Speed up

Want your results faster?

Options to speed up segmentation:

- Use folders with sequential PNGs or TIFs:

Implementation for reading images for segmentation differs on datatype. For large image volumes it is much faster to use sequential .PNGs or sequential .TIFs (instead of a single .tif or .h5 stack).

- Use a multi-GPU instance:

To reduce wall-time speed up the processing you can use a multi-GPU instance. The segmentation will work parallelized and number of GPUs translates into a proportionally decrease of processing time. However read/write speed of the disk may limit a fully proportional speed up. Training of 1fm, 3fm, 5fm models on multiple GPUs is enabled reducing the processing time by a factor of 3. While training on multiple GPUs is also enabled, the data transfer between GPUs is limiting the overall processing speed.

- Setting the speed parameter for segmentation:

This speeds up processing, but potentially with a negative effect on accuracy (speed of 1 equals highest accuracy). This should be tested on a small sample, since it can save a lot of time for large segmentation tasks. Speed of 2 and 4 (twice and ~four times as fast result in a slightly less accurate result) A value of 10 is expected to result in a somewhat reduced accuracy, but the speed-up is >10 times faster. --augspeed: supported values 1,2,4 or 10

- Using only a single model (1fm, 3fm or 5fm) for segmentation:

A single can be used instead of the 3 models. This together with the speed-up value of 10 will result in fastest processing (however may be significantly less precise)

Estimated time gain can be 3* 16 = 48 times (compared to the original publication).