search

JinYize / self2self_pytorch

BSD 3-Clause "New" or "Revised" License
29 stars 2 forks source link

Did it use noisy image? #2

Open SparkElf opened 1 year ago

SparkElf commented 1 year ago

Hello,I would be appereciate if you could explain why the original self2self.py didn't use noisy image to train the model and if I used the noisy image , the result went wrong：the image seemd too white than the ground truth. image image

hzy-del commented 1 year ago

Hello,I would be appereciate if you could explain why the original self2self.py didn't use noisy image to train the model and if I used the noisy image , the result went wrong：the image seemd too white than the ground truth. image image

I also have this problem.

HappyYuji commented 1 year ago

Hello,I would be appereciate if you could explain why the original self2self.py didn't use noisy image to train the model and if I used the noisy image , the result went wrong：the image seemd too white than the ground truth. image image

I also have this problem.

I also have thin problem

QilnnGuo commented 3 weeks ago

Hello,I would be appereciate if you could explain why the original self2self.py didn't use noisy image to train the model and if I used the noisy image , the result went wrong：the image seemd too white than the ground truth. image image

Recently I also tried to reproduce the code, and I think I could answer your problem. You may notice that the author here used the mask for the input with multiplier 0.7 when training, which is unnecessary because the original tensorflow version used tf.nn.dropout(x, keep_drop) which will apply a multiplier 1/keep_drop. And the reason for your brighter output may be caused by that: the code here did not apply the multiplier 0.7 when evaluating despite training input.