BP mode

[lschm]

Hi,
on my data --bp mode gives far superior results, compared to both blind spot size 1 and 3. All other parameters are identical for training and testing. In another post you mentioned, that this should give worse performance. Am I missing something?

I am using GCaMP8m, imaging at 15 Hz. Tried doing a patch_size of 11, 22, 22 to match the slow imaging speed, but that only made it worse.

Code for non-bp denoising:
Train:
patch_size = [61, 22, 22]
patch_interval = [1, 11, 11]
batch_size = 16 # lower it if memory exceeds.
bs_size = 1 # modify if you changed bs_size when training.
bp_mode = False
model = SUPPORT(in_channels=61, mid_channels=[64, 128, 256, 512, 1024], depth=5, blind_conv_channels=64, one_by_one_channels=[32, 16], last_layer_channels=[64, 32, 16], bs_size=bs_size, bp=bp_mode).cuda()

Test:
python -m src.train --exp_name MF1_opt --noisy_data F:\MF1\ --is_folder --results_dir F:\deepSupport\trainedModel --patch_size 61 22 22 --bs_size 1 1 --n_epochs 25 --logging_interval_batch 10000 --unet_channels 64 128 256 512 1024 --patch_interval 11

Thanks,
Lena