Open AXu0511 opened 1 year ago
We tried and updated both the full- and half-resolution code. The pre-trained half-resolution models are also released and can be found in the readme.
Hi there,
Thanks for the work. I tried re-training the neurite OASIS dataset, and the full-resolution training collapses after a few iterations. Here is what the training log looks like:
I have used the default parameters provided in the file. Did you face any similar problems while training? Thanks in advance for the help.
@rohitrango, the corrected weights can be found in the command.txt, with --data_labda 1.0 (not used in code actually) --smth_labda 0.01 (is different from the one in train.py) --mask__labda 1.0 (as default in the train.py)
loss = 1 loss1 + 0.01 loss2 + 1.0 * loss3
Yes, I noticed that data_labda
was not used, and I added it back (using 1.0) as of now.
I've changed smth_labda
and start_channels
appropriately from command.txt
. I also wanted to use cross correlation instead of L2 loss (using_l2
= 2), is that fine or does it make the training unstable?
I've started training with the config you mentioned, fingers crossed!
@rohitrango
use cross correlation instead of L2 loss (using_l2 = 2), is that fine?
Yes. For NCC to achieve a satisfying Dice, the smth_labda should be much larger than 0.01 used for MSE. I vaguely remember we tuned a few sets of parameters like --data_labda 1.0 --smth_labda 0.1 --masklabda 1.0 or --data_labda 1.0 --smth_labda 0.5 --masklabda 1.0, but the optimal Dice of NCC is 1%-2% lower than that of MSE for this specific dataset.
Sounds good. Thanks for the info.
Hello, thank you very much for your outstanding work, I was recently reproducing your results. Did you train the half-resolution displacements directly or train the a full-resolution displacement first then downsamples the displacement? By the way, could you be so kind to share the OASIS version code?