vmulti/compensateMulti/degrainN/writevec/readvec for vsmvtools

[feisty2]

def getvectors (src, supersearch, tr=6, pel=4, dct=5, thsad=400):
    core        = vs.get_core ()
    supersoft   = core.mv.Super (src, pel=pel, chroma=False, hpad=32, vpad=32, pelclip=supersearch, sharp=2, rfilter=4, levels=0)
    supersharp  = core.mv.Super (src, pel=pel, chroma=False, hpad=32, vpad=32, pelclip=supersearch, sharp=2, rfilter=2, levels=0)
    def search (isb, delta):
        vectors = core.mv.Analyse (supersoft, isb=isb, overlap=16, blksize=32, search=3, chroma=False, truemotion=True, delta=delta, trymany=True, searchparam=16, pelsearch=16, dct=0, levels=0, divide=2, badrange=-24)
        vectors = core.mv.Recalculate (supersoft, vectors, overlap=8, blksize=16, thsad=thsad//2, chroma=False, truemotion=True, search=3, searchparam=16, dct=dct, smooth=1, divide=2)
        vectors = core.mv.Recalculate (supersharp, vectors, overlap=4, blksize=8, thsad=thsad//2, chroma=False, truemotion=True, search=3, searchparam=16, dct=dct, smooth=1, divide=2)
        vectors = core.mv.Recalculate (supersharp, vectors, overlap=2, blksize=4, thsad=thsad//2, chroma=False, truemotion=True, search=3, searchparam=16, dct=dct, smooth=1, divide=0)
        return vectors
    bv          = [search (True, i) for i in range (tr, 0, -1)]
    fv          = [search (False, i) for i in range (1, tr+1)]
    vmulti      = bv + fv
    vmulti      = core.std.Interleave (vmulti)
    return vmulti

def compensatemulti (src, comp, superclip, vmulti, tr=6, pel=4, thsad=400, thscd1=248, thscd2=130):
    core        = vs.get_core ()
    super       = core.mv.Super (comp, pel=pel, chroma=False, hpad=32, vpad=32, pelclip=superclip, sharp=2, rfilter=2, levels=0)
    def compensate (delta):
        vectors = core.std.SelectEvery (vmulti, tr*2, delta)
        filter  = core.mv.Compensate (src, super, vectors, thsad=thsad, thscd1=thscd1, thscd2=thscd2)
        return filter
    bcomp       = [compensate (i) for i in range (0, tr)]
    fcomp       = [compensate (i) for i in range (tr, 2*tr)]
    compmulti   = bcomp + [src] + fcomp
    compmulti   = core.std.Interleave (compmulti)
    return compmulti

def degrainn (src, comp, superclip, vmulti, tr=6, pel=4, thsad=400, thscd1=248, thscd2=130, full=True):
    core        = vs.get_core ()
    super       = core.mv.Super (comp, pel=pel, chroma=False, hpad=32, vpad=32, pelclip=superclip, sharp=2, rfilter=2, levels=0)
    def MDG1 (a):
        bv      = core.std.SelectEvery (vmulti, tr*2, tr-1-a)
        fv      = core.std.SelectEvery (vmulti, tr*2, tr+a)
        MDG     = core.mv.Degrain1 (src, super, bv, fv, thsad=thsad, thscd1=thscd1, thscd2=thscd2, plane=0)
        return MDG
    MDGMulti    = [MDG1 (i) for i in range (0, tr)]
    MDGMulti    = core.fmtc.bitdepth (core.std.Interleave (MDGMulti), bits=32, flt=True, fulls=full, fulld=full, dmode=1)
    def MDGMerge (start=None, a=2):
        start   = core.std.Merge (core.std.SelectEvery (MDGMulti, tr, 0), core.std.SelectEvery (MDGMulti, tr, 1), 0.5) if start is None else start
        merge   = core.std.Merge (start, core.std.SelectEvery (MDGMulti, tr, a), 1/(a+1))
        a       = a+1
        clip    = merge if a == tr else MDGMerge (start=merge, a=a)
        return clip
    return MDGMerge ()

def writevec (vec, logout):
    core = vs.get_core ()
    w    = vec.get_frame (0).width
    with open (logout, "w") as f:
         print (w, file=f)
    vec  = core.std.CropAbs (vec, width=w, height=1)
    return vec

def readvec (vec, login):
    core = vs.get_core ()
    with open (login, "r") as f:
         w = int (f.readline ())
    vec  = core.raws.Source (vec, w, 1, src_fmt="Y8")
    return vec

works with a "tr" parameter, like in avisynth mvtools
degrainN takes tr >= 3 only, go mv.Degrain1/2 if tr<3, and degrainN ain't the exact extension of degrain1/2/3, it's a more convolved (averaged) variant, but gets you very similar results compared to the original one.

writevec=MStoreVec in avisynth MVTools
readvec=MRestoreVec in avisynth MVTools

[Myrsloik]

This is where you simplify the script to figure out what's causing it.

[feisty2]

bump after update

[feisty2]

update: rewrote the merging part of degrainn, it won't lose precision for nothing now

[feisty2]

update:
new: writevec/readvec

[MonoS]

I suggest you to add something like that

def writevecs(vecs, files, log):
	w = vecs[0].get_frame(0).width
	
	f = [0 for x in range(len(files))]
	for numFile in range(len(files)):
		f[numFile] = open(files[numFile], "wb")

	for i in range(vecs[0].num_frames):
		if (i+1) % 100 == 0:
			print(i+1, "/", vecs[0].num_frames)
		
		for numVec in range(len(vecs)):
			frame = vecs[numVec].get_frame(i)
			view = frame.get_read_array(0)
			
			for y in range(len(view)):
				f[numVec].write(view[y])
	
	for numFile in range(len(files)):
		f[numFile].close()
	
	open (log, "w").write (repr(w))

I made a little test and this gave me about 2x speedup to the single write version.
I don't know how to optimize python code, but maybe it can be further improved.

The function seems to works properly, tried saving the vector file with the single file version, reading it with rawsource and saving 4 vector files using this function [if you try to compare it saving the vector twice you get different md5 due to some initialized memory]

[Myrsloik]

Actually there's another trick you can use to make it doable with vspipe which will definitely beat any python based solution.

1. Get the width of the vector clip from the first frame
2. Use CropAbs(vector_clip, width=frame0width, height=1)
3. vspipe script.vpy vectors.bin -p

alternative 3. Use clip.output() which should also be faster at this point


The key here is that CropAbs can make a variable dimension clip have fixed dimensions (as long as the input frames are big enough to not leave blank parts in the output).

[feisty2]

done, looks super

[splinter98]

I'd tweak the single readvecs and writevecs to the following, you want to ensure you close the files after writing otherwise someone may have issues with too many files open (by using a context manager, the the file closes automatically when you exit the with block).

def writevec(vec, logout):
    core = vs.get_core()
    w = vec.get_frame(0).width
    vec = core.std.CropAbs(vec, width=w, height=1)
    with open(logout, "w") as f:
        f.write(repr(w))
    return vec

def readvec(vec, login):
    core = vs.get_core()
    with open(login, "r") as f:
        w = int(f.read())
    vec = core.raws.Source(vec, w, 1, src_fmt="Y8")
    return vec

Personally I'd also swap f.write(repr(w)) with print(w, file=f) as then you get a new line in the file and f.read() with f.readline() as if the file provided is incorrect and is large you won't fill your memory with nearly as much data before int returns an error.

[MonoS]

Using Myrsloik suggestion i created this.
It's about 25% faster than my previous writevecs [i didn't made any test]

The resulting clip must be saved as a raw file.

ReadVecs works using the same indexes used writing.
Maybe opening the vecs file outside of the function could improve performance, i didn't made extensive testing [and don't appear to be faster than feisty2 readvec implementation]

def WriteVecs(vecs, prefix):
	core = vs.get_core()
	
	w = vecs[0].get_frame(0).width
	
	for numVec in range(len(vecs)):
		vecs[numVec] = core.std.CropAbs(vecs[numVec], width=w, height=1)
	
	v = core.std.StackVertical(vecs)
	
	log = open(prefix + ".len", "w")
	log.write(repr(w))
	
	return v

def ReadVecs(index, prefix, h):
	core = vs.get_core()
	
	f = open(prefix + ".len", "r")
	w = int(f.read())
	f.close()
	
	vecs = core.raws.Source(prefix + ".vec", w, h, src_fmt="Y8")
	
	v = core.std.CropAbs(vecs, y=index, height=1, width=w)
	
	return v

[feisty2]

@MonoS
why putting all that extra work to make the 1xn clip mxn,

[MonoS]

Originally Posted by feisty2

@MonoS
why putting all that extra work to make the 1xn clip mxn,

Because i wanted an easy way to save all the vector in 1 call and have a nice speedup [this last one was not expected]

If you want to save only one vector clip than your method is the best [simple code, no strange modification to the clip, etc], but if you want to save more vector clips you must do it by hand and imho it's annoying

[feisty2]

update: degrainn always outputs float point clips now, to minimize the rounding error of the merging process

[splinter98]

Originally Posted by MonoS

Using Myrsloik suggestion i created this.
It's about 25% faster than my previous writevecs [i didn't made any test]

...

def WriteVecs(vecs, prefix):
	core = vs.get_core()
	
	w = vecs[0].get_frame(0).width
	
	for numVec in range(len(vecs)):
		vecs[numVec] = core.std.CropAbs(vecs[numVec], width=w, height=1)
	
	v = core.std.StackVertical(vecs)
	
	log = open(prefix + ".len", "w")
	log.write(repr(w))
	
	return v
...

You can squeeze even more performance out of this function by making it a little more pythonic, and also removes the mutation of vecs.

def WriteVecs(vecs, prefix):
    core = vs.get_core()
    w = vecs[0].width
    v = core.std.StackVertical([core.std.CropAbs(vec, width=w, height=1) for vec in vecs])
    with open(prefix + ".len", "w") as f:
        print(w, file=f)
    return v

[MonoS]

Originally Posted by splinter98

You can squeeze even more performance out of this function by making it a little more pythonic, and also removes the mutation of vecs.

Oh, that nice, thanks [i hope that my supreme skillfulness on python come through my code]

[feisty2]

update: fixed a "limit" bug, just found out not like in avisynth mvtools, "no limit" in mv.degrain does not stay constant as "255", actually bit depth related (2 ^ bitdepth - 1), simplified the merging expression, should run a little bit faster now

[Myrsloik]

Originally Posted by feisty2

update: fixed a "limit" bug, just found out not like in avisynth mvtools, "no limit" in mv.degrain does not stay constant as "255", actually bit depth related (2 ^ bitdepth - 1), simplified the merging expression, should run a little bit faster now

Generally speaking Expr is slower than anything else that does similar things. Its only advantage is that it can do things no other filter can sometimes...

start = core.std.Expr ([core.std.SelectEvery (MDGMulti, tr, 0), core.std.SelectEvery (MDGMulti, tr, 1)], ["x y + 2 /"]) if start is None else start
is faster if written as
core.std.Merge(core.std.SelectEvery (MDGMulti, tr, 0), clipb=core.std.SelectEvery (MDGMulti, tr, 1))

Likewise
merge = core.std.Expr ([start, core.std.SelectEvery (MDGMulti, tr, a)], ["x {a} * y + {a} 1 + /".format (a=a)])
can be accomplished faster by fiddling with the weight of the merge.

[feisty2]

update: replaced Expr with Merge, should work faster now