elevate project to the next level

[bodokaiser]

1. Precision

To increase image precision on MRI images we would need to:

• [x] read slices from MINC2 as np.float64
• [x] update network models to use weights, etc. as LongTensors
• [x] update scripts/patch.py to save images as TIFFs
• [x] update MNIBITEFolder to load TIFF patches

2. Filtering

To support our enhanced loss function we need to create a binary mask from US thresholding. That mask can also be used to filter too small US images (saves us one calculation) in the training loop.

• [x] create helper function which creates binary mask
• [x] filter provided (mr, us) batch in training loop according to binary mask

3. Enhanced Loss

• [x] use l2 loss
• [x] find better place for calculating mask
• [ ] support batches
• [x] use matrix multiplication with mask
• [x] normalize loss by dividing through pixels

Bugs

• [ ] data loader varnishes after one epoch
• [x] use same dataset for show images and training

[bodokaiser]

I just noticed that our ElasticNet implementation is numerically unstable (result yields nan's).

[bodokaiser]

I found the problem with why data varnishes after one epoch. The reason is that MNIBITENative calls to Normalize() overwrite the images such that with a second normalize (over the original value range) it has lost its precision. I try to find a way that MINC2 loads the corresponding image slice directly over h5py.File each time.

[bodokaiser]

Rebased with master.