Vapoursynth

[lansing]

Originally Posted by TheFluff

Well, I dunno. Aside from DeleteFrames taking a list of ints, not a single int (I think) that kinda looks like it should work (maybe; I'm a bit unsure of what'll happen on the last frame of the clip). I don't think "source = abc.avi" is what's actually in your script though. What does happen when you try to run it?

There's no error, it skips the first return and just do the second return, which give me back the source clip.

[TheFluff]

Are you sure about that? I suspect the problem might be that calling deleteframes there doesn't do what you think it does. It won't change the length of the clip returned by frameeval, I'm pretty darn sure. What I'd expect it to do is to replace even numbered frames with the following frame.

[lansing]

Originally Posted by TheFluff

Are you sure about that? I suspect the problem might be that calling deleteframes there doesn't do what you think it does. It won't change the length of the clip returned by frameeval, I'm pretty darn sure. What I'd expect it to do is to replace even numbered frames with the following frame.

Oh yes it did replace the even number frame with the odd one.

So I can't delete frames with arbitrary functions using FrameEval?

[Myrsloik]

Originally Posted by lansing

Oh yes it did replace the even number frame with the odd one.

So I can't delete frames with arbitrary functions using FrameEval?

If you're clever you can reorder things. Note that frameeval takes a clip argument that's only used to set the output format and length. Adjist the total number of frames there.

Probably what you wanted to do. Halve the framecount and you're done.

def DeleteEven(n, clip):	
	return clip[n // 2]

Can also be expressed as clip[::2] without resorting to frameeval at all.

[TheFluff]

Originally Posted by lansing

Oh yes it did replace the even number frame with the odd one.

So I can't delete frames with arbitrary functions using FrameEval?

You can't change the length of the clip, no. When you use FrameEval, the frame number requested from the clip you return from the evaluator function doesn't change. When the client application requests (say) frame 4, and your evaluator function gets n = 4 as an argument and returns a clip that's the same as the original but with frame 4 deleted, FrameEval requests frame 4 from that clip, but since you deleted the original frame 4, the frame now numbered 4 in that clip is number 5 in the original clip. That's why even numbered frames get replaced with the following even one.

The thing Myrsloik did relies on the fact that requesting a frame beyond the end of a clip just returns the last frame, so you get back a clip with the same length as the original, but every frame after the halfway point is just the same as the last.

[Selur]

Is there something like ffmpegs shuffleframes in Vapoursynth? (A filter which allows so swap and drop frames by a given pattern?)

[Myrsloik]

Originally Posted by Selur

Is there something like ffmpegs shuffleframes in Vapoursynth? (A filter which allows so swap and drop frames by a given pattern?)

See frameeval. Especially the most recent posts. A simple callback and you can apply any remapping.

[sneaker_ger]

Isn't it easier with SelectEvery?

[Myrsloik]

Originally Posted by sneaker_ger

Isn't it easier with SelectEvery?

If your pattern can be expressed with it. Yes. But it gets kinda iffy for 1000+ frame cycles and can't do stuff like drop all prime number frames.

[Selur]

SelectEvery could do the dropping, but not the shuffling, or am I misunderstanding the filter?

[Myrsloik]

Originally Posted by Selur

SelectEvery could do the dropping, but not the shuffling, or am I misunderstanding the filter?

Correct, no shuffle.

[sneaker_ger]

I think Selur means "shuffle" as in "re-order with fixed pattern", not "re-order randomly".

So it should be able to do everything ffmpeg's shuffleframes can do.

[Selur]

Yes, given a pattern like for example "0 2 7 5 4 3 1 6 8" would do some crazy reordering on each set of 9 frames.

[Myrsloik]

Originally Posted by Selur

Yes, given a pattern like for example "0 2 7 5 4 3 1 6 8" would do some crazy reordering on each set of 9 frames.

I think that works. Try it and see. I probably made selectevery take the order into account.

[feisty2]

@Myrsloik
can you write something to convert a vaporsynth video clip to a numpy array and a reverse conversion like-wise? so, like
forward conversion: video clip -> numpy.array((framecount, height, width, channel))
reverse conversion: numpy.array((framecount, height, width, channel)) -> video clip
the conversion should only support GRAYS/RGBS/YUV444PS inputs obviously

the thing is... I wanna pre-process my image samples with vaporsynth before feeding them to keras
I could use vspipe to generate raw files and load them with numpy.memmap for now, but it sucks

[Myrsloik]

All frames are exposed as a memoryview. Just use get_frame(). I don't know the exact python way but I'm very certain you can create numpy arrays from memoryviews. Or maybe it already is one.

For the reverse you can probably use modifyframe (see the insane filter writing in python thread). No idea what the speed will be like, probably not that bad as long as you find the appropriate.

Or do it all in a single modifyframe callback. I have no idea what kind of preprocessing you want to do.

[feisty2]

Originally Posted by Myrsloik

I have no idea what kind of preprocessing you want to do.

say I wanna do a super resolution neural net with keras, first I would need to generate the corresponding low resolution version of the high resolution samples, and for that I need resampling filters.
there's scipy.misc.imresize but it's pretty rudimentary comparing to fmtc, it has very few resampling kernels, doesn't even have spline! and you can't tweak stuff like "taps" or whatever for each kernel, with fmtc I got control over everything, I can even create my own resampling kernel or resize under linear light
so I want vaporsynth (filters) to do stuff like that

concretely, something like

def GenLowResSamples(samples):
    core = vs.get_core()
    samples = ArrayToVideo(samples) #reverse conversion
    samples = core.fmtc.resample(samples, sample.width//2, sample.height//2, kernel="sinc", taps=128, fulls=True, fulld=True)
    samples = core.fmtc.resample(samples, sample.width*2, sample.height*2, kernel="sinc", taps=128, fulls=True, fulld=True)
    samples = VideoToArray(samples) #forward conversion
    return samples

[lansing]

Hi I'm trying to frame serve a video from Adobe Premiere to Vapoursynth using this frame server called advanced frame server, when I try to grab it with avisource, the editor crashed and I got this error "setVideoInfo: The frame rate specified by AVISource must be a reduced fraction. (Instead, it is 23976/1000.)", and there's no option in avisource to change it.

[Selur]

I think that works. Try it and see. I probably made selectevery take the order into account.

Works fine! Thanks to myrsloik and sneaker_ger for the help and pointing to SelectEvery() !

[shekh]

Tried to implement r210 DIB. Have issue with VapourSynth-R38

r = core.avisource.AVISource('rgb.avi')
r = core.resize.Point(r, format=vs.RGB30)
r.set_output()

Source is 784x400 but appears to have 832*4 pitch from vs (I derived it from total size and it works).
Is this error on output or there is some rule for it?