search

zhixuhao / unet

unet for image segmentation
MIT License
4.59k stars 2k forks source link

Reducing size of training set will make output pure gray picture. #74

Open CreepZzy opened 6 years ago

CreepZzy commented 6 years ago

When I used my dataset to train the u-net, I found a weird phenomenon. When I used 16 images to train, the predict result will show a image like mask, a black and white image. However, when I used 14 images to train the model, the predict results became pure gray images. Did anyone face the same problem or have some solutions to this problem?

helwilliams commented 6 years ago

yes I have the same problem, I have 26 input ultrasound images and I get a low contrast image warning, if anyones solved this problem any information would be most helpful. Thanks

CreepZzy commented 5 years ago

yes I have the same problem, I have 26 input ultrasound images and I get a low contrast image warning, if anyones solved this problem any information would be most helpful. Thanks

I found that the problem may resulted from randomness. Reduce the epochs and try to do the training multiple times. Sometimes it gave pure gray results, somtimes it can produce meaningful results when the loss function decrease significantly in the first epoch... So maybe you can run the code more times.

jizhang02 commented 5 years ago

hello, Why the input data size is different from the output size?