batch load or line load

[s0rrymaker77777]

batch load or line load

[gengshan-y]

When switching to --batch_load, you want to change --nsample (number of sampled pxs per image) as well.
For instance, --batch_size 8 --nsample 128, which ensures the same total number of sampled pixels.

Note you are expected to still see slightly worse shape and camera poses, given the total number of updated cameras per iter is fewer than --line_load. My hunch is using CNN to parameterize cameras will improve results in this case, as in LASR.

[s0rrymaker77777]

Thank you for your prompt reply! I found nsample is 6 when batch_size is 256. So how to calculate the related nsample when batch_size changes? Thanks!

[gengshan-y]

So nsample * batch_size gives you the total number of pixels sampled in a batch, which should roughly stays the same

[s0rrymaker77777]

Thanks for your suggestion! I have tried to change --nsample and the performance is close to --lineload. I will also try to parameterize cameras using CNN in the future. Thanks!

[kts707]

Hi @s0rrymaker77777 @gengshan-y , what --nsample and --batch_size did you use to get the similar performance as --lineload?

Also I tried to switch to batch load with --nsample 128 --batch_size 12 to make sure nsample * batch_size stays the same as in lineload, however the training is extremely slow compared to line load. Did your batch load training also become very slow?

[gengshan-y]

If I remember correctly, --nsample 192 --batch_size 8 is 1.5-2x slower and gives slightly worse performance.

[kts707]

Thanks!