Changing the GridSize of segnet

Autonomous Machines Jetson & Embedded Systems Jetson Xavier NX
1

How do I change the GridSize of segnet? I just got the grid size with GetGridSize like grid_width, grid_height = net.GetGridSize(), and executed with linear filter. I want to increase the resolution of mask image. It might be dropped the performance.

2

Hi @Nakhyun, the raw grid size is inherent to the model and can’t be altered after training. It is a ratio of the input resolution of the model. So to increase the grid size, you would need to re-train the model on higher resolution (which would in turn slow down the processing).

For example, if you check the grid sizes of cityscapes-512x256, cityscapes-1024x512, and cityscapes-2048x1024, the grid sizes increase with the resolution of the model.

Alternatively can pass in a larger buffer to net.Mask() and it will use interpolation to up-scale the mask to match the size of your buffer.

3

Thanks dusty_nv. It is clear to me.
Is there any example for retraining the segnet? I just ran the pre-trained model.

4

There isn’t a tutorial about it, but the training code is here: https://github.com/dusty-nv/pytorch-segmentation

I have not tested the segmentation training on Jetson, I train it on PC (it is rather intensive, and the datasets are large)

To train FCN-Resnet18, you would also need to use my fork of torchvision (the v0.3.0 branch): https://github.com/dusty-nv/vision/tree/v0.3.0

5

In train.py of pytorch-segmentation, is the gridsize of network directly related to --resolution or --width, --height?
When I ran the pre-trained model, the grid size is about 64x32(cityscapes-2048x1024), 36x16(1024x512), 16x8(512x256).

parser.add_argument('--resolution', default=320, type=int, metavar='N',
                    help='NxN resolution used for scaling the training dataset (default: 320x320) '
                     'to specify a non-square resolution, use the --width and --height options')
parser.add_argument('--width', default=argparse.SUPPRESS, type=int, metavar='X',
                    help='desired width of the training dataset. if this option is not set, --resolution will be used')
parser.add_argument('--height', default=argparse.SUPPRESS, type=int, metavar='Y',
                    help='desired height of the training dataset. if this option is not set, --resolution will be used')
6

Yes that is correct, the grid size will be a ratio of the input resolution. For FCN-ResNet18, I believe the grid size is 32X smaller than the input size.