aimansnigdha / ambiguous-medical-image-segmentation-using-diffusion-models Goto Github PK

I've found that the MSE of the model increases after adding the loss of ANM and ACM, so what it does is make the mask blurry

Thank you for your excellent work. Can you check this script?

During training, the while loop in the 'run_loop' function does not seem to terminate because self. lr_ annel_ Steps=0, not self. lr_ annel_ Steps are always true.

the code is below:
segmentation_train.py
def create_argparser():
defaults = dict(
data_dir="./data/training",
schedule_sampler="uniform",
lr=1e-4,
weight_decay=0.0,
lr_anneal_steps=0,
batch_size=1,
microbatch=-1, # -1 disables microbatches
ema_rate="0.9999", # comma-separated list of EMA values
log_interval=100,
save_interval=5000,
resume_checkpoint='',#'"./results/pretrainedmodel.pt",
use_fp16=False,
fp16_scale_growth=1e-3,
)
train_util.py
class TrainLoop:
def init(
self,
*,
model,
classifier,
diffusion,
data,
dataloader,
prior,
posterior,
batch_size,
microbatch,
lr,
ema_rate,
log_interval,
save_interval,
resume_checkpoint,
use_fp16=False,
fp16_scale_growth=1e-3,
schedule_sampler=None,
weight_decay=0.0,
lr_anneal_steps=0,
):

def run_loop(self):
    i = 0
    totseg = 0
    totcls = 0
    totrec=0
    data_iter = iter(self.dataloader)
    while (
        not self.lr_anneal_steps
        or self.step + self.resume_step < self.lr_anneal_steps
    ):

The code is written for binary segmentation. Please direct me on how to modify it for the multi-class segmentation ?

Hi ! According the code, it seems that L_vlb in the loss function is just consist of the KL divergence of the postprior and prior at time step t and not the KL divergence at all of the time steps. Is that True?

I am very grateful for the authors to bring a novel paper to us. But when I tried to reproduce your work, I met some problem. Could you provide the processed data about KIDC and CI score code?

Thanks for your excellent work! Could you please provide your processed data?

Hi

Could you please release the code for the proposed CI score as well?

Thanks in advance!

Best,
Lukas

Hi, thank you for sharing this impressive work! I have a question: how can I train the model using a custom loss function that compares the generated image with the ground truth? Thank you in advance.