YOLOV4 configs file

Accelerated Computing Intelligent Video Analytics TAO Toolkit
1

TensorRT Version7.2.1.6
Quadro RTX 5000 dual GPU
Driver Version: 455.23.05
CUDA Version: 11.1
Ubuntu 18.04
python 3.6
Yolo_v4

nvidia/tao/tao-toolkit-tf:
docker_registry: nvcr.io
docker_tag: v3.21.08-py3

I am training a custom yoloV4 model using transfer learning toolkit
I am facing few problems while building model

If you use your own dataset, you will need to run the code below to generate the best anchor shape

!tlt yolo_v4 kmeans -l $DATA_DOWNLOAD_DIR/training/label_2
-i $DATA_DOWNLOAD_DIR/training/image_2
-n 9
-x 1248
-y 384

The anchor shape generated by this script is sorted. Write the first 3 into small_anchor_shape in the config

file. Write middle 3 into mid_anchor_shape. Write last 3 into big_anchor_shape.

-x,-y are for shape of the image
here actually my data is of two different image shapes so how should i get the anchors

2.https://catalog.ngc.nvidia.com/orgs/nvidia/models/tlt_vehicletypenet

here in this documentation

Input

RGB Images of dimensions: 224 X 224 X 3 (W x H x C) Channel Ordering of the Input: NCHW, where N = Batch Size, C = number of channels (3), H = Height of images (224), W = Width of the images (224) Input scale: None Mean subtraction: [103.939, 116.779, 123.68]

so should i reshape my data according to this shape 224 X 224 X 3?

3.yolov4 gives many options for pretrained model is resnet18 is vehical-net?

3

For your case, please check which model size you want to train. See output_width and output_height you set in the training spec. Then, resize the bboxes in all the labels files. Then run kmeans to get the anchors.

The vehicleTypeNet is a classification model. It is not related to yolov4 model. May I know why you ask VehicleTypenet?

I cannot understand. May I know what are the “options” . The pretrained models in ngc are not related to vehicles. They are trained via OpenImage dataset.

1 Like
4

So i am gonna build a yoloV4 detection model for vehicles with 12 classes ,i have around 11k data, using transfer learning toolkit .

the Jupiter notebook YoloV4 consists the following option to download the pre trained model to build

!ngc registry model list nvidia/tlt_pretrained_object_detection:*

  1. vgg19
  2. vgg16
  3. squeezenet
  4. resnet50
  5. resnet34
  6. resnet18
  7. resnet101
  8. resnet10
  9. mobilenetv2
  10. mobilenetv1
  11. googlenet
  12. efficientnet_b1_swish
  13. efficientnet_b1_relu
  14. efficientnet_b0_swish
  15. efficientnet_b0_relu
  16. Darknet53
  17. Darknet19
  18. cspdarknet53
  19. cspdarknet19

which pretrained model is best for my use case?

5

These are pretrained models for different backbones. Please consider the combination of mAP and fps. It depends on your requirement. You can have a try for resnet18.

1 Like
6

Thanks

7

and what is the preferred amount of data required ?

8

There is no rule for the amount of data. More dataset will result in more training time. I think you can start with your 11k data.

1 Like
10

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.