Skip to content

scale

Doubler

Bases: ABC

Source code in vodesfunc/scale.py 
40
41
42
43
44
45
46
47
48
49
50
51
52
class Doubler(ABC):
    kwargs: dict[str, Any]
    """Arguments passed to the internal scale function"""

    def __init__(self, **kwargs: Any) -> None:
        self.kwargs = kwargs

    @abstractmethod
    def double(self, clip: vs.VideoNode) -> vs.VideoNode:
        """
        Returns doubled clip
        """
        pass

kwargs: dict[str, Any] = kwargs instance-attribute

Arguments passed to the internal scale function

double(clip) abstractmethod

Returns doubled clip

Source code in vodesfunc/scale.py 
47
48
49
50
51
52
@abstractmethod
def double(self, clip: vs.VideoNode) -> vs.VideoNode:
    """
    Returns doubled clip
    """
    pass

GenericOnnxScaler dataclass

Bases: GenericScaler

Generic scaler class for an onnx model.

Source code in vodesfunc/scale.py 
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
@dataclass
class GenericOnnxScaler(GenericScaler):
    """
    Generic scaler class for an onnx model.
    """

    model: SPathLike | None = None
    """Path to the model."""
    backend: Any | None = None
    """
    vs-mlrt backend. Will attempt to autoselect the most suitable one with fp16=True if None.\n
    In order of trt > cuda > nncn > cpu.
    """
    tiles: int | tuple[int, int] | None = None
    """Splits up the frame into multiple tiles. Helps if you're lacking in vram but models may behave differently."""

    tilesize: int | tuple[int, int] | None = None
    overlap: int | tuple[int, int] | None = None

    _static_kernel_radius = 2

    @inject_self
    def scale(self, clip: vs.VideoNode, width: int, height: int, shift: tuple[float, float] = (0, 0), **kwargs: KwargsT) -> vs.VideoNode:
        if self.backend is None:
            self.backend = autoselect_backend()

        clip_format = get_video_format(clip)
        if clip_format.subsampling_h != 0 or clip_format.subsampling_w != 0:
            raise ValueError("GenericOnnxScaler: This scaler requires non subsampled input!")

        wclip, og_depth = expect_bits(clip, 32)

        from vsmlrt import inference, calc_tilesize, init_backend

        if self.overlap is None:
            overlap_w = overlap_h = 8
        else:
            overlap_w, overlap_h = (self.overlap, self.overlap) if isinstance(self.overlap, int) else self.overlap

        (tile_w, tile_h), (overlap_w, overlap_h) = calc_tilesize(
            tiles=self.tiles,
            tilesize=self.tilesize,
            width=wclip.width,
            height=wclip.height,
            multiple=1,
            overlap_w=overlap_w,
            overlap_h=overlap_h,
        )

        if tile_w % 1 != 0 or tile_h % 1 != 0:
            raise ValueError(f"GenericOnnxScaler: tile size must be divisible by 1 ({tile_w}, {tile_h})")

        backend = init_backend(backend=self.backend, trt_opt_shapes=(tile_w, tile_h))

        scaled = inference(
            wclip, network_path=str(SPath(self.model).resolve()), backend=backend, overlap=(overlap_w, overlap_h), tilesize=(tile_w, tile_h)
        )
        scaled = self._finish_scale(scaled, wclip, width, height, shift)
        return depth(scaled, og_depth)

backend: Any | None = None class-attribute instance-attribute

vs-mlrt backend. Will attempt to autoselect the most suitable one with fp16=True if None.

In order of trt > cuda > nncn > cpu.

model: SPathLike | None = None class-attribute instance-attribute

Path to the model.

tiles: int | tuple[int, int] | None = None class-attribute instance-attribute

Splits up the frame into multiple tiles. Helps if you're lacking in vram but models may behave differently.

Lanczos_PreSS

Bases: Lanczos

Convenience class to pass to a dehalo function. This serves the same purpose as NNEDI to double and reverse using point. Except it is a quite a bit faster and (if using opencl) takes a lot of load off the GPU.

Source code in vodesfunc/scale.py 
26
27
28
29
30
31
32
33
34
35
36
37
class Lanczos_PreSS(Lanczos):
    """
    Convenience class to pass to a dehalo function.
    This serves the same purpose as NNEDI to double and reverse using point.
    Except it is a quite a bit faster and (if using opencl) takes a lot of load off the GPU.
    """

    @inject_self.init_kwargs.clean
    def scale(self, clip: vs.VideoNode, width: int, height: int, shift: tuple[float, float] = (0, 0), **kwargs) -> vs.VideoNode:
        if width != clip.width * 2 or height != clip.height * 2:
            raise ValueError("Lanczos_PreSS: You're probably not using this correctly.")
        return Lanczos.scale(clip, width, height, (-0.25, -0.25))

NNEDI_Doubler

Bases: Doubler

Source code in vodesfunc/scale.py 
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
class NNEDI_Doubler(Doubler):
    ediargs: dict[str, Any]
    opencl: bool

    def __init__(self, opencl: bool = True, nns: int = 4, nsize: int = 4, qual: int = 2, pscrn: int = 1, **kwargs) -> None:
        """
        Simple utility class for doubling a clip using znedi or nnedi3cl (also fixes the shift)

        :param opencl:          Will use nnedi3cl if True and znedi3 if False
        """
        self.ediargs = {"qual": qual, "nsize": nsize, "nns": nns, "pscrn": pscrn}
        self.ediargs.update(**kwargs)
        self.opencl = opencl

    def double(self, clip: vs.VideoNode, correct_shift: bool = True) -> vs.VideoNode:
        y = get_y(clip)
        # nnedi3cl needs padding, to avoid issues on edges (https://slow.pics/c/QcJef38u)
        if self.opencl:
            (left, right, top, bottom) = mod_padding(y, 2, 2)
            width = clip.width + left + right
            height = clip.height + top + bottom
            pad = y.resize.Point(width, height, src_left=-left, src_top=-top, src_width=width, src_height=height).std.Transpose()
            doubled_y = pad.nnedi3cl.NNEDI3CL(dh=True, dw=True, field=0, **self.ediargs).std.Transpose()
            doubled_y = doubled_y.std.Crop(left * 2, right * 2, top * 2, bottom * 2)
        else:
            doubled_y = (
                depth(y, 16)
                .znedi3.nnedi3(dh=True, field=0, **self.ediargs)
                .std.Transpose()
                .znedi3.nnedi3(dh=True, field=0, **self.ediargs)
                .std.Transpose()
            )
            doubled_y = depth(doubled_y, get_depth(clip))

        if correct_shift:
            doubled_y = doubled_y.resize.Bicubic(src_top=0.5, src_left=0.5)

        return doubled_y.std.CopyFrameProps(y)

__init__(opencl=True, nns=4, nsize=4, qual=2, pscrn=1, **kwargs)

Simple utility class for doubling a clip using znedi or nnedi3cl (also fixes the shift)

Parameters:

Name Type Description Default
opencl bool

Will use nnedi3cl if True and znedi3 if False

 True
Source code in vodesfunc/scale.py 
59
60
61
62
63
64
65
66
67
def __init__(self, opencl: bool = True, nns: int = 4, nsize: int = 4, qual: int = 2, pscrn: int = 1, **kwargs) -> None:
    """
    Simple utility class for doubling a clip using znedi or nnedi3cl (also fixes the shift)

    :param opencl:          Will use nnedi3cl if True and znedi3 if False
    """
    self.ediargs = {"qual": qual, "nsize": nsize, "nns": nns, "pscrn": pscrn}
    self.ediargs.update(**kwargs)
    self.opencl = opencl

Waifu2x_Doubler

Bases: Doubler

Source code in vodesfunc/scale.py 
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
class Waifu2x_Doubler(Doubler):
    backend: Any
    kwargs: KwargsT
    w2xargs: KwargsT = {}

    def __init__(
        self,
        cuda: bool | str | None = None,
        fp16: bool = True,
        num_streams: int = 1,
        tiles: int | tuple[int, int] | None = None,
        model: int = 6,
        **kwargs,
    ) -> None:
        """
        Simple utility class for doubling a clip using Waifu2x

        :param cuda:            ORT-Cuda if True, NCNN-VK or CPU (depending on what you have installed) if False, TRT if some string
                                Automatically chosen and tuned when None
        :param fp16:            Uses 16 bit floating point internally if True.
        :param num_streams:     Amount of streams to use for Waifu2x; Sacrifices a lot of vram for a speedup.
        :param tiles:           Splits up the upscaling process into multiple tiles.
                                You will likely have to use atleast `2` if you have less than 16 GB of VRAM.
        :param model:           Model to use from vsmlrt.
        :param kwargs:          Args that get passed to both the Backend and actual scaling function.
        """
        from vsmlrt import Backend, backendT

        self.kwargs = {"num_streams": num_streams, "fp16": fp16}

        if "backend" in kwargs.keys():
            cuda = False

        # Partially stolen from setsu but removed useless stuff that is default in mlrt already and added version checks
        if cuda is None:
            nv = get_nvidia_version()
            cuda = nv is not None
            try:
                if nv is not None and not hasattr(core, "trt") and hasattr(core, "ort"):
                    self.kwargs.update({"use_cuda_graph": True})
                else:
                    props: KwargsT = core.trt.DeviceProperties(kwargs.get("device_id", 0))  # type: ignore
                    version_props: KwargsT = core.trt.Version()  # type: ignore

                    vram = props.get("total_global_memory", 0)
                    trt_version = float(version_props.get("tensorrt_version", 0))

                    cuda = "trt"

                    presumedArgs = KwargsT(
                        workspace=vram / (1 << 22) if vram else None,
                        use_cuda_graph=True,
                        use_cublas=True,
                        use_cudnn=trt_version < 8400,
                        heuristic=trt_version >= 8500,
                        output_format=int(fp16),
                    )

                    # Swinunet doesn't like forced 16. Further testing for the other models needed.
                    if model <= 6:
                        presumedArgs.update({"tf32": not fp16, "force_fp16": fp16})

                    self.kwargs.update(presumedArgs)
            except:
                cuda = nv is not None

        self.w2xargs = KwargsT(
            model=model,
            tiles=tiles,
            preprocess=kwargs.pop("preprocess", True),
            scale=kwargs.pop("scale", 2),
            tilesize=kwargs.pop("tilesize", None),
            overlap=kwargs.pop("overlap", None),
        )

        self.kwargs.update(kwargs)

        if cuda is False:
            backend = kwargs.pop("backend", None)
            if backend and isinstance(backend, backendT):
                self.backend = backend
            else:
                if hasattr(core, "ncnn"):
                    self.backend = Backend.NCNN_VK(**self.kwargs)
                else:
                    self.kwargs.pop("device_id")
                    self.backend = Backend.ORT_CPU(**self.kwargs) if hasattr(core, "ort") else Backend.OV_CPU(**self.kwargs)
        elif cuda is True:
            self.backend = Backend.ORT_CUDA(**self.kwargs) if hasattr(core, "ort") else Backend.OV_GPU(**self.kwargs)
        else:
            self.backend = Backend.TRT(**self.kwargs)
        self.kwargs = kwargs
        self.model = model

    def double(self, clip: vs.VideoNode) -> vs.VideoNode:
        from vsmlrt import Waifu2x

        pre = depth(clip, 32).std.Limiter()

        (left, right, top, bottom) = mod_padding(pre)
        width = pre.width + left + right
        height = pre.height + top + bottom
        pad = pre.resize.Point(width, height, src_left=-left, src_top=-top, src_width=width, src_height=height)

        # Model 0 wants a gray input
        needs_gray = self.w2xargs.get("model", 6) == 0
        was_444 = pre.format.color_family == vs.YUV and pre.format.subsampling_w == 0 and pre.format.subsampling_h == 0 and not needs_gray  # type: ignore

        if was_444:
            pad = Catrom().resample(pad, format=vs.RGBS, matrix=Matrix.RGB, matrix_in=Matrix.from_video(pre))
        elif needs_gray is True:
            pad = get_y(pad)
        else:
            pad = get_y(pad).std.ShufflePlanes(0, vs.RGB)

        up = Waifu2x(pad, noise=-1, backend=self.backend, **self.w2xargs)

        if was_444:
            up = Catrom().resample(up, format=vs.YUV444PS, matrix=Matrix.from_video(pre), matrix_in=Matrix.RGB)
        elif needs_gray is False:
            up = up.std.ShufflePlanes(0, vs.GRAY)

        up = up.std.Crop(left * 2, right * 2, top * 2, bottom * 2)

        # Only Model 6 has the tint
        if self.w2xargs.get("model", 6) == 6:
            up = up.std.Expr("x 0.5 255 / +")

        return depth(up, get_depth(clip)).std.CopyFrameProps(pre)

__init__(cuda=None, fp16=True, num_streams=1, tiles=None, model=6, **kwargs)

Simple utility class for doubling a clip using Waifu2x

Parameters:

Name Type Description Default
cuda bool | str | None

ORT-Cuda if True, NCNN-VK or CPU (depending on what you have installed) if False, TRT if some string Automatically chosen and tuned when None

 None
fp16 bool

Uses 16 bit floating point internally if True.

 True
num_streams int

Amount of streams to use for Waifu2x; Sacrifices a lot of vram for a speedup.

 1
tiles int | tuple[int, int] | None

Splits up the upscaling process into multiple tiles. You will likely have to use atleast 2 if you have less than 16 GB of VRAM.

 None
model int

Model to use from vsmlrt.

 6
kwargs

Args that get passed to both the Backend and actual scaling function.

 {}
Source code in vodesfunc/scale.py 
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
def __init__(
    self,
    cuda: bool | str | None = None,
    fp16: bool = True,
    num_streams: int = 1,
    tiles: int | tuple[int, int] | None = None,
    model: int = 6,
    **kwargs,
) -> None:
    """
    Simple utility class for doubling a clip using Waifu2x

    :param cuda:            ORT-Cuda if True, NCNN-VK or CPU (depending on what you have installed) if False, TRT if some string
                            Automatically chosen and tuned when None
    :param fp16:            Uses 16 bit floating point internally if True.
    :param num_streams:     Amount of streams to use for Waifu2x; Sacrifices a lot of vram for a speedup.
    :param tiles:           Splits up the upscaling process into multiple tiles.
                            You will likely have to use atleast `2` if you have less than 16 GB of VRAM.
    :param model:           Model to use from vsmlrt.
    :param kwargs:          Args that get passed to both the Backend and actual scaling function.
    """
    from vsmlrt import Backend, backendT

    self.kwargs = {"num_streams": num_streams, "fp16": fp16}

    if "backend" in kwargs.keys():
        cuda = False

    # Partially stolen from setsu but removed useless stuff that is default in mlrt already and added version checks
    if cuda is None:
        nv = get_nvidia_version()
        cuda = nv is not None
        try:
            if nv is not None and not hasattr(core, "trt") and hasattr(core, "ort"):
                self.kwargs.update({"use_cuda_graph": True})
            else:
                props: KwargsT = core.trt.DeviceProperties(kwargs.get("device_id", 0))  # type: ignore
                version_props: KwargsT = core.trt.Version()  # type: ignore

                vram = props.get("total_global_memory", 0)
                trt_version = float(version_props.get("tensorrt_version", 0))

                cuda = "trt"

                presumedArgs = KwargsT(
                    workspace=vram / (1 << 22) if vram else None,
                    use_cuda_graph=True,
                    use_cublas=True,
                    use_cudnn=trt_version < 8400,
                    heuristic=trt_version >= 8500,
                    output_format=int(fp16),
                )

                # Swinunet doesn't like forced 16. Further testing for the other models needed.
                if model <= 6:
                    presumedArgs.update({"tf32": not fp16, "force_fp16": fp16})

                self.kwargs.update(presumedArgs)
        except:
            cuda = nv is not None

    self.w2xargs = KwargsT(
        model=model,
        tiles=tiles,
        preprocess=kwargs.pop("preprocess", True),
        scale=kwargs.pop("scale", 2),
        tilesize=kwargs.pop("tilesize", None),
        overlap=kwargs.pop("overlap", None),
    )

    self.kwargs.update(kwargs)

    if cuda is False:
        backend = kwargs.pop("backend", None)
        if backend and isinstance(backend, backendT):
            self.backend = backend
        else:
            if hasattr(core, "ncnn"):
                self.backend = Backend.NCNN_VK(**self.kwargs)
            else:
                self.kwargs.pop("device_id")
                self.backend = Backend.ORT_CPU(**self.kwargs) if hasattr(core, "ort") else Backend.OV_CPU(**self.kwargs)
    elif cuda is True:
        self.backend = Backend.ORT_CUDA(**self.kwargs) if hasattr(core, "ort") else Backend.OV_GPU(**self.kwargs)
    else:
        self.backend = Backend.TRT(**self.kwargs)
    self.kwargs = kwargs
    self.model = model