Add some benchmarks and information

This worsens CPU-only times by some five percent,
but can also make GPU-accelerated runtime twice as fast.

LICENSE

@@ -1,4 +1,4 @@
-Copyright (c) 2023, Přemysl Eric Janouch <p@janouch.name>
+Copyright (c) 2023 - 2024, Přemysl Eric Janouch <p@janouch.name>
 
 Permission to use, copy, modify, and/or distribute this software for any
 purpose with or without fee is hereby granted.

deeptagger/CMakeLists.txt

@@ -0,0 +1,20 @@
+# Ubuntu 20.04 LTS
+cmake_minimum_required (VERSION 3.16)
+project (deeptagger VERSION 0.0.1 LANGUAGES CXX)
+
+# Hint: set ONNXRuntime_ROOT to a directory with a pre-built GitHub release.
+# (Useful for development, otherwise you may need to adjust the rpath.)
+set (CMAKE_MODULE_PATH "${PROJECT_SOURCE_DIR}")
+
+find_package (ONNXRuntime REQUIRED)
+find_package (PkgConfig REQUIRED)
+pkg_check_modules (GM REQUIRED GraphicsMagick++)
+
+add_executable (deeptagger deeptagger.cpp)
+target_compile_features (deeptagger PRIVATE cxx_std_17)
+target_include_directories (deeptagger PRIVATE
+	${GM_INCLUDE_DIRS} ${ONNXRuntime_INCLUDE_DIRS})
+target_link_directories (deeptagger PRIVATE
+	${GM_LIBRARY_DIRS})
+target_link_libraries (deeptagger PRIVATE
+	${GM_LIBRARIES} ${ONNXRuntime_LIBRARIES})

deeptagger/FindONNXRuntime.cmake

@@ -0,0 +1,11 @@
+# Public Domain
+
+find_path (ONNXRuntime_INCLUDE_DIRS onnxruntime_c_api.h
+	PATH_SUFFIXES onnxruntime)
+find_library (ONNXRuntime_LIBRARIES NAMES onnxruntime)
+
+include (FindPackageHandleStandardArgs)
+FIND_PACKAGE_HANDLE_STANDARD_ARGS (ONNXRuntime DEFAULT_MSG
+	ONNXRuntime_INCLUDE_DIRS ONNXRuntime_LIBRARIES)
+
+mark_as_advanced (ONNXRuntime_LIBRARIES ONNXRuntime_INCLUDE_DIRS)

deeptagger/README.adoc

@@ -0,0 +1,130 @@
+deeptagger
+==========
+
+This is an automatic image tagger/classifier written in C++,
+primarily targeting various anime models.
+
+Unfortunately, you will still need Python 3, as well as some luck, to prepare
+the models, achieved by running download.sh.  You will need about 20 gigabytes
+of space for this operation.
+
+"WaifuDiffusion v1.4" models are officially distributed with ONNX model exports
+that do not support symbolic batch sizes.  The script attempts to fix this
+by running custom exports.
+
+You're invited to change things to suit your particular needs.
+
+Getting it to work
+------------------
+To build the evaluator, install a C++ compiler, CMake, and development packages
+of GraphicsMagick and ONNX Runtime.
+
+Prebuilt ONNX Runtime can be most conveniently downloaded from
+https://github.com/microsoft/onnxruntime/releases[GitHub releases].
+Remember to also install CUDA packages, such as _nvidia-cudnn_ on Debian,
+if you plan on using the GPU-enabled options.
+
+ $ cmake -DONNXRuntime_ROOT=/path/to/onnxruntime -B build
+ $ cmake --build build
+ $ ./download.sh
+ $ build/deeptagger models/deepdanbooru-v3-20211112-sgd-e28.model image.jpg
+
+Very little effort is made to make the project compatible with non-POSIX
+systems.
+
+Options
+-------
+--batch 1::
+	This program makes use of batches by decoding and preparing multiple images
+	in parallel before sending them off to models.
+	Batching requires appropriate models.
+--cpu::
+	Force CPU inference, which is usually extremely slow.
+--debug::
+	Increase verbosity.
+--options "CUDAExecutionProvider;device_id=0"::
+	Set various ONNX Runtime execution provider options.
+--pipe::
+	Take input filenames from the standard input.
+--threshold 0.1::
+	Output weight threshold.  Needs to be set very high on ML-Danbooru models.
+
+Model benchmarks
+----------------
+These were measured on a machine with GeForce RTX 4090 (24G),
+and Ryzen 9 7950X3D (32 threads), on a sample of 704 images,
+which took over eight hours.
+
+There is room for further performance tuning.
+
+GPU inference
+~~~~~~~~~~~~~
+[cols="<,>,>", options=header]
+|===
+|Model|Batch size|Time
+|ML-Danbooru Caformer dec-5-97527|16|OOM
+|WD v1.4 ViT v2 (batch)|16|19 s
+|DeepDanbooru|16|21 s
+|WD v1.4 SwinV2 v2 (batch)|16|21 s
+|WD v1.4 ViT v2 (batch)|4|27 s
+|WD v1.4 SwinV2 v2 (batch)|4|30 s
+|DeepDanbooru|4|31 s
+|ML-Danbooru TResNet-D 6-30000|16|31 s
+|WD v1.4 MOAT v2 (batch)|16|31 s
+|WD v1.4 ConvNeXT v2 (batch)|16|32 s
+|WD v1.4 ConvNeXTV2 v2 (batch)|16|36 s
+|ML-Danbooru TResNet-D 6-30000|4|39 s
+|WD v1.4 ConvNeXT v2 (batch)|4|39 s
+|WD v1.4 MOAT v2 (batch)|4|39 s
+|WD v1.4 ConvNeXTV2 v2 (batch)|4|43 s
+|WD v1.4 ViT v2|1|43 s
+|WD v1.4 ViT v2 (batch)|1|43 s
+|ML-Danbooru Caformer dec-5-97527|4|48 s
+|DeepDanbooru|1|53 s
+|WD v1.4 MOAT v2|1|53 s
+|WD v1.4 ConvNeXT v2|1|54 s
+|WD v1.4 MOAT v2 (batch)|1|54 s
+|WD v1.4 SwinV2 v2|1|54 s
+|WD v1.4 SwinV2 v2 (batch)|1|54 s
+|WD v1.4 ConvNeXT v2 (batch)|1|56 s
+|WD v1.4 ConvNeXTV2 v2|1|56 s
+|ML-Danbooru TResNet-D 6-30000|1|58 s
+|WD v1.4 ConvNeXTV2 v2 (batch)|1|58 s
+|ML-Danbooru Caformer dec-5-97527|1|73 s
+|===
+
+CPU inference
+~~~~~~~~~~~~~
+[cols="<,>,>", options=header]
+|===
+|Model|Batch size|Time
+|DeepDanbooru|16|45 s
+|DeepDanbooru|4|54 s
+|DeepDanbooru|1|88 s
+|ML-Danbooru TResNet-D 6-30000|4|139 s
+|ML-Danbooru TResNet-D 6-30000|16|162 s
+|ML-Danbooru TResNet-D 6-30000|1|167 s
+|WD v1.4 ConvNeXT v2|1|208 s
+|WD v1.4 ConvNeXT v2 (batch)|4|226 s
+|WD v1.4 ConvNeXT v2 (batch)|16|238 s
+|WD v1.4 ConvNeXTV2 v2|1|245 s
+|WD v1.4 ConvNeXTV2 v2 (batch)|4|268 s
+|WD v1.4 ViT v2 (batch)|16|270 s
+|WD v1.4 ConvNeXT v2 (batch)|1|272 s
+|WD v1.4 SwinV2 v2 (batch)|4|277 s
+|WD v1.4 ViT v2 (batch)|4|277 s
+|WD v1.4 ConvNeXTV2 v2 (batch)|16|294 s
+|WD v1.4 SwinV2 v2 (batch)|1|300 s
+|WD v1.4 SwinV2 v2|1|302 s
+|WD v1.4 SwinV2 v2 (batch)|16|305 s
+|WD v1.4 MOAT v2 (batch)|4|307 s
+|WD v1.4 ViT v2|1|308 s
+|WD v1.4 ViT v2 (batch)|1|311 s
+|WD v1.4 ConvNeXTV2 v2 (batch)|1|312 s
+|WD v1.4 MOAT v2|1|332 s
+|WD v1.4 MOAT v2 (batch)|16|335 s
+|WD v1.4 MOAT v2 (batch)|1|339 s
+|ML-Danbooru Caformer dec-5-97527|4|637 s
+|ML-Danbooru Caformer dec-5-97527|16|689 s
+|ML-Danbooru Caformer dec-5-97527|1|829 s
+|===

deeptagger/bench-interpret.sh

@@ -0,0 +1,51 @@
+#!/bin/sh -e
+parse() {
+	awk 'BEGIN {
+		OFS = FS = "\t"
+	} {
+		name = $1
+		path = $2
+		cpu = $3 != ""
+		batch = $4
+		time = $5
+
+		if (path ~ "/batch-")
+			name = name " (batch)"
+		else if (name ~ /^WD / && batch > 1)
+			next
+	} {
+		group = name FS cpu FS batch
+		if (lastgroup != group) {
+			if (lastgroup)
+				print lastgroup, mintime
+
+			lastgroup = group
+			mintime = time
+		} else {
+			if (mintime > time)
+				mintime = time
+		}
+	} END {
+		print lastgroup, mintime
+	}' "${BENCH_LOG:-bench.out}"
+}
+
+cat <<END
+GPU inference
+~~~~~~~~~~~~~
+[cols="<,>,>", options=header]
+|===
+|Model|Batch size|Time
+$(parse | awk -F'\t' 'BEGIN { OFS = "|" }
+	!$2 { print "", $1, $3, $4 " s" }' | sort -t'|' -nk4)
+|===
+
+CPU inference
+~~~~~~~~~~~~~
+[cols="<,>,>", options=header]
+|===
+|Model|Batch size|Time
+$(parse | awk -F'\t' 'BEGIN { OFS = "|" }
+	$2 { print "", $1, $3, $4 " s" }' | sort -t'|' -nk4)
+|===
+END

deeptagger/bench.sh

@@ -0,0 +1,38 @@
+#!/bin/sh -e
+if [ $# -lt 2 ] || ! [ -x "$1" ]
+then
+	echo "Usage: $0 DEEPTAGGER FILE..."
+	echo "Run this after using download.sh, from the same directory."
+	exit 1
+fi
+
+runner=$1
+shift
+log=bench.out
+: >$log
+
+run() {
+	opts=$1 batch=$2 model=$3
+	shift 3
+
+	for i in $(seq 1 3)
+	do
+		start=$(date +%s)
+		"$runner" $opts -b "$batch" -t 0.75 "$model" "$@" >/dev/null || :
+		end=$(date +%s)
+		printf '%s\t%s\t%s\t%s\t%s\n' \
+			"$name" "$model" "$opts" "$batch" "$((end - start))" | tee -a $log
+	done
+}
+
+for model in models/*.model
+do
+	name=$(sed -n 's/^name=//p' "$model")
+	run ""     1 "$model" "$@"
+	run ""     4 "$model" "$@"
+	run ""    16 "$model" "$@"
+
+	run --cpu  1 "$model" "$@"
+	run --cpu  4 "$model" "$@"
+	run --cpu 16 "$model" "$@"
+done

deeptagger/deeptagger.cpp

@@ -0,0 +1,744 @@
+#include <getopt.h>
+#include <Magick++.h>
+#include <onnxruntime_cxx_api.h>
+#ifdef __APPLE__
+#include <coreml_provider_factory.h>
+#endif
+
+#include <algorithm>
+#include <condition_variable>
+#include <filesystem>
+#include <fstream>
+#include <iostream>
+#include <mutex>
+#include <queue>
+#include <regex>
+#include <set>
+#include <stdexcept>
+#include <string>
+#include <thread>
+#include <tuple>
+
+#include <cstdio>
+#include <cstdint>
+#include <climits>
+
+static struct {
+	bool cpu = false;
+	int debug = 0;
+	long batch = 1;
+	float threshold = 0.1;
+
+	// Execution provider name → Key → Value
+	std::map<std::string, std::map<std::string, std::string>> options;
+} g;
+
+// --- Configuration -----------------------------------------------------------
+
+// Arguably, input normalization could be incorporated into models instead.
+struct Config {
+	std::string name;
+	enum class Shape {NHWC, NCHW} shape = Shape::NHWC;
+	enum class Channels {RGB, BGR} channels = Channels::RGB;
+	bool normalize = false;
+	enum class Pad {WHITE, EDGE, STRETCH} pad = Pad::WHITE;
+	int size = -1;
+	bool sigmoid = false;
+
+	std::vector<std::string> tags;
+};
+
+static void
+read_tags(const std::string &path, std::vector<std::string> &tags)
+{
+	std::ifstream f(path);
+	f.exceptions(std::ifstream::badbit);
+	if (!f)
+		throw std::runtime_error("cannot read tags");
+
+	std::string line;
+	while (std::getline(f, line)) {
+		if (!line.empty() && line.back() == '\r')
+			line.erase(line.size() - 1);
+		tags.push_back(line);
+	}
+}
+
+static void
+read_field(Config &config, std::string key, std::string value)
+{
+	if (key == "name") {
+		config.name = value;
+	} else if (key == "shape") {
+		if      (value == "nhwc")    config.shape = Config::Shape::NHWC;
+		else if (value == "nchw")    config.shape = Config::Shape::NCHW;
+		else throw std::invalid_argument("bad value for: " + key);
+	} else if (key == "channels") {
+		if      (value == "rgb")     config.channels = Config::Channels::RGB;
+		else if (value == "bgr")     config.channels = Config::Channels::BGR;
+		else throw std::invalid_argument("bad value for: " + key);
+	} else if (key == "normalize") {
+		if      (value == "true")    config.normalize = true;
+		else if (value == "false")   config.normalize = false;
+		else throw std::invalid_argument("bad value for: " + key);
+	} else if (key == "pad") {
+		if      (value == "white")   config.pad = Config::Pad::WHITE;
+		else if (value == "edge")    config.pad = Config::Pad::EDGE;
+		else if (value == "stretch") config.pad = Config::Pad::STRETCH;
+		else throw std::invalid_argument("bad value for: " + key);
+	} else if (key == "size") {
+		config.size = std::stoi(value);
+	} else if (key == "interpret") {
+		if      (value == "false")   config.sigmoid = false;
+		else if (value == "sigmoid") config.sigmoid = true;
+		else throw std::invalid_argument("bad value for: " + key);
+	} else {
+		throw std::invalid_argument("unsupported config key: " + key);
+	}
+}
+
+static void
+read_config(Config &config, const char *path)
+{
+	std::ifstream f(path);
+	f.exceptions(std::ifstream::badbit);
+	if (!f)
+		throw std::runtime_error("cannot read configuration");
+
+	std::regex re(R"(^\s*([^#=]+?)\s*=\s*([^#]*?)\s*(?:#|$))",
+		std::regex::optimize);
+	std::smatch m;
+
+	std::string line;
+	while (std::getline(f, line)) {
+		if (std::regex_match(line, m, re))
+			read_field(config, m[1].str(), m[2].str());
+	}
+
+	read_tags(
+		std::filesystem::path(path).replace_extension("tags"), config.tags);
+}
+
+// --- Data preparation --------------------------------------------------------
+
+static float *
+image_to_nhwc(float *data, Magick::Image &image, Config::Channels channels)
+{
+	unsigned int width = image.columns();
+	unsigned int height = image.rows();
+
+	auto pixels = image.getConstPixels(0, 0, width, height);
+	switch (channels) {
+	case Config::Channels::RGB:
+		for (unsigned int y = 0; y < height; y++) {
+			for (unsigned int x = 0; x < width; x++) {
+				auto pixel = *pixels++;
+				*data++ = ScaleQuantumToChar(pixel.red);
+				*data++ = ScaleQuantumToChar(pixel.green);
+				*data++ = ScaleQuantumToChar(pixel.blue);
+			}
+		}
+		break;
+	case Config::Channels::BGR:
+		for (unsigned int y = 0; y < height; y++) {
+			for (unsigned int x = 0; x < width; x++) {
+				auto pixel = *pixels++;
+				*data++ = ScaleQuantumToChar(pixel.blue);
+				*data++ = ScaleQuantumToChar(pixel.green);
+				*data++ = ScaleQuantumToChar(pixel.red);
+			}
+		}
+	}
+	return data;
+}
+
+static float *
+image_to_nchw(float *data, Magick::Image &image, Config::Channels channels)
+{
+	unsigned int width = image.columns();
+	unsigned int height = image.rows();
+
+	auto pixels = image.getConstPixels(0, 0, width, height), pp = pixels;
+	switch (channels) {
+	case Config::Channels::RGB:
+		for (unsigned int y = 0; y < height; y++)
+			for (unsigned int x = 0; x < width; x++)
+				*data++ = ScaleQuantumToChar((*pp++).red);
+		pp = pixels;
+		for (unsigned int y = 0; y < height; y++)
+			for (unsigned int x = 0; x < width; x++)
+				*data++ = ScaleQuantumToChar((*pp++).green);
+		pp = pixels;
+		for (unsigned int y = 0; y < height; y++)
+			for (unsigned int x = 0; x < width; x++)
+				*data++ = ScaleQuantumToChar((*pp++).blue);
+		break;
+	case Config::Channels::BGR:
+		for (unsigned int y = 0; y < height; y++)
+			for (unsigned int x = 0; x < width; x++)
+				*data++ = ScaleQuantumToChar((*pp++).blue);
+		pp = pixels;
+		for (unsigned int y = 0; y < height; y++)
+			for (unsigned int x = 0; x < width; x++)
+				*data++ = ScaleQuantumToChar((*pp++).green);
+		pp = pixels;
+		for (unsigned int y = 0; y < height; y++)
+			for (unsigned int x = 0; x < width; x++)
+				*data++ = ScaleQuantumToChar((*pp++).red);
+	}
+	return data;
+}
+
+static Magick::Image
+load(const std::string filename,
+	const Config &config, int64_t width, int64_t height)
+{
+	Magick::Image image;
+	try {
+		image.read(filename);
+	} catch (const Magick::Warning &warning) {
+		if (g.debug)
+			fprintf(stderr, "%s: %s\n", filename.c_str(), warning.what());
+	}
+
+	image.autoOrient();
+
+	Magick::Geometry adjusted(width, height);
+	switch (config.pad) {
+	case Config::Pad::EDGE:
+	case Config::Pad::WHITE:
+		adjusted.greater(true);
+		break;
+	case Config::Pad::STRETCH:
+		adjusted.aspect(false);
+	}
+
+	image.resize(adjusted, Magick::LanczosFilter);
+
+	// The GraphicsMagick API doesn't offer any good options.
+	if (config.pad == Config::Pad::EDGE) {
+		MagickLib::SetImageVirtualPixelMethod(
+			image.image(), MagickLib::EdgeVirtualPixelMethod);
+
+		auto x = (int64_t(image.columns()) - width) / 2;
+		auto y = (int64_t(image.rows()) - height) / 2;
+		auto source = image.getConstPixels(x, y, width, height);
+		std::vector<MagickLib::PixelPacket>
+			pixels(source, source + width * height);
+
+		Magick::Image edged(Magick::Geometry(width, height), "black");
+		edged.classType(Magick::DirectClass);
+		auto target = edged.setPixels(0, 0, width, height);
+		memcpy(target, pixels.data(), pixels.size() * sizeof pixels[0]);
+		edged.syncPixels();
+
+		image = edged;
+	}
+
+	// Center it in a square patch of white, removing any transparency.
+	// image.extent() could probably be used to do the same thing.
+	Magick::Image white(Magick::Geometry(width, height), "white");
+	auto x = (white.columns() - image.columns()) / 2;
+	auto y = (white.rows() - image.rows()) / 2;
+	white.composite(image, x, y, Magick::OverCompositeOp);
+	white.fileName(filename);
+
+	if (g.debug > 2)
+		white.display();
+
+	return white;
+}
+
+// --- Inference ---------------------------------------------------------------
+
+static void
+run(std::vector<Magick::Image> &images, const Config &config,
+	Ort::Session &session, std::vector<int64_t> shape)
+{
+	// For consistency, this value may be bumped to always be g.batch,
+	// but it does not seem to have an effect on anything.
+	shape[0] = images.size();
+
+	Ort::AllocatorWithDefaultOptions allocator;
+	auto tensor = Ort::Value::CreateTensor<float>(
+		allocator, shape.data(), shape.size());
+
+	auto input_len = tensor.GetTensorTypeAndShapeInfo().GetElementCount();
+	auto input_data = tensor.GetTensorMutableData<float>(), pi = input_data;
+	for (int64_t i = 0; i < images.size(); i++) {
+		switch (config.shape) {
+		case Config::Shape::NCHW:
+			pi = image_to_nchw(pi, images.at(i), config.channels);
+			break;
+		case Config::Shape::NHWC:
+			pi = image_to_nhwc(pi, images.at(i), config.channels);
+		}
+	}
+	if (config.normalize) {
+		pi = input_data;
+		for (size_t i = 0; i < input_len; i++)
+			*pi++ /= 255.0;
+	}
+
+	std::string input_name =
+		session.GetInputNameAllocated(0, allocator).get();
+	std::string output_name =
+		session.GetOutputNameAllocated(0, allocator).get();
+
+	std::vector<const char *> input_names = {input_name.c_str()};
+	std::vector<const char *> output_names = {output_name.c_str()};
+
+	auto outputs = session.Run(Ort::RunOptions{},
+		input_names.data(), &tensor, input_names.size(),
+		output_names.data(), output_names.size());
+	if (outputs.size() != 1 || !outputs[0].IsTensor()) {
+		fprintf(stderr, "Wrong output\n");
+		return;
+	}
+
+	auto output_len = outputs[0].GetTensorTypeAndShapeInfo().GetElementCount();
+	auto output_data = outputs.front().GetTensorData<float>(), po = output_data;
+	if (output_len != shape[0] * config.tags.size()) {
+		fprintf(stderr, "Tags don't match the output\n");
+		return;
+	}
+
+	for (size_t i = 0; i < images.size(); i++) {
+		for (size_t t = 0; t < config.tags.size(); t++) {
+			float value = *po++;
+			if (config.sigmoid)
+				value = 1 / (1 + std::exp(-value));
+			if (value > g.threshold) {
+				printf("%s\t%.2f\t%s\n", images.at(i).fileName().c_str(),
+					value, config.tags.at(t).c_str());
+			}
+		}
+	}
+}
+
+// - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
+
+static void
+parse_options(const std::string &options)
+{
+	auto semicolon = options.find(";");
+	auto name = options.substr(0, semicolon);
+	auto sequence = options.substr(semicolon);
+
+	std::map<std::string, std::string> kv;
+	std::regex re(R"(;*([^;=]+)=([^;=]+))", std::regex::optimize);
+	std::sregex_iterator it(sequence.begin(), sequence.end(), re), end;
+	for (; it != end; ++it)
+		kv[it->str(1)] = it->str(2);
+	g.options.insert_or_assign(name, std::move(kv));
+}
+
+static std::tuple<std::vector<const char *>, std::vector<const char *>>
+unpack_options(const std::string &provider)
+{
+	std::vector<const char *> keys, values;
+	if (g.options.count(provider)) {
+		for (const auto &kv : g.options.at(provider)) {
+			keys.push_back(kv.first.c_str());
+			values.push_back(kv.second.c_str());
+		}
+	}
+	return {keys, values};
+}
+
+static void
+add_providers(Ort::SessionOptions &options)
+{
+	auto api = Ort::GetApi();
+	auto v_providers = Ort::GetAvailableProviders();
+	std::set<std::string> providers(v_providers.begin(), v_providers.end());
+
+	if (g.debug) {
+		printf("Providers:");
+		for (const auto &it : providers)
+			printf(" %s", it.c_str());
+		printf("\n");
+	}
+
+	// There is a string-based AppendExecutionProvider() method,
+	// but it cannot be used with all providers.
+	// TODO: Make it possible to disable providers.
+	// TODO: Providers will deserve some performance tuning.
+
+	if (g.cpu)
+		return;
+
+#ifdef __APPLE__
+	if (providers.count("CoreMLExecutionProvider")) {
+		try {
+			Ort::ThrowOnError(
+				OrtSessionOptionsAppendExecutionProvider_CoreML(options, 0));
+		} catch (const std::exception &e) {
+			fprintf(stderr, "CoreML unavailable: %s\n", e.what());
+		}
+	}
+#endif
+
+#if TENSORRT
+	// TensorRT should be the more performant execution provider, however:
+	//  - it is difficult to set up (needs logging in to download),
+	//  - with WD v1.4 ONNX models, one gets "Your ONNX model has been generated
+	//    with INT64 weights, while TensorRT does not natively support INT64.
+	//    Attempting to cast down to INT32." and that's not nice.
+	if (providers.count("TensorrtExecutionProvider")) {
+		OrtTensorRTProviderOptionsV2* tensorrt_options = nullptr;
+		Ort::ThrowOnError(api.CreateTensorRTProviderOptions(&tensorrt_options));
+		auto [keys, values] = unpack_options("TensorrtExecutionProvider");
+		if (!keys.empty()) {
+			Ort::ThrowOnError(api.UpdateTensorRTProviderOptions(
+				tensorrt_options, keys.data(), values.data(), keys.size()));
+		}
+
+		try {
+			options.AppendExecutionProvider_TensorRT_V2(*tensorrt_options);
+		} catch (const std::exception &e) {
+			fprintf(stderr, "TensorRT unavailable: %s\n", e.what());
+		}
+		api.ReleaseTensorRTProviderOptions(tensorrt_options);
+	}
+#endif
+
+	// See CUDA-ExecutionProvider.html for documentation.
+	if (providers.count("CUDAExecutionProvider")) {
+		OrtCUDAProviderOptionsV2* cuda_options = nullptr;
+		Ort::ThrowOnError(api.CreateCUDAProviderOptions(&cuda_options));
+		auto [keys, values] = unpack_options("CUDAExecutionProvider");
+		if (!keys.empty()) {
+			Ort::ThrowOnError(api.UpdateCUDAProviderOptions(
+				cuda_options, keys.data(), values.data(), keys.size()));
+		}
+
+		try {
+			options.AppendExecutionProvider_CUDA_V2(*cuda_options);
+		} catch (const std::exception &e) {
+			fprintf(stderr, "CUDA unavailable: %s\n", e.what());
+		}
+		api.ReleaseCUDAProviderOptions(cuda_options);
+	}
+
+	if (providers.count("ROCMExecutionProvider")) {
+		OrtROCMProviderOptions rocm_options = {};
+		auto [keys, values] = unpack_options("ROCMExecutionProvider");
+		if (!keys.empty()) {
+			Ort::ThrowOnError(api.UpdateROCMProviderOptions(
+				&rocm_options, keys.data(), values.data(), keys.size()));
+		}
+
+		try {
+			options.AppendExecutionProvider_ROCM(rocm_options);
+		} catch (const std::exception &e) {
+			fprintf(stderr, "ROCM unavailable: %s\n", e.what());
+		}
+	}
+
+	// The CPU provider is the default fallback, if everything else fails.
+}
+
+// - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
+
+struct Thumbnailing {
+	std::mutex input_mutex;
+	std::condition_variable input_cv;
+	std::queue<std::string> input;      // All input paths
+	int work = 0;                       // Number of images requested
+
+	std::mutex output_mutex;
+	std::condition_variable output_cv;
+	std::vector<Magick::Image> output;  // Processed images
+	int done = 0;                       // Finished worker threads
+};
+
+static void
+thumbnail(const Config &config, int64_t width, int64_t height,
+	Thumbnailing &ctx)
+{
+	while (true) {
+		std::unique_lock<std::mutex> input_lock(ctx.input_mutex);
+		ctx.input_cv.wait(input_lock,
+			[&]{ return ctx.input.empty() || ctx.work; });
+		if (ctx.input.empty())
+			break;
+
+		auto path = ctx.input.front();
+		ctx.input.pop();
+		ctx.work--;
+		input_lock.unlock();
+
+		Magick::Image image;
+		try {
+			image = load(path, config, width, height);
+			if (height != image.rows() || width != image.columns())
+				throw std::runtime_error("tensor mismatch");
+
+			std::unique_lock<std::mutex> output_lock(ctx.output_mutex);
+			ctx.output.push_back(image);
+			output_lock.unlock();
+			ctx.output_cv.notify_all();
+		} catch (const std::exception &e) {
+			fprintf(stderr, "%s: %s\n", path.c_str(), e.what());
+
+			std::unique_lock<std::mutex> input_lock(ctx.input_mutex);
+			ctx.work++;
+			input_lock.unlock();
+			ctx.input_cv.notify_all();
+		}
+	}
+
+	std::unique_lock<std::mutex> output_lock(ctx.output_mutex);
+	ctx.done++;
+	output_lock.unlock();
+	ctx.output_cv.notify_all();
+}
+
+// - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
+
+static std::string
+print_shape(const Ort::ConstTensorTypeAndShapeInfo &info)
+{
+	std::vector<const char *> names(info.GetDimensionsCount());
+	info.GetSymbolicDimensions(names.data(), names.size());
+
+	auto shape = info.GetShape();
+	std::string result;
+	for (size_t i = 0; i < shape.size(); i++) {
+		if (shape[i] < 0)
+			result.append(names.at(i));
+		else
+			result.append(std::to_string(shape[i]));
+		result.append(" x ");
+	}
+	if (!result.empty())
+		result.erase(result.size() - 3);
+	return result;
+}
+
+static void
+print_shapes(const Ort::Session &session)
+{
+	Ort::AllocatorWithDefaultOptions allocator;
+	for (size_t i = 0; i < session.GetInputCount(); i++) {
+		std::string name = session.GetInputNameAllocated(i, allocator).get();
+		auto info = session.GetInputTypeInfo(i);
+		auto shape = print_shape(info.GetTensorTypeAndShapeInfo());
+		printf("Input: %s: %s\n", name.c_str(), shape.c_str());
+	}
+	for (size_t i = 0; i < session.GetOutputCount(); i++) {
+		std::string name = session.GetOutputNameAllocated(i, allocator).get();
+		auto info = session.GetOutputTypeInfo(i);
+		auto shape = print_shape(info.GetTensorTypeAndShapeInfo());
+		printf("Output: %s: %s\n", name.c_str(), shape.c_str());
+	}
+}
+
+static void
+infer(Ort::Env &env, const char *path, const std::vector<std::string> &images)
+{
+	Config config;
+	read_config(config, path);
+
+	Ort::SessionOptions session_options;
+	add_providers(session_options);
+
+	Ort::Session session = Ort::Session(env,
+		std::filesystem::path(path).replace_extension("onnx").c_str(),
+		session_options);
+
+	if (g.debug)
+		print_shapes(session);
+
+	if (session.GetInputCount() != 1 || session.GetOutputCount() != 1) {
+		fprintf(stderr, "Invalid input or output shape\n");
+		exit(EXIT_FAILURE);
+	}
+
+	auto input_info = session.GetInputTypeInfo(0);
+	auto shape = input_info.GetTensorTypeAndShapeInfo().GetShape();
+	if (shape.size() != 4) {
+		fprintf(stderr, "Incompatible input tensor format\n");
+		exit(EXIT_FAILURE);
+	}
+	if (shape.at(0) > 1) {
+		fprintf(stderr, "Fixed batching not supported\n");
+		exit(EXIT_FAILURE);
+	}
+	if (shape.at(0) >= 0 && g.batch > 1) {
+		fprintf(stderr, "Requested batching for a non-batching model\n");
+		exit(EXIT_FAILURE);
+	}
+
+	int64_t *height = {}, *width = {}, *channels = {};
+	switch (config.shape) {
+	case Config::Shape::NCHW:
+		channels = &shape[1];
+		height = &shape[2];
+		width = &shape[3];
+		break;
+	case Config::Shape::NHWC:
+		height = &shape[1];
+		width = &shape[2];
+		channels = &shape[3];
+		break;
+	}
+
+	// Variable dimensions don't combine well with batches.
+	if (*height < 0)
+		*height = config.size;
+	if (*width < 0)
+		*width = config.size;
+	if (*channels != 3 || *height < 1 || *width < 1) {
+		fprintf(stderr, "Incompatible input tensor format\n");
+		return;
+	}
+
+	// By only parallelizing image loads here during batching,
+	// they never compete for CPU time with inference.
+	Thumbnailing ctx;
+	for (const auto &path : images)
+		ctx.input.push(path);
+
+	auto workers = g.batch;
+	if (auto threads = std::thread::hardware_concurrency())
+		workers = std::min(workers, long(threads));
+	for (auto i = workers; i--; )
+		std::thread(thumbnail, std::ref(config), *width, *height,
+			std::ref(ctx)).detach();
+
+	while (true) {
+		std::unique_lock<std::mutex> input_lock(ctx.input_mutex);
+		ctx.work = g.batch;
+		input_lock.unlock();
+		ctx.input_cv.notify_all();
+
+		std::unique_lock<std::mutex> output_lock(ctx.output_mutex);
+		ctx.output_cv.wait(output_lock,
+			[&]{ return ctx.output.size() == g.batch || ctx.done == workers; });
+
+		if (!ctx.output.empty()) {
+			run(ctx.output, config, session, shape);
+			ctx.output.clear();
+		}
+		if (ctx.done == workers)
+			break;
+	}
+}
+
+int
+main(int argc, char *argv[])
+{
+	auto invocation_name = argv[0];
+	auto print_usage = [=] {
+		fprintf(stderr,
+			"Usage: %s [-b BATCH] [--cpu] [-d] [-o EP;KEY=VALUE...] "
+			"[-t THRESHOLD] MODEL { --pipe | [IMAGE...] }\n", invocation_name);
+	};
+
+	static option opts[] = {
+		{"batch", required_argument, 0, 'b'},
+		{"cpu", no_argument, 0, 'c'},
+		{"debug", no_argument, 0, 'd'},
+		{"help", no_argument, 0, 'h'},
+		{"options", required_argument, 0, 'o'},
+		{"pipe", no_argument, 0, 'p'},
+		{"threshold", required_argument, 0, 't'},
+		{nullptr, 0, 0, 0},
+	};
+
+	bool pipe = false;
+	while (1) {
+		int option_index = 0;
+		auto c = getopt_long(argc, const_cast<char *const *>(argv),
+			"b:cdho:pt:", opts, &option_index);
+		if (c == -1)
+			break;
+
+		char *end = nullptr;
+		switch (c) {
+		case 'b':
+			errno = 0, g.batch = strtol(optarg, &end, 10);
+			if (errno || *end || g.batch < 1 || g.batch > SHRT_MAX) {
+				fprintf(stderr, "Batch size must be a positive number\n");
+				exit(EXIT_FAILURE);
+			}
+			break;
+		case 'c':
+			g.cpu = true;
+			break;
+		case 'd':
+			g.debug++;
+			break;
+		case 'h':
+			print_usage();
+			return 0;
+		case 'o':
+			parse_options(optarg);
+			break;
+		case 'p':
+			pipe = true;
+			break;
+		case 't':
+			errno = 0, g.threshold = strtod(optarg, &end);
+			if (errno || *end || !std::isfinite(g.threshold) ||
+				g.threshold < 0 || g.threshold > 1) {
+				fprintf(stderr, "Threshold must be a number within 0..1\n");
+				exit(EXIT_FAILURE);
+			}
+			break;
+		default:
+			print_usage();
+			return 1;
+		}
+	}
+
+	argv += optind;
+	argc -= optind;
+
+	// TODO: There's actually no need to slurp all the lines up front.
+	std::vector<std::string> paths;
+	if (pipe) {
+		if (argc != 1) {
+			print_usage();
+			return 1;
+		}
+
+		std::string line;
+		while (std::getline(std::cin, line))
+			paths.push_back(line);
+	} else {
+		if (argc < 1) {
+			print_usage();
+			return 1;
+		}
+
+		paths.assign(argv + 1, argv + argc);
+	}
+
+	// Load batched images in parallel (the first is for GM, the other for IM).
+	if (g.batch > 1) {
+		auto value = std::to_string(
+			std::max(std::thread::hardware_concurrency() / g.batch, 1L));
+		setenv("OMP_NUM_THREADS", value.c_str(), true);
+		setenv("MAGICK_THREAD_LIMIT", value.c_str(), true);
+	}
+
+	// XXX: GraphicsMagick initializes signal handlers here,
+	// one needs to use MagickLib::InitializeMagickEx()
+	// with MAGICK_OPT_NO_SIGNAL_HANDER to prevent that.
+	//
+	// ImageMagick conveniently has the opposite default.
+	Magick::InitializeMagick(nullptr);
+
+	OrtLoggingLevel logging = g.debug > 1
+		? ORT_LOGGING_LEVEL_VERBOSE
+		: ORT_LOGGING_LEVEL_WARNING;
+
+	// Creating an environment before initializing providers in order to avoid:
+	// "Attempt to use DefaultLogger but none has been registered."
+	Ort::Env env(logging, invocation_name);
+	infer(env, argv[0], paths);
+	return 0;
+}

deeptagger/download.sh

@@ -0,0 +1,161 @@
+#!/bin/sh -e
+# Requirements: Python ~ 3.11, curl, unzip, git-lfs, awk
+#
+# This script downloads a bunch of models into the models/ directory,
+# after any necessary transformations to run them using the deeptagger binary.
+#
+# Once it succeeds, feel free to remove everything but *.{model,tags,onnx}
+git lfs install
+mkdir -p models
+cd models
+
+# Create a virtual environment for model conversion.
+#
+# If any of the Python stuff fails,
+# retry from within a Conda environment with a different version of Python.
+export VIRTUAL_ENV=$(pwd)/venv
+export TF_ENABLE_ONEDNN_OPTS=0
+if ! [ -f "$VIRTUAL_ENV/ready" ]
+then
+	python3 -m venv "$VIRTUAL_ENV"
+	#"$VIRTUAL_ENV/bin/pip3" install tensorflow[and-cuda]
+	"$VIRTUAL_ENV/bin/pip3" install tf2onnx 'deepdanbooru[tensorflow]'
+	touch "$VIRTUAL_ENV/ready"
+fi
+
+status() {
+	echo "$(tput bold)-- $*$(tput sgr0)"
+}
+
+# Using the deepdanbooru package makes it possible to use other models
+# trained with the project.
+deepdanbooru() {
+	local name=$1 url=$2
+	status "$name"
+
+	local basename=$(basename "$url")
+	if ! [ -e "$basename" ]
+	then curl -LO "$url"
+	fi
+
+	local modelname=${basename%%.*}
+	if ! [ -d "$modelname" ]
+	then unzip -d "$modelname" "$basename"
+	fi
+
+	if ! [ -e "$modelname.tags" ]
+	then ln "$modelname/tags.txt" "$modelname.tags"
+	fi
+
+	if ! [ -d "$modelname.saved" ]
+	then "$VIRTUAL_ENV/bin/python3" - "$modelname" "$modelname.saved" <<-'END'
+		import sys
+		import deepdanbooru.project as ddp
+		model = ddp.load_model_from_project(
+			project_path=sys.argv[1], compile_model=False)
+		model.export(sys.argv[2])
+	END
+	fi
+
+	if ! [ -e "$modelname.onnx" ]
+	then "$VIRTUAL_ENV/bin/python3" -m tf2onnx.convert \
+		--saved-model "$modelname.saved" --output "$modelname.onnx"
+	fi
+
+	cat > "$modelname.model" <<-END
+		name=$name
+		shape=nhwc
+		channels=rgb
+		normalize=true
+		pad=edge
+	END
+}
+
+# ONNX preconversions don't have a symbolic first dimension, thus doing our own.
+wd14() {
+	local name=$1 repository=$2
+	status "$name"
+
+	local modelname=$(basename "$repository")
+	if ! [ -d "$modelname" ]
+	then git clone "https://huggingface.co/$repository"
+	fi
+
+	# Though link the original export as well.
+	if ! [ -e "$modelname.onnx" ]
+	then ln "$modelname/model.onnx" "$modelname.onnx"
+	fi
+
+	if ! [ -e "$modelname.tags" ]
+	then awk -F, 'NR > 1 { print $2 }' "$modelname/selected_tags.csv" \
+		> "$modelname.tags"
+	fi
+
+	cat > "$modelname.model" <<-END
+		name=$name
+		shape=nhwc
+		channels=bgr
+		normalize=false
+		pad=white
+	END
+
+	if ! [ -e "batch-$modelname.onnx" ]
+	then "$VIRTUAL_ENV/bin/python3" -m tf2onnx.convert \
+		--saved-model "$modelname" --output "batch-$modelname.onnx"
+	fi
+
+	if ! [ -e "batch-$modelname.tags" ]
+	then ln "$modelname.tags" "batch-$modelname.tags"
+	fi
+
+	if ! [ -e "batch-$modelname.model" ]
+	then ln "$modelname.model" "batch-$modelname.model"
+	fi
+}
+
+# These models are an undocumented mess, thus using ONNX preconversions.
+mldanbooru() {
+	local name=$1 basename=$2
+	status "$name"
+
+	if ! [ -d ml-danbooru-onnx ]
+	then git clone https://huggingface.co/deepghs/ml-danbooru-onnx
+	fi
+
+	local modelname=${basename%%.*}
+	if ! [ -e "$basename" ]
+	then ln "ml-danbooru-onnx/$basename"
+	fi
+
+	if ! [ -e "$modelname.tags" ]
+	then awk -F, 'NR > 1 { print $1 }' ml-danbooru-onnx/tags.csv \
+		> "$modelname.tags"
+	fi
+
+	cat > "$modelname.model" <<-END
+		name=$name
+		shape=nchw
+		channels=rgb
+		normalize=true
+		pad=stretch
+		size=640
+		interpret=sigmoid
+	END
+}
+
+status "Downloading models, beware that git-lfs doesn't indicate progress"
+
+deepdanbooru DeepDanbooru \
+	'https://github.com/KichangKim/DeepDanbooru/releases/download/v3-20211112-sgd-e28/deepdanbooru-v3-20211112-sgd-e28.zip'
+
+#wd14 'WD v1.4 ViT v1'        'SmilingWolf/wd-v1-4-vit-tagger'
+wd14 'WD v1.4 ViT v2'        'SmilingWolf/wd-v1-4-vit-tagger-v2'
+#wd14 'WD v1.4 ConvNeXT v1'   'SmilingWolf/wd-v1-4-convnext-tagger'
+wd14 'WD v1.4 ConvNeXT v2'   'SmilingWolf/wd-v1-4-convnext-tagger-v2'
+wd14 'WD v1.4 ConvNeXTV2 v2' 'SmilingWolf/wd-v1-4-convnextv2-tagger-v2'
+wd14 'WD v1.4 SwinV2 v2'     'SmilingWolf/wd-v1-4-swinv2-tagger-v2'
+wd14 'WD v1.4 MOAT v2'       'SmilingWolf/wd-v1-4-moat-tagger-v2'
+
+# As suggested by author https://github.com/IrisRainbowNeko/ML-Danbooru-webui
+mldanbooru 'ML-Danbooru Caformer dec-5-97527' 'ml_caformer_m36_dec-5-97527.onnx'
+mldanbooru 'ML-Danbooru TResNet-D 6-30000' 'TResnet-D-FLq_ema_6-30000.onnx'