todo.txt

np.lstsq for replacing cgls
use matplotlib
plt.figure(), plt.imshow(abs(im), cmap='gray'), plt.show()

plot distribution of hashkey
9 filters, one for each location of fourier

Train on several images, test, find minimal
numba, make code faster
strength with normalization
calculate distance between test and result
structure tensor for computing angles

larger sets
train on differrent 3d object
pixel type with other hash
spread out set to train
Try without normalization
Try dividing by mean for each pach filter

Try just cropping fourier
crop and zero padding for regular images/knee images crop a square shape (2/3) and a circle masking (again 2/3)
test for knee images
upscalling by 1.5 (knees)
Bicubic upscalling

imagenet train on regular images, test on mri
try adjusting crop factor to 1/2
try normalizing over patch filter (sum all values in filter, divide by 1)
try absolute value before filter calculations

visualize different filters
imagenet testing
bugfix weird pixels in zero version
sanity check for training, total mean sq error should go down
	run test on each image in train and find total mse
try adjusting patchsize
play around with some variables to get improvements
work on unit tests
email mikki aboout summer plans cc frank

consider removing pixeltype for zeropadding

Try considering scaling factors for fft
consider using bicubic
Possibly add bias term for each filter
Get all angles

train on single line image
consider adding filter for each pixel in patch, sum all outputs
Try more angles/coherence/strength
try changing patch size

Concatonate images to print
Make presentation for Wednesday
-Why, problem, goal
-Raisr
-differences between this and other methods (RAISR, zeropadding MRI)
-results
Test on MRI images


Try visualizing filter better with log
Only train on center slices (125-225)
try several different sigmas
retest with new gradient size

1 page abstract
-why is it better
-tests needed
train on regular images
use non trained images for testing
use sigma = 1
adding noise
send email to keys@eecs.berkeley.edu cc Miki and Frank
look at abstracts ISMRM.org
-Less than 750 words
-5 figs max
https://cds.ismrm.org/protected/18MPresentations/abstracts/3365.html

training on real images filters for bilinear vs zeropadding
training on natural images vs MRI
filters from regular images applied to MR vs MR onto MR