-custom-pretrained-models-using-tlt/ as reference .

here is the spec file i have used and label file
label file :-

person 0.0 2 0.0 215 137 357 500 0.00 0.00 0.00 0.00 0.00 0.00 0.00
person 0.0 2 0.0 478 409 625 636 0.00 0.00 0.00 0.00 0.00 0.00 0.00
face 0.0 2 0.0 243 134 300 192 0.00 0.00 0.00 0.00 0.00 0.00 0.00
face 0.0 2 0.0 480 414 574 505 0.00 0.00 0.00 0.00 0.00 0.00 0.00

tfrecord spec file :- con.txt (294 Bytes)

output of tf record convert :- ut.txt (4.7 KB)

Traning spec file :- train_v2.txt (4.0 KB)

for training i have also tried to add bag in the spec file but that too not working

command used for training :- tlt-train detectnet_v2 -e train_v2.txt -r output/ -k tlt_encode -n peoplenet_resnet34

output from training command :- output.txt (31.3 KB)

dataset structure :- |--dataset_root

|-- images
|-- 000000.jpg
|-- 000001.jpg
|-- xxxxxx.jpg
|-- labels
|-- 000000.txt
|-- 000001.txt
|-- xxxxxx.txt

Please let me know what am i doing wrong .

3

If train for two classes, please set only two classes in the training spec. Please delete all the info about the third class.

4

Hello @Morganh , thank you for the reply , i have tried with two classes as well .

Here is the training spec file for that :- train_v2.txt (3.3 KB)

still got the same error .

5

Hello @Morganh, i found the issue in my training spec file bbox_rasterizer_config was missing after i added that training is running .

thanks for the help .