ClikaRT::processor::ImageProcessorConfig

class

Header: ClikaRT/processor/config.h

Settable configuration for an ImageProcessor (convert-rgb → resize → center-crop → rescale → normalize → pad). Default-constructed to the documented HuggingFace defaults (resize + rescale by 1/255 + normalize on; crop + pad off); tune with the fluent setters, or round-trip a preprocessor_config.json through from_json / to_json. Move-only.

Static member functions

from_json()

static ImageProcessorConfig from_json(std::string_view text)

Parse a config from preprocessor_config.json text (flat, or nesting an "image_processor" block). Malformed text raises ClikaRT::Error.

The accepted schema is ClikaRT's own canonical config: the HuggingFace preprocessing keys plus the ClikaRT extension keys (target_dtype, resize_before_cast). Two semantics to be aware of when feeding it a HuggingFace-authored file:

• A missing key takes ClikaRT's documented default (this class's default-constructed state), NOT the per-processor-class default the HuggingFace image_processor_type would imply. Fields a HuggingFace class hardcodes rather than serializes must be set explicitly.
• A key outside the accepted schema is ignored for the parse but logged as a warning naming it, so a mis-spelled or unsupported key is a visible diagnostic, never a silently different pipeline.
• A size / crop_size / pad_size value is one of: a bare integer (for size, the shortest edge, aspect preserved, the CLIP-style "size": 224; for crop_size / pad_size, a square), a [height, width] list (an exact target), {height, width}, {shortest_edge}, {shortest_edge, longest_edge}, or {max_height, max_width}. Any other value form (a one-element list, a string, an object with another key set) is refused with an error naming the supported forms, never read as "no size".

Declared in ClikaRT/processor/config.h, line 80

Member functions

ImageProcessorConfig()

ImageProcessorConfig()

A default configuration, usable with no config file.

Declared in ClikaRT/processor/config.h, line 50

ImageProcessorConfig(ImageProcessorConfig)

ImageProcessorConfig(ImageProcessorConfig&&) noexcept

move-construct (the source becomes an empty shell)

Declared in ClikaRT/processor/config.h, line 51

operator=(ImageProcessorConfig)

ImageProcessorConfig& operator=(ImageProcessorConfig&&) noexcept

move-assign (the source becomes an empty shell)

Declared in ClikaRT/processor/config.h, line 52

ImageProcessorConfig(ImageProcessorConfig)

ImageProcessorConfig(const ImageProcessorConfig&)

copy-construct: a DEEP copy (independent settings)

Declared in ClikaRT/processor/config.h, line 53

operator=(ImageProcessorConfig)

ImageProcessorConfig& operator=(const ImageProcessorConfig&)

copy-assign: a deep copy

Declared in ClikaRT/processor/config.h, line 54

~ImageProcessorConfig()

~ImageProcessorConfig()

destructor

Declared in ClikaRT/processor/config.h, line 55

to_json()

std::string to_json(int indent = 2) const

Serialize to preprocessor_config.json text (indent < 0 = compact), ClikaRT's canonical form, with every field (defaults included) materialized. Every field round-trips through from_json. Infallible.

Declared in ClikaRT/processor/config.h, line 86

resize_shortest_edge(int64_t)

ImageProcessorConfig& resize_shortest_edge(std::int64_t edge)

Resize the shortest edge to edge, preserving aspect (turns resize on): the {shortest_edge} size form; a bare integer size in a config file means the same. The other edge is truncated, never rounded, so a 480×640 image at 224 becomes 224×298.

Declared in ClikaRT/processor/config.h, line 94

resize_shortest_edge(int64_t, int64_t)

ImageProcessorConfig& resize_shortest_edge(std::int64_t edge, std::int64_t longest_cap)

Resize the shortest edge to edge and cap the longest at longest_cap (the DETR-style aspect-preserving fit, the {shortest_edge, longest_edge} size form; turns resize on). When the cap binds, the shortest edge is re-derived from it: a 1080×1920 image at 800/1333 becomes 750×1333.

Declared in ClikaRT/processor/config.h, line 99

resize_to()

ImageProcessorConfig& resize_to(std::int64_t height, std::int64_t width)

Resize to an exact height×width (turns resize on): the {height, width} / [height, width] size forms.

Declared in ClikaRT/processor/config.h, line 102

resize_within()

ImageProcessorConfig& resize_within(std::int64_t max_height, std::int64_t max_width)

Fit inside max_height×max_width preserving aspect, the {max_height, max_width} size form (turns resize on). Scales up as well as down: the more constrained side lands exactly on its bound, the other is truncated; a 300×400 image within 100×60 becomes 45×60.

Declared in ClikaRT/processor/config.h, line 107

set_do_resize()

ImageProcessorConfig& set_do_resize(bool on)

resize step on/off (the resize_* setters turn it on)

Declared in ClikaRT/processor/config.h, line 108

resample(ImageResample)

ImageProcessorConfig& resample(ImageResample mode)

The resize resampling filter (PIL vocabulary). Default bicubic.

Declared in ClikaRT/processor/config.h, line 110

resample(ImageResample, bool)

ImageProcessorConfig& resample(ImageResample mode, bool antialias)

Filter + antialias in one call (the resize quality pair): mode picks the filter, antialias the smoothing.

Declared in ClikaRT/processor/config.h, line 113

set_antialias()

ImageProcessorConfig& set_antialias(bool on)

Antialiased resize on/off. Default ON, the PIL behavior every HuggingFace image processor exhibits; turn off only to trade fidelity for speed on large downscales.

Declared in ClikaRT/processor/config.h, line 117

center_crop_to()

ImageProcessorConfig& center_crop_to(std::int64_t height, std::int64_t width)

Center-crop to height×width (turns crop on).

Declared in ClikaRT/processor/config.h, line 119

set_do_center_crop()

ImageProcessorConfig& set_do_center_crop(bool on)

center-crop step on/off (center_crop_to turns it on)

Declared in ClikaRT/processor/config.h, line 120

pad_to()

ImageProcessorConfig& pad_to(std::int64_t height, std::int64_t width)

Pad (detection-style) bottom/right with zeros to height×width (turns pad on).

Declared in ClikaRT/processor/config.h, line 123

set_do_pad()

ImageProcessorConfig& set_do_pad(bool on)

pad step on/off (pad_to turns it on)

Declared in ClikaRT/processor/config.h, line 124

rescale_by()

ImageProcessorConfig& rescale_by(double factor)

Scale pixel values by factor (e.g. 1/255). Turns rescale on.

Declared in ClikaRT/processor/config.h, line 126

set_do_rescale()

ImageProcessorConfig& set_do_rescale(bool on)

rescale step on/off (rescale_by turns it on)

Declared in ClikaRT/processor/config.h, line 127

normalize()

ImageProcessorConfig& normalize(ClikaRT::Span<const float> mean, ClikaRT::Span<const float> std)

Per-channel normalization (subtract mean, divide by std). mean/std are length 1 or 3; turns normalize on. The spans are read during the call.

Declared in ClikaRT/processor/config.h, line 130

set_do_normalize()

ImageProcessorConfig& set_do_normalize(bool on)

normalize step on/off (normalize turns it on)

Declared in ClikaRT/processor/config.h, line 131

set_do_convert_rgb()

ImageProcessorConfig& set_do_convert_rgb(bool on)

force 1/4-channel input to RGB before the pipeline (default on)

Declared in ClikaRT/processor/config.h, line 132

set_target_dtype()

ImageProcessorConfig& set_target_dtype(DataType dtype)

The dtype the pipeline produces (and works in for rescale/normalize). The loaded image is cast to it exactly once, or not at all when it already matches. Default Float32.

Declared in ClikaRT/processor/config.h, line 136

set_resize_before_cast()

ImageProcessorConfig& set_resize_before_cast(bool on)

Resize/crop in the LOADED dtype and cast to target_dtype afterward, the HuggingFace/PIL order (default true). Set false to cast to target_dtype FIRST and resize in that dtype: the portable path for a device whose resize kernel does not cover the loaded integer dtype.

Declared in ClikaRT/processor/config.h, line 141

do_resize()

bool do_resize() const

whether the resize step runs (default on)

Declared in ClikaRT/processor/config.h, line 144

resample()

ImageResample resample() const

The resampling filter. A config whose PIL code has no filter here (e.g. LANCZOS) reads back Bicubic, the filter the pipeline actually runs.

Declared in ClikaRT/processor/config.h, line 147

antialias()

bool antialias() const

whether resize antialiases (default on)

Declared in ClikaRT/processor/config.h, line 148

do_center_crop()

bool do_center_crop() const

whether center-crop runs (default off)

Declared in ClikaRT/processor/config.h, line 149

do_rescale()

bool do_rescale() const

whether rescale runs (default on)

Declared in ClikaRT/processor/config.h, line 150

rescale_factor()

double rescale_factor() const

the rescale multiplier (default 1/255)

Declared in ClikaRT/processor/config.h, line 151

do_normalize()

bool do_normalize() const

whether normalize runs (default on)

Declared in ClikaRT/processor/config.h, line 152

do_convert_rgb()

bool do_convert_rgb() const

whether RGB conversion runs (default on)

Declared in ClikaRT/processor/config.h, line 153

do_pad()

bool do_pad() const

whether detection-pad runs (default off)

Declared in ClikaRT/processor/config.h, line 154

target_dtype()

DataType target_dtype() const

the pipeline output/work dtype (default Float32)

Declared in ClikaRT/processor/config.h, line 155

resize_before_cast()

bool resize_before_cast() const

whether geometry runs before the dtype cast (default true)

Declared in ClikaRT/processor/config.h, line 156