Files
scene-actor-extraction/src/gallery/gallery_store.cpp
Claude 7db40f430d GR-004: bind galleries to the embedder that built them
A gallery is only valid for the embedder that produced its vectors. Cosine
similarities across models are meaningless but *look* plausible, so the mistake
is silent and every measurement taken afterwards is suspect. Stamp the embedder
identity into the gallery at build; verify it at every load.

The stamp is the model file's basename plus the SHA-256 of its bytes (plus
embed_dim). The hash decides, the name explains. A name alone is a promise
rather than a fact — models get re-exported and overwritten in place under an
unchanged filename, which is exactly the case where the weights differ and
nothing else does. A hash alone is correct but unactionable in an error message.
SHA-256 is derived from the artefact, needs no registry kept current, and costs
~0.1s for a 250MB ONNX, memoised per process.

Mismatch is a hard error in every mode, with no bypass, naming both sides.

Unstamped legacy galleries warn loudly and proceed: unknown is not known-bad,
and hard-failing every pre-existing gallery would turn the check into something
people disable rather than trust. --require-gallery-stamp (or
SAE_REQUIRE_GALLERY_STAMP=1, which propagates to subprocesses) promotes that to
a hard error — the mode measurement work should run in. scripts/stamp_gallery.py
re-binds an existing gallery with no re-embedding, so "warn" is a cheap state to
leave rather than a permanent one.

Embedding dumps carry the same stamp: a replay has no live embedder, so the dump
is the embedder as far as the gallery is concerned. Derived galleries inherit
their source's stamp; --merge and the JSON gallery merge check before writing,
since one file holding two embedding spaces cannot be untangled afterwards.

Verified in: scene_analyze, scene_preview, the sae_kpn matcher binding,
replay.py, optimize.py (once per film at startup, before the first evaluation),
movienet_eval.py and both merge paths.

Stamp logic lives in src/gallery/embedder_stamp.{hpp,cpp} and its Python twin
scripts/sae_stamp.py, kept dependency-light so replay subprocesses do not pay
sae_gallery's requests/Pillow import to ask whether two models match.

Tests: 12 new cases in test_gallery_store.cpp covering the comparison logic,
both round trips, and the SHA-256 vectors that guarantee the C++ and hashlib
stamps agree. No ONNX or GPU required.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-07-30 18:35:46 +02:00

270 lines
11 KiB
C++

#include "gallery_store.hpp"
#include <nlohmann/json.hpp>
#include <H5Cpp.h>
#include <chrono>
#include <fstream>
#include <iostream>
#include <stdexcept>
using json = nlohmann::json;
// ── HDF5 (see gallery_store.hpp for the full layout) ─────────────────────────
// A 170MB gallery JSON parses in ~18s (nlohmann). The same data as HDF5 loads in
// ~1s — a big win for the optimizer, which reloads the gallery per replay subprocess.
static bool ends_with(const std::string& s, const std::string& suf) {
return s.size() >= suf.size() &&
s.compare(s.size() - suf.size(), suf.size(), suf) == 0;
}
static std::vector<std::string> read_str_dataset(H5::H5File& file, const char* name, hsize_t a) {
H5::DataSet ds = file.openDataSet(name);
H5::StrType st = ds.getStrType();
std::vector<std::string> out(a);
if (st.isVariableStr()) {
std::vector<char*> raw(a);
ds.read(raw.data(), st);
for (hsize_t i = 0; i < a; ++i) { out[i] = raw[i] ? raw[i] : ""; }
H5::DataSpace sp = ds.getSpace();
H5Dvlen_reclaim(st.getId(), sp.getId(), H5P_DEFAULT, raw.data());
}
return out;
}
static ActorGallery load_gallery_hdf5(const std::string& path) {
std::cerr << "[gallery] loading " << path << " (HDF5)..." << std::flush;
auto t0 = std::chrono::steady_clock::now();
H5::H5File file(path, H5F_ACC_RDONLY);
H5::DataSet emb_ds = file.openDataSet("embeddings");
hsize_t dims[2];
emb_ds.getSpace().getSimpleExtentDims(dims); // [N, 512]
const hsize_t N = dims[0];
if (dims[1] != 512) throw std::runtime_error("gallery HDF5: embedding dim != 512");
std::vector<float> flat(N * 512);
emb_ds.read(flat.data(), H5::PredType::NATIVE_FLOAT);
H5::DataSet off_ds = file.openDataSet("offset");
hsize_t adim[1];
off_ds.getSpace().getSimpleExtentDims(adim);
const hsize_t A = adim[0];
std::vector<int64_t> offset(A);
off_ds.read(offset.data(), H5::PredType::NATIVE_INT64);
std::vector<int32_t> count(A);
file.openDataSet("count").read(count.data(), H5::PredType::NATIVE_INT32);
auto imdb = read_str_dataset(file, "imdb_id", A);
auto tmdb = read_str_dataset(file, "tmdb_id", A);
auto jf = read_str_dataset(file, "jellyfin_id", A);
auto name = read_str_dataset(file, "name", A);
std::vector<std::string> src_images;
if (file.nameExists("source_images"))
src_images = read_str_dataset(file, "source_images", N);
ActorGallery gallery;
gallery.actors.reserve(A);
for (hsize_t a = 0; a < A; ++a) {
ActorGallery::Actor actor;
actor.imdb_id = imdb[a]; actor.tmdb_id = tmdb[a];
actor.jellyfin_id = jf[a]; actor.name = name[a];
for (int32_t e = 0; e < count[a]; ++e) {
hsize_t row = offset[a] + e;
Embedding emb;
std::copy_n(flat.data() + row * 512, 512, emb.begin());
actor.embeddings.push_back(emb);
if (!src_images.empty())
actor.source_images.push_back(src_images[row]);
}
gallery.actors.push_back(std::move(actor));
}
/// TRACES: GR-004 | SR-001
// Absent /embedder group == a gallery written before model binding existed.
// It stays readable; verify_gallery_embedder() decides what that means.
if (file.nameExists("embedder")) {
H5::Group eg = file.openGroup("embedder");
H5::StrType str(H5::PredType::C_S1, H5T_VARIABLE);
if (eg.attrExists("model_name"))
eg.openAttribute("model_name").read(str, gallery.embedder.model_name);
if (eg.attrExists("model_sha256"))
eg.openAttribute("model_sha256").read(str, gallery.embedder.model_sha256);
if (eg.attrExists("embed_dim"))
eg.openAttribute("embed_dim").read(H5::PredType::NATIVE_INT32,
&gallery.embedder.embed_dim);
}
if (file.nameExists("calibration")) {
H5::Group cal = file.openGroup("calibration");
cal.openAttribute("a").read(H5::PredType::NATIVE_FLOAT, &gallery.calib_a);
cal.openAttribute("b").read(H5::PredType::NATIVE_FLOAT, &gallery.calib_b);
int8_t valid = 0;
cal.openAttribute("valid").read(H5::PredType::NATIVE_INT8, &valid);
gallery.calib_valid = valid != 0;
cal.openAttribute("hash").read(H5::PredType::NATIVE_UINT64, &gallery.calib_hash);
}
auto t1 = std::chrono::steady_clock::now();
std::cerr << " built " << A << " actors / " << N << " embeddings in "
<< std::chrono::duration<double>(t1 - t0).count() << "s";
if (gallery.calib_hash != 0)
std::cerr << " (calibration cached: a=" << gallery.calib_a
<< " b=" << gallery.calib_b << " valid=" << gallery.calib_valid << ")";
std::cerr << "\n";
return gallery;
}
static void write_str_dataset(H5::H5File& file, const char* name,
const std::vector<std::string>& values) {
H5::StrType str_t(H5::PredType::C_S1, H5T_VARIABLE);
hsize_t n = values.size();
H5::DataSpace space(1, &n);
H5::DataSet ds = file.createDataSet(name, str_t, space);
std::vector<const char*> raw(n);
for (hsize_t i = 0; i < n; ++i) raw[i] = values[i].c_str();
ds.write(raw.data(), str_t);
}
static void save_gallery_hdf5(const std::string& path, const ActorGallery& gallery) {
H5::H5File file(path, H5F_ACC_TRUNC);
std::vector<float> flat;
std::vector<int64_t> offset;
std::vector<int32_t> count;
std::vector<std::string> imdb, tmdb, jf, name, src_images;
int64_t row = 0;
for (const auto& a : gallery.actors) {
offset.push_back(row);
count.push_back(static_cast<int32_t>(a.embeddings.size()));
row += static_cast<int64_t>(a.embeddings.size());
for (size_t i = 0; i < a.embeddings.size(); ++i) {
flat.insert(flat.end(), a.embeddings[i].begin(), a.embeddings[i].end());
src_images.push_back(i < a.source_images.size() ? a.source_images[i] : "");
}
imdb.push_back(a.imdb_id); tmdb.push_back(a.tmdb_id);
jf.push_back(a.jellyfin_id); name.push_back(a.name);
}
hsize_t N = flat.size() / 512;
hsize_t emb_dims[2] = {N, 512};
H5::DataSpace emb_space(2, emb_dims);
file.createDataSet("embeddings", H5::PredType::NATIVE_FLOAT, emb_space)
.write(flat.data(), H5::PredType::NATIVE_FLOAT);
hsize_t A = gallery.actors.size();
H5::DataSpace a_space(1, &A);
file.createDataSet("offset", H5::PredType::NATIVE_INT64, a_space)
.write(offset.data(), H5::PredType::NATIVE_INT64);
file.createDataSet("count", H5::PredType::NATIVE_INT32, a_space)
.write(count.data(), H5::PredType::NATIVE_INT32);
write_str_dataset(file, "imdb_id", imdb);
write_str_dataset(file, "tmdb_id", tmdb);
write_str_dataset(file, "jellyfin_id", jf);
write_str_dataset(file, "name", name);
write_str_dataset(file, "source_images", src_images);
/// TRACES: GR-004 | SR-001
// Bind the file to the embedder that produced its vectors. Written only when
// known — an empty stamp must round-trip as "unstamped", not as a stamp
// claiming an unnamed model.
if (!gallery.embedder.empty()) {
H5::Group eg = file.createGroup("embedder");
H5::DataSpace scalar(H5S_SCALAR);
H5::StrType str(H5::PredType::C_S1, H5T_VARIABLE);
eg.createAttribute("model_name", str, scalar).write(str, gallery.embedder.model_name);
eg.createAttribute("model_sha256", str, scalar).write(str, gallery.embedder.model_sha256);
eg.createAttribute("embed_dim", H5::PredType::NATIVE_INT32, scalar)
.write(H5::PredType::NATIVE_INT32, &gallery.embedder.embed_dim);
}
if (gallery.calib_hash != 0) {
H5::Group cal = file.createGroup("calibration");
H5::DataSpace scalar(H5S_SCALAR);
cal.createAttribute("a", H5::PredType::NATIVE_FLOAT, scalar)
.write(H5::PredType::NATIVE_FLOAT, &gallery.calib_a);
cal.createAttribute("b", H5::PredType::NATIVE_FLOAT, scalar)
.write(H5::PredType::NATIVE_FLOAT, &gallery.calib_b);
int8_t valid = gallery.calib_valid ? 1 : 0;
cal.createAttribute("valid", H5::PredType::NATIVE_INT8, scalar)
.write(H5::PredType::NATIVE_INT8, &valid);
cal.createAttribute("hash", H5::PredType::NATIVE_UINT64, scalar)
.write(H5::PredType::NATIVE_UINT64, &gallery.calib_hash);
}
std::cerr << "[gallery] saved " << A << " actors / " << N
<< " embeddings to " << path << " (HDF5)\n";
}
ActorGallery load_gallery(const std::string& path) {
if (ends_with(path, ".h5") || ends_with(path, ".hdf5"))
return load_gallery_hdf5(path);
std::ifstream f(path);
if (!f.is_open())
throw std::runtime_error("load_gallery: cannot open " + path);
std::cerr << "[gallery] loading " << path << "..." << std::flush;
auto t0 = std::chrono::steady_clock::now();
json j;
f >> j;
auto t1 = std::chrono::steady_clock::now();
std::cerr << " parsed JSON in "
<< std::chrono::duration<double>(t1 - t0).count() << "s\n";
ActorGallery gallery;
/// TRACES: GR-004 | SR-001
// Optional top-level "embedder" object, matching the HDF5 /embedder group.
// Written by the JSON-era helper scripts; absent in anything older.
if (j.contains("embedder") && j.at("embedder").is_object()) {
const auto& je = j.at("embedder");
gallery.embedder.model_name = je.value("model_name", "");
gallery.embedder.model_sha256 = je.value("model_sha256", "");
gallery.embedder.embed_dim = je.value("embed_dim", 512);
}
for (const auto& ja : j.at("actors")) {
ActorGallery::Actor actor;
actor.imdb_id = ja.value("imdb_id", "");
actor.tmdb_id = ja.value("tmdb_id", "");
// older make_jellyfin_gallery.py galleries used "jellyfin_person_id"
actor.jellyfin_id = ja.value("jellyfin_id", ja.value("jellyfin_person_id", ""));
actor.name = ja.at("name").get<std::string>();
if (ja.contains("source_images"))
actor.source_images = ja.at("source_images").get<std::vector<std::string>>();
for (const auto& je : ja.at("embeddings")) {
Embedding emb = je.get<Embedding>();
actor.embeddings.push_back(emb);
}
gallery.actors.push_back(std::move(actor));
}
size_t n_emb = 0;
for (const auto& actor : gallery.actors) n_emb += actor.embeddings.size();
auto t2 = std::chrono::steady_clock::now();
std::cerr << "[gallery] built " << gallery.actors.size() << " actors / "
<< n_emb << " embeddings in "
<< std::chrono::duration<double>(t2 - t1).count() << "s\n";
return gallery;
}
// Always writes HDF5. If `path` doesn't already end in .h5/.hdf5, the
// extension is replaced (galleries are never written as JSON anymore).
void save_gallery(const std::string& path, const ActorGallery& gallery) {
std::string out_path = path;
if (!ends_with(out_path, ".h5") && !ends_with(out_path, ".hdf5")) {
auto dot = out_path.find_last_of('.');
out_path = (dot == std::string::npos ? out_path : out_path.substr(0, dot)) + ".h5";
std::cerr << "[gallery] save_gallery: writing HDF5 to " << out_path
<< " (galleries are no longer written as JSON)\n";
}
save_gallery_hdf5(out_path, gallery);
}