Random resizing - downsize only

[mmartin56]

Hi @AlexeyAB

The option random=0 means you can train on larger mini-batches, which, according to you (#4386 ), leads to a higher mAP. random=1 means data augmentation, which I guess also leads to a higher mAP, but for a different reason (model generalisation).

Which one is the best option?

Could only downsizing images be a better compromise? This way we get more data augmentation than random=0, and mini-batches of the same size - it seems better than random=0.

That could be achieved by something like: if ( l.random == 2 && random_val > 1 ) random_val = 1/random_val; after the lines

            printf("Resizing\n");
            float random_val = rand_scale(1.4);    // *x or /x

[AlexeyAB]

I think we can add this feature optionally - ranges of risizing:

• random=0 - without resizing
• random=1 - resizing /1.4 - x1.4 (default values)
• random=2.5 - resizing /2.0 - x2.0 (custom values)
• random=0.5 - resizing /2.0 - x1.0 (downsize only)

[mmartin56]

Thanks! My code if it helps:

            if ( l.random == 1 && random_val > 1 ) random_val = 1/random_val;
            int dim_w = roundl(random_val*init_w / 32 + 1) * 32;
            int dim_h = roundl(random_val*init_h / 32 + 1) * 32;
            if ( l.random == 1 )
            {
                dim_w = fmin( dim_w, init_w );
                dim_h = fmin( dim_h, init_h );
            }
            if (avg_loss < 0) // at the beginning
            {
                if ( l.random == 1 )
                {
                    dim_w = init_w;
                    dim_h = init_h;
                }
                else
                {
                    dim_w = roundl(1.4*init_w / 32 + 1) * 32;
                    dim_h = roundl(1.4*init_h / 32 + 1) * 32;
                }
            }

so random=1 is downsizing only, and any random>1 is the normal resizing (up and down).

[iraadit]

Hi @mmartin56

Did you see any improvement by downsizing only?

Thanks

[mmartin56]

Hi @iraadit, as far as I can tell I haven't seen any drop in performance when only downsizing - but I haven't done proper tests to be honest.

That allowed me to increase the batch size, which also hasn't resulted in a significant increase in performance - all in all probably not much of a difference. :)