Repository navigation
Randomization of transforms per image per batch - #231
Conversation
|
Changes look fine to me for rand_color_jitter. Performance can be improved by doing some sort of kernel aggregation and selective image level filtering within the kernel. Once we have proved that rand_color_jitter is good for model accuracy/froc scores, can look into optimizing performance for this. |
gigony
left a comment
gigony
left a comment
There was a problem hiding this comment.
Looks good to me.
Have some minor suggestions.
| np_output = spt.rand_image_flip(arr_batch, prob=1.0, spatial_axis=(2, 3)) | ||
| np_output = spt.rand_image_flip(arr_batch, | ||
| prob=1.0, | ||
| spatial_axis=(2, 3), |
There was a problem hiding this comment.
Parameter input is not consistent. Here it uses (2, 3), but below code, it uses [2, 3].
| img: Any, | ||
| spatial_axis: tuple(), | ||
| prob: float = 0.1 | ||
| prob: float = 0.1, |
There was a problem hiding this comment.
What would be the good default value for this?
Having 0.5 for the default value looks more useful here?
Co-authored-by: Gigon Bae <gigony@gmail.com>
|
Generally seems reasonable. Do we want to test Edit: Filed issue ( #259 ) to track this. |
|
@gpucibot merge |
gigony
changed the title
|
Can one of the admins verify this patch? |
As per #225
Randomness updated to be applied to per image in the batch. Existing functionality of applying randomness on entire batch at once is unchanged.