rand_conv

This repo shows how to create on-demand random colour augmentations by convolving an image with random conv2d filters. This has shown to improve the performance in RL as well as image classification by https://arxiv.org/abs/1910.05396 and https://openreview.net/pdf?id=BVSM0x3EDK6

The filter weights are initialised from a normal distribution with standard deviation of 1 / sqrt(C_in) * kernel_size. Since we are dealing with 3 channel input images C_in is 3 and we also want output to be 3 channel.

# create the random_conv filter 
m = nn.Conv2d(3, 3, kernel_size, stride=1, padding=kernel_size//2, bias=False).cuda()

std_normal = 1 / (np.sqrt(3) * kernel_size)

m.weight = torch.nn.Parameter(torch.normal(mean=torch.zeros_like(m.weight), 
                                            std=torch.ones_like(m.weight)*std_normal))

However, as mentioned in the paper https://arxiv.org/abs/1910.05396, using only randomised inputs complicates training. Therefore, we do blending as suggested

As shown in the paper https://openreview.net/pdf?id=BVSM0x3EDK6, varying the parameters (kernel size and the blending), we get a wide variety of augmentations

Dependencies

Install the following via pip

fire
matplotlib
pprintpp
torch

How to run

The function call arguments are made via fire. Therefore, to understand what arguments to be passed, we'd recommend to do this first

python rand_conv.py --help

It will show you the following help for commandline arguments

NAME
    rand_conv.py - A function to apply random convolution filters to the image 
                   input (H x W x 3) ->rand_conv-> output (H x W x 3)

SYNOPSIS
    rand_conv.py <flags>

DESCRIPTION
    @param: kernel_size is the size of the convolution filter to be used 
    @param: weight_init is the initalisation scheme (whether it's normal or xavier etc.) 
    @param: alpha is the blending parameter i.e. alpha*input_image + (1-alpha)*convolved_image 
    @param: save_images is for saving the blended images

FLAGS
    --kernel_size=KERNEL_SIZE
    --weight_init=WEIGHT_INIT
    --alpha=ALPHA
    --save_images=SAVE_IMAGES
(END)

Results

Running with --kernel_size=3 and --alpha=0.7

Running with --kernel_size=3 and --alpha=0.3

Running with --kernel_size=11 and --alpha=0.3