Transfer learning not working - yolov3 - tao toolkit

Accelerated Computing Intelligent Video Analytics TAO Toolkit
, ,
#1

Please provide the following information when requesting support.

• Hardware (T4/V100/Xavier/Nano/etc) : X86_64 GPU Machine
• Network Type (Detectnet_v2/Faster_rcnn/Yolo_v4/LPRnet/Mask_rcnn/Classification/etc) : YOLOv3
• TLT Version (Please run “tlt info --verbose” and share “docker_tag” here)
• Training spec file(If have, please share here):
training_config.txt (2.0 KB)

• How to reproduce the issue ? (This is for errors. Please share the command line and the detailed log here.)

Transfer learning is not working as expected,
Steps followed:

  1. Model: TAO Pretrained Object Detection | NVIDIA NGC

  2. Dataset: The PASCAL Visual Object Classes Challenge 2012 (VOC2012)

  3. Trained with model(refer traininig_config file attached) {classes: bike,bottle,car,cycle,dog,person}

  4. Inferencing is happening and predicting the above classes.
    -----Transfer learning-----

  5. Again trained the model only with custom microcontroller datasets.
    training_config.txt (2.3 KB)

  6. Model is predicting the newly trained microcontrollers but not predicting the previously trained weights {car,dog,bike etc…}

Question:

  • Is there any steps i need to follow for transfer learning?(model should predict already trained classes along with newly added classes)
  • Any configuration i need to update to retain/load pretarined weights?
#3

Does this custom microcontroller datasets contain car/dog/bike ? If not, need to add car/dog/bike dataset since the training_config.txt contains these classes.