Yolov4_tiny with darknet19/53

Hi ,

Does yolov4_tiny support darknet19/53 ?

Yes.

Well, It is giving an unsupported backbone error.

ValueError: Unsupported backbone model: darknet19

Primary job terminated normally, but 1 process returned
a non-zero exit code. Per user-direction, the job has been aborted.


mpirun.real detected that one or more processes exited with non-zero status, thus causing
the job to be terminated. The first process to do so was:

Sorry, after checking, for YOLOv4-tiny, only cspdarknet_tiny and cspdarknet_tiny_3l are supported. The former arch has two heads(stride 8 and 16), while the latter has three heads(stride 8, 16, and 32).

More info is in YOLOv4-tiny — TAO Toolkit 3.21.11 documentation

For cspdarknet_tiny_3l I am unable to download the model using the ngc command.

±------±------±------±------±------±------±------±------±------+
| Versi | Accur | Epoch | Batch | GPU | Memor | File | Statu | Creat |
| on | acy | s | Size | Model | y Foo | Size | s | ed |
| | | | | | tprin | | | Date |
| | | | | | t | | | |
±------±------±------±------±------±------±------±------±------+
| vgg19 | 77.56 | 80 | 1 | V100 | 153.7 | 153.7 | UPLOA | Aug |
| | | | | | | 2 MB | D_COM | 18, |
| | | | | | | | PLETE | 2021 |
| vgg16 | 77.17 | 80 | 1 | V100 | 113.2 | 113.1 | UPLOA | Aug |
| | | | | | | 6 MB | D_COM | 18, |
| | | | | | | | PLETE | 2021 |
| squee | 65.13 | 80 | 1 | V100 | 6.5 | 6.46 | UPLOA | Aug |
| zenet | | | | | | MB | D_COM | 18, |
| | | | | | | | PLETE | 2021 |
| resne | 77.91 | 80 | 1 | V100 | 294.2 | 294.2 | UPLOA | Aug |
| t50 | | | | | | MB | D_COM | 18, |
| | | | | | | | PLETE | 2021 |
| resne | 77.04 | 80 | 1 | V100 | 170.7 | 170.6 | UPLOA | Aug |
| t34 | | | | | | 5 MB | D_COM | 18, |
| | | | | | | | PLETE | 2021 |
| resne | 76.74 | 80 | 1 | V100 | 89.0 | 88.96 | UPLOA | Aug |
| t18 | | | | | | MB | D_COM | 18, |
| | | | | | | | PLETE | 2021 |
| resne | 77.78 | 80 | 1 | V100 | 576.3 | 576.3 | UPLOA | Aug |
| t101 | | | | | | 3 MB | D_COM | 18, |
| | | | | | | | PLETE | 2021 |
| resne | 74.38 | 80 | 1 | V100 | 38.3 | 38.31 | UPLOA | Aug |
| t10 | | | | | | MB | D_COM | 18, |
| | | | | | | | PLETE | 2021 |
| mobil | 72.75 | 80 | 1 | V100 | 5.0 | 5.01 | UPLOA | Aug |
| enet_ | | | | | | MB | D_COM | 18, |
| v2 | | | | | | | PLETE | 2021 |
| mobil | 79.5 | 80 | 1 | V100 | 26.2 | 26.22 | UPLOA | Aug |
| enet_ | | | | | | MB | D_COM | 18, |
| v1 | | | | | | | PLETE | 2021 |
| googl | 77.11 | 80 | 1 | V100 | 47.6 | 47.64 | UPLOA | Aug |
| enet | | | | | | MB | D_COM | 18, |
| | | | | | | | PLETE | 2021 |
| effic | 77.11 | 80 | 1 | V100 | 26.8 | 26.78 | UPLOA | Aug |
| ientn | | | | | | MB | D_COM | 18, |
| et_b1 | | | | | | | PLETE | 2021 |
| _swis | | | | | | | | |
| h | | | | | | | | |
| effic | 77.11 | 80 | 1 | V100 | 26.8 | 26.78 | UPLOA | Aug |
| ientn | | | | | | MB | D_COM | 18, |
| et_b1 | | | | | | | PLETE | 2021 |
| _relu | | | | | | | | |
| darkn | 76.44 | 80 | 1 | V100 | 311.7 | 311.6 | UPLOA | Aug |
| et53 | | | | | | 8 MB | D_COM | 18, |
| | | | | | | | PLETE | 2021 |
| darkn | 77.52 | 80 | 1 | V100 | 152.8 | 152.8 | UPLOA | Aug |
| et19 | | | | | | 2 MB | D_COM | 18, |
| | | | | | | | PLETE | 2021 |
| cspda | 77.1 | 80 | 1 | V100 | 28.6 | 28.57 | UPLOA | Nov |
| rknet | | | | | | MB | D_COM | 23, |
| _tiny | | | | | | | PLETE | 2021 |
| cspda | 76.44 | 80 | 1 | V100 | 103.0 | 102.9 | UPLOA | Sep |
| rknet | | | | | | 9 MB | D_COM | 10, |
| 53 | | | | | | | PLETE | 2021 |
| cspda | 77.52 | 80 | 1 | V100 | 62.9 | 62.86 | UPLOA | Sep |
| rknet | | | | | | MB | D_COM | 10, |
| 19 | | | | | | | PLETE | 2021 |
±------±------±------±------±------±------±------±------±------+

Could you share your command? Thanks.

what command?
Are you asking for this one !ngc registry model list nvidia/tao/pretrained_object_detection:*? OR

!ngc registry model download-version nvidia/tao/pretrained_object_detection:cspdarknet_tiny_3l
–dest $LOCAL_EXPERIMENT_DIR/pretrained_cspdarknet_tiny_3l

I need to only download the “cspdarknet_tiny_3l” architecture using the above command but how?

Thanks for the info. I will check.

when you try to list all the object detection models using “!ngc registry model list nvidia/tao/pretrained_object_detection:*” … " cspdarknet_tiny_3l" is not even displayed

You can directly use cspdarknet_tiny.hdf5.

wget https://api.ngc.nvidia.com/v2/models/nvidia/tao/pretrained_object_detection/versions/cspdarknet_tiny/files/cspdarknet_tiny.hdf5

or

!ngc registry model download-version nvidia/tao/pretrained_object_detection:cspdarknet_tiny

I have it for “cspdarknet_tiny”, I need it for " cspdarknet_tiny_3l"

Yes, you can use the same pretrained model for cspdarknet_tiny_3l.

No there is no model for " cspdarknet_tiny_3l", you can try and list the models from your end and check.
No such model.

Are you saying that I can use the “cspdarknet_tiny” model for " cspdarknet_tiny_3l" too?

Yes, I also find that there is not cspdarknet_tiny_3l.hdf5 in ngc.
But as mentioned you can use cspdarknet_tiny.hdf5 instead.

I can use the same “cspdarknet_tiny” model for both the yolov4_tiny arch but the only difference between the two are the strides right?

Only heads are different.

okay, Thank you!

What image resolution does Yolov4_tiny support? Should it be a square image or else multiple of 16?