Xiaomeng Li
Xiaomeng Li
Hi, I changed all `nn.BatchNorm2d` to `nn.BatchNorm3d` in the resnet.py file and it works now. Run cifar.py at epoch 120, it achieves 94.00 on the test dataset. Maybe it is...
@YU1ut Thanks for your explanation. I noticed that you copy the BN parameters from self.tmp_model instead of self.model. Could you please explain why is this? ``` for ema_param, tmp_param in...
@Kennard123661 I think you are correct. This is a bug in the code. Another way is: > jittered_clip = [] for img in clip: jittered_img = img.copy() for func in...
The problem is in the data loader part, I guess. Check if the slicing operation excels the array boundary.
Since tumor label is 2. np.where(livertumor == 2) used to find the location of tumors.
After liver segmentation, we just get the liver mask. The data is also in 512*512. We random crop it into 224, and then feed into the network for training.
you can download from challenge website. https://competitions.codalab.org/competitions/17094
It is not necessary to train Resnet, since the coarse segmentation is only used in the inference stage. Training takes around 3 days using 4 12GB TITANXP GPU.
The functions are used to find the liver, tumor locations, so that we can generate training samples containing much more tumors.
The liver mask you load is the one in the "test data", however, I think you load the "train data" for testing?