Retrain pruned model for Frcnn in TLT

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

Trying to retrain the pruned model and my spec file is attached.retrain_spec_resnet34.log (3.6 KB)

According to the tutorial here, I need to set regularizer type is ‘NO_REG’. But I have error

 File "/usr/local/bin/tlt-train-g1", line 8, in <module>
    sys.exit(main())
  File "./common/magnet_train.py", line 33, in main
  File "./faster_rcnn/scripts/train.py", line 48, in main
  File "./faster_rcnn/spec_loader/spec_wrapper.py", line 36, in __init__
  File "./faster_rcnn/spec_loader/spec_wrapper.py", line 575, in validate_spec
  File "./faster_rcnn/spec_loader/spec_wrapper.py", line 756, in validate_training_config
AssertionError: 
        Regularizer can only be L1(l1), or L2(l2), got no_reg

Why I have error in Frcnn?

3

Could you try “none”?
See https://docs.nvidia.com/metropolis/TLT/tlt-getting-started-guide/index.html#training_configuration

The regularizer type. Can be either ‘L1’(L1 regularizer), ‘L2’(L2 regularizer), or ‘none’(No regularizer).

4

I still have error as

Traceback (most recent call last):
  File "/usr/local/bin/tlt-train-g1", line 8, in <module>
    sys.exit(main())
  File "./common/magnet_train.py", line 33, in main
  File "./faster_rcnn/scripts/train.py", line 48, in main
  File "./faster_rcnn/spec_loader/spec_wrapper.py", line 36, in __init__
  File "./faster_rcnn/spec_loader/spec_wrapper.py", line 575, in validate_spec
  File "./faster_rcnn/spec_loader/spec_wrapper.py", line 756, in validate_training_config
AssertionError: 
        Regularizer can only be L1(l1), or L2(l2), got none

I am retraining pruned model for Frcnn with TLT docker version 2.0.

5

Sorry, unfortunately, TLT 2.0_dp docker does not support no regularizer.
TLT 1.0 supports no regularizer.

For temp workaround, in TLT 2.0 training spec, please set a small L2 reg like 1e-10 to essentially disable it for now.
We will fix it in next release.

6

Thank you