-
Notifications
You must be signed in to change notification settings - Fork 0
/
todo
34 lines (24 loc) · 990 Bytes
/
todo
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
1. remove Link in the setupPath file
2. handle the find nothing problem in preprocess
3. handle projection size larger than 28888
check the wavelet, it is very suspicious
Noticed that:
1. continuation.
2. the direct implementation continue decreasing the cost function, while
the fourier transform implementation stack at certain point, although the
two share the same path before the point.
The RSE can reach unblievable precision. 4e-5; for less than 1000 iterations
The line search for FISTA speed up the convergence in terms of number of outer
iterations.
FISTA_NNL1 and FISTA_ADMM_NNL1 are almost equivalent, which means with/without
line search, their convergence curve are the same.
Restarting in FISTA is not better.
todo for denoise_bound_mod
use more agressive step size
use C implementation
use a better mask implementation
make solver to support complex number
write a c++ version of the algorithm
multi threads?
gpu?
!!!make sure that bhc works!