Datasets:

b4ph
/

clap-retrieval-benchmark

"""Shared fixtures and metrics for the audio-text retrieval harness.
Protocol (v1, pinned):
  * Query sets come from HF datasets whose audio is embedded in the parquet,
    so nothing is downloaded from YouTube, Kaggle, or Zenodo.
  * One "clip" is a unique row-group id. Its captions are the distinct caption
    strings that appear for that id.
  * Text-to-audio (T2A): every caption is a query; the only relevant item is its
    own clip. Standard single-positive AP.
  * Audio-to-text (A2T): every clip is a query; all captions of that clip are
    relevant. R@k counts a hit if any relevant caption lands in the top k.
  * Similarity is cosine on L2-normalized embeddings.
"""
from __future__ import annotations

import json
import time
from pathlib import Path

import numpy as np

DATASETS = {
    # repo, split, glob of only the shards for that split, id/text/audio columns.
    # The glob matters: loading the repo by split name still fetches every split's
    # shards (11 GB for AudioCaps), while these globs pull only the eval shards.
    "clotho": {
        "repo": "CLAPv2/Clotho",
        "split": "test",
        "files": "data/test/*.parquet",
        "id_col": "index",
        "text_col": "text",
        "audio_col": "audio",
        "captions_per_clip": 1,
    },
    "audiocaps": {
        "repo": "CLAPv2/audiocaps",
        "split": "test",
        "files": "data/test/*.parquet",
        "id_col": "index",
        "text_col": "text",
        "audio_col": "audio",
        "captions_per_clip": 1,
    },
    # Same audio as clotho5 but only the first reference, so the effect of caption
    # multiplicity can be separated from the effect of the audio itself.
    "clotho1": {
        "repo": "gijs/clotho",
        "split": "test",
        "files": None,
        "id_col": "file_name",
        "caption_cols": ["caption_1"],
        "audio_col": "audio",
        "captions_per_clip": 1,
    },
    # The five references joined into one string, matching what CLAPv2/Clotho ships.
    # Same audio and same words as clotho5, so it isolates text length from audio.
    "clotho5cat": {
        "repo": "gijs/clotho",
        "split": "test",
        "files": None,
        "id_col": "file_name",
        "caption_cols": ["caption_1", "caption_2", "caption_3", "caption_4", "caption_5"],
        "join_captions": " ",
        "audio_col": "audio",
        "captions_per_clip": 1,
    },
    # Five references per clip on the same audio as clotho1: this is the fixture
    # that puts the harness on the published protocol and exercises multi-positive AP.
    "clotho5": {
        "repo": "gijs/clotho",
        "split": "test",
        "files": None,  # parquet native; load by split name
        "id_col": "file_name",
        "caption_cols": ["caption_1", "caption_2", "caption_3", "caption_4", "caption_5"],
        "audio_col": "audio",
        "captions_per_clip": 5,
    },
}

SAMPLE_RATE = 48_000  # CLAP feature extractor expects 48 kHz


def load_rows(name: str, limit: int | None = None, revision: str | None = None):
    from datasets import load_dataset

    from datasets import Audio

    spec = DATASETS[name]
    if spec.get("files"):
        # Pull only the shards for this split, then treat them as one split.
        ds = load_dataset(spec["repo"], data_files=spec["files"], split="train",
                          revision=revision)
    else:
        ds = load_dataset(spec["repo"], split=spec["split"], revision=revision)
    # Keep audio as raw encoded bytes and decode with soundfile below. datasets>=3
    # wants torchcodec for its own decoder, and the torchcodec wheel tracks a newer
    # CUDA runtime than the cu128 image provides.
    ds = ds.cast_column(spec["audio_col"], Audio(decode=False))
    if limit:
        ds = ds.select(range(min(limit, len(ds))))
    return ds, spec


def _clip_key(row, spec) -> str:
    """Identity of a clip is its audio, not its row id.
    CLAPv2/audiocaps repeats one clip across caption rows and its `index` column is
    a shard label, not a clip id, so grouping on it collapses the whole split into
    one clip. Hash the encoded audio instead.
    """
    import hashlib

    audio = row.get(spec["audio_col"])
    if isinstance(audio, dict) and audio.get("bytes"):
        return "clip_" + hashlib.sha1(audio["bytes"]).hexdigest()[:16]
    if isinstance(audio, dict) and audio.get("path"):
        return "clip_" + hashlib.sha1(str(audio["path"]).encode()).hexdigest()[:16]
    return str(row[spec["id_col"]])


def group_clips(ds, spec) -> list[dict]:
    """Collapse rows into unique clips, first-seen order.
    Two fixture shapes are supported:
      * caption_cols present  -> one row per clip carrying N caption columns
        (clotho5: caption_1..caption_5), so every clip keeps all references;
      * text_col               -> one caption per row, collapsed by clip identity.
    """
    audio_col = spec["audio_col"]
    caption_cols = spec.get("caption_cols")
    text_col = spec.get("text_col")
    order: list[str] = []
    clips: dict[str, dict] = {}
    for row in ds:
        cid = _clip_key(row, spec)
        if cid not in clips:
            clips[cid] = {"clip_id": cid, "captions": [], "audio": row[audio_col]}
            order.append(cid)
        if caption_cols:
            texts = [(row.get(c) or "").strip() for c in caption_cols]
            if spec.get("join_captions"):
                # one caption built from all references, the CLAPv2 convention
                joined = spec["join_captions"].join(t for t in texts if t)
                texts = [joined]
        else:
            texts = [(row.get(text_col) or "").strip()]
        for text in texts:
            if text and text not in clips[cid]["captions"]:
                clips[cid]["captions"].append(text)
    return [clips[c] for c in order if clips[c]["captions"]]


def fixture_summary(clips: list[dict]) -> dict:
    """Counts plus a stable hash, so a v2 run can be compared to the fixture spec."""
    import hashlib

    h = hashlib.sha1()
    for c in clips:
        h.update(("|".join([c["clip_id"], *c["captions"]])).encode())
        h.update(b"\x00")
    caps = [len(c["captions"]) for c in clips]
    return {
        "clips": len(clips),
        "captions": int(sum(caps)),
        "min_captions_per_clip": int(min(caps)) if caps else 0,
        "max_captions_per_clip": int(max(caps)) if caps else 0,
        "fixture_sha": h.hexdigest()[:16],
    }


def to_mono_48k(audio, orig_sr: int) -> np.ndarray:
    """audio: {'bytes','path'} (decode=False), {'array','sampling_rate'}, or np array."""
    import io

    import librosa
    import soundfile as sf

    if isinstance(audio, dict) and audio.get("bytes"):
        arr, orig_sr = sf.read(io.BytesIO(audio["bytes"]), dtype="float32", always_2d=False)
        arr = np.asarray(arr, dtype=np.float32)
    elif isinstance(audio, dict):
        arr = np.asarray(audio["array"], dtype=np.float32)
        orig_sr = int(audio.get("sampling_rate", orig_sr))
    else:
        arr = np.asarray(audio, dtype=np.float32)
    if arr.ndim > 1:
        arr = arr.mean(axis=1)
    if orig_sr != SAMPLE_RATE:
        arr = librosa.resample(arr, orig_sr=orig_sr, target_sr=SAMPLE_RATE)
    return np.ascontiguousarray(arr, dtype=np.float32)


def l2norm(x: np.ndarray) -> np.ndarray:
    n = np.linalg.norm(x, axis=-1, keepdims=True)
    return x / np.clip(n, 1e-12, None)


def retrieval_metrics(sim: np.ndarray, relevance: list[set[int]], ks=(1, 5, 10),
                      per_query: bool = False, clusters: list[int] | None = None) -> dict:
    """sim: (n_query, n_target), higher is better. relevance: per-query target indices.
    With per_query=True the result also carries the two arrays a bootstrap needs:
    the 1-based rank of the first relevant item for every query, and the AP of
    every query. R@k and MRR are recoverable from the ranks, mAP from the APs.
    """
    n_query = sim.shape[0]
    # stable sort so ties break by target index rather than by memory order
    ranks = np.argsort(-sim, axis=1, kind="stable")
    out = {}
    for k in ks:
        hits = 0
        for i in range(n_query):
            top = ranks[i, :k]
            if any(int(j) in relevance[i] for j in top):
                hits += 1
        out[f"r@{k}"] = hits / n_query if n_query else 0.0

    aps: list[float] = []
    first_ranks: list[int] = []
    for i in range(n_query):
        rel = relevance[i]
        if not rel:
            continue
        found = 0
        prec_sum = 0.0
        first = 0
        for pos, j in enumerate(ranks[i], start=1):
            if int(j) in rel:
                found += 1
                if first == 0:
                    first = pos
                prec_sum += found / pos
                if found == len(rel):
                    break
        aps.append(prec_sum / len(rel))
        first_ranks.append(first if first else len(ranks[i]) + 1)
    out["map"] = float(np.mean(aps)) if aps else 0.0
    out["n_queries"] = n_query
    if per_query:
        out["first_hit_rank"] = first_ranks
        out["ap"] = aps
        # Queries inside one clip are not independent (five captions describe the
        # same audio), so a bootstrap has to resample clips, not queries.
        out["cluster"] = list(clusters) if clusters is not None else list(range(n_query))
    return out


def build_problem(clips: list[dict], audio_emb: np.ndarray, text_emb: np.ndarray,
                  per_query: bool = False):
    """Return (t2a_metrics, a2t_metrics) from per-clip and per-caption embeddings."""
    clip_index = {c["clip_id"]: i for i, c in enumerate(clips)}
    cap_owner: list[int] = []
    for i, c in enumerate(clips):
        cap_owner.extend([i] * len(c["captions"]))
    assert len(cap_owner) == text_emb.shape[0], (len(cap_owner), text_emb.shape)

    a = l2norm(audio_emb)
    t = l2norm(text_emb)

    # T2A: each caption queries for its own clip
    sim_t2a = t @ a.T
    rel_t2a = [{owner} for owner in cap_owner]
    t2a = retrieval_metrics(sim_t2a, rel_t2a, per_query=per_query, clusters=cap_owner)

    # A2T: each clip queries for any of its captions
    sim_a2t = a @ t.T
    rel_a2t = [set(j for j, o in enumerate(cap_owner) if o == i) for i in range(len(clips))]
    a2t = retrieval_metrics(sim_a2t, rel_a2t, per_query=per_query)

    return t2a, a2t, clip_index


def write_result(path: Path, payload: dict) -> None:
    path.parent.mkdir(parents=True, exist_ok=True)
    # newline="\n" keeps artifacts byte-identical across platforms; the default
    # would emit CRLF on Windows and break byte-level review of the release.
    path.write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n",
                    encoding="utf-8", newline="\n")
    print(f"[harness] wrote {path}")


def now() -> str:
    return time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())

	"""Shared fixtures and metrics for the audio-text retrieval harness.

	Protocol (v1, pinned):
	* Query sets come from HF datasets whose audio is embedded in the parquet,
	so nothing is downloaded from YouTube, Kaggle, or Zenodo.
	* One "clip" is a unique row-group id. Its captions are the distinct caption
	strings that appear for that id.
	* Text-to-audio (T2A): every caption is a query; the only relevant item is its
	own clip. Standard single-positive AP.
	* Audio-to-text (A2T): every clip is a query; all captions of that clip are
	relevant. R@k counts a hit if any relevant caption lands in the top k.
	* Similarity is cosine on L2-normalized embeddings.
	"""
	from __future__ import annotations

	import json
	import time
	from pathlib import Path

	import numpy as np

	DATASETS = {
	# repo, split, glob of only the shards for that split, id/text/audio columns.
	# The glob matters: loading the repo by split name still fetches every split's
	# shards (11 GB for AudioCaps), while these globs pull only the eval shards.
	"clotho": {
	"repo": "CLAPv2/Clotho",
	"split": "test",
	"files": "data/test/*.parquet",
	"id_col": "index",
	"text_col": "text",
	"audio_col": "audio",
	"captions_per_clip": 1,
	},
	"audiocaps": {
	"repo": "CLAPv2/audiocaps",
	"split": "test",
	"files": "data/test/*.parquet",
	"id_col": "index",
	"text_col": "text",
	"audio_col": "audio",
	"captions_per_clip": 1,
	},
	# Same audio as clotho5 but only the first reference, so the effect of caption
	# multiplicity can be separated from the effect of the audio itself.
	"clotho1": {
	"repo": "gijs/clotho",
	"split": "test",
	"files": None,
	"id_col": "file_name",
	"caption_cols": ["caption_1"],
	"audio_col": "audio",
	"captions_per_clip": 1,
	},
	# The five references joined into one string, matching what CLAPv2/Clotho ships.
	# Same audio and same words as clotho5, so it isolates text length from audio.
	"clotho5cat": {
	"repo": "gijs/clotho",
	"split": "test",
	"files": None,
	"id_col": "file_name",
	"caption_cols": ["caption_1", "caption_2", "caption_3", "caption_4", "caption_5"],
	"join_captions": " ",
	"audio_col": "audio",
	"captions_per_clip": 1,
	},
	# Five references per clip on the same audio as clotho1: this is the fixture
	# that puts the harness on the published protocol and exercises multi-positive AP.
	"clotho5": {
	"repo": "gijs/clotho",
	"split": "test",
	"files": None, # parquet native; load by split name
	"id_col": "file_name",
	"caption_cols": ["caption_1", "caption_2", "caption_3", "caption_4", "caption_5"],
	"audio_col": "audio",
	"captions_per_clip": 5,
	},
	}

	SAMPLE_RATE = 48_000 # CLAP feature extractor expects 48 kHz


	def load_rows(name: str, limit: int \| None = None, revision: str \| None = None):
	from datasets import load_dataset

	from datasets import Audio

	spec = DATASETS[name]
	if spec.get("files"):
	# Pull only the shards for this split, then treat them as one split.
	ds = load_dataset(spec["repo"], data_files=spec["files"], split="train",
	revision=revision)
	else:
	ds = load_dataset(spec["repo"], split=spec["split"], revision=revision)
	# Keep audio as raw encoded bytes and decode with soundfile below. datasets>=3
	# wants torchcodec for its own decoder, and the torchcodec wheel tracks a newer
	# CUDA runtime than the cu128 image provides.
	ds = ds.cast_column(spec["audio_col"], Audio(decode=False))
	if limit:
	ds = ds.select(range(min(limit, len(ds))))
	return ds, spec


	def _clip_key(row, spec) -> str:
	"""Identity of a clip is its audio, not its row id.

	CLAPv2/audiocaps repeats one clip across caption rows and its `index` column is
	a shard label, not a clip id, so grouping on it collapses the whole split into
	one clip. Hash the encoded audio instead.
	"""
	import hashlib

	audio = row.get(spec["audio_col"])
	if isinstance(audio, dict) and audio.get("bytes"):
	return "clip_" + hashlib.sha1(audio["bytes"]).hexdigest()[:16]
	if isinstance(audio, dict) and audio.get("path"):
	return "clip_" + hashlib.sha1(str(audio["path"]).encode()).hexdigest()[:16]
	return str(row[spec["id_col"]])


	def group_clips(ds, spec) -> list[dict]:
	"""Collapse rows into unique clips, first-seen order.

	Two fixture shapes are supported:
	* caption_cols present -> one row per clip carrying N caption columns
	(clotho5: caption_1..caption_5), so every clip keeps all references;
	* text_col -> one caption per row, collapsed by clip identity.
	"""
	audio_col = spec["audio_col"]
	caption_cols = spec.get("caption_cols")
	text_col = spec.get("text_col")
	order: list[str] = []
	clips: dict[str, dict] = {}
	for row in ds:
	cid = _clip_key(row, spec)
	if cid not in clips:
	clips[cid] = {"clip_id": cid, "captions": [], "audio": row[audio_col]}
	order.append(cid)
	if caption_cols:
	texts = [(row.get(c) or "").strip() for c in caption_cols]
	if spec.get("join_captions"):
	# one caption built from all references, the CLAPv2 convention
	joined = spec["join_captions"].join(t for t in texts if t)
	texts = [joined]
	else:
	texts = [(row.get(text_col) or "").strip()]
	for text in texts:
	if text and text not in clips[cid]["captions"]:
	clips[cid]["captions"].append(text)
	return [clips[c] for c in order if clips[c]["captions"]]


	def fixture_summary(clips: list[dict]) -> dict:
	"""Counts plus a stable hash, so a v2 run can be compared to the fixture spec."""
	import hashlib

	h = hashlib.sha1()
	for c in clips:
	h.update(("\|".join([c["clip_id"], *c["captions"]])).encode())
	h.update(b"\x00")
	caps = [len(c["captions"]) for c in clips]
	return {
	"clips": len(clips),
	"captions": int(sum(caps)),
	"min_captions_per_clip": int(min(caps)) if caps else 0,
	"max_captions_per_clip": int(max(caps)) if caps else 0,
	"fixture_sha": h.hexdigest()[:16],
	}


	def to_mono_48k(audio, orig_sr: int) -> np.ndarray:
	"""audio: {'bytes','path'} (decode=False), {'array','sampling_rate'}, or np array."""
	import io

	import librosa
	import soundfile as sf

	if isinstance(audio, dict) and audio.get("bytes"):
	arr, orig_sr = sf.read(io.BytesIO(audio["bytes"]), dtype="float32", always_2d=False)
	arr = np.asarray(arr, dtype=np.float32)
	elif isinstance(audio, dict):
	arr = np.asarray(audio["array"], dtype=np.float32)
	orig_sr = int(audio.get("sampling_rate", orig_sr))
	else:
	arr = np.asarray(audio, dtype=np.float32)
	if arr.ndim > 1:
	arr = arr.mean(axis=1)
	if orig_sr != SAMPLE_RATE:
	arr = librosa.resample(arr, orig_sr=orig_sr, target_sr=SAMPLE_RATE)
	return np.ascontiguousarray(arr, dtype=np.float32)


	def l2norm(x: np.ndarray) -> np.ndarray:
	n = np.linalg.norm(x, axis=-1, keepdims=True)
	return x / np.clip(n, 1e-12, None)


	def retrieval_metrics(sim: np.ndarray, relevance: list[set[int]], ks=(1, 5, 10),
	per_query: bool = False, clusters: list[int] \| None = None) -> dict:
	"""sim: (n_query, n_target), higher is better. relevance: per-query target indices.

	With per_query=True the result also carries the two arrays a bootstrap needs:
	the 1-based rank of the first relevant item for every query, and the AP of
	every query. R@k and MRR are recoverable from the ranks, mAP from the APs.
	"""
	n_query = sim.shape[0]
	# stable sort so ties break by target index rather than by memory order
	ranks = np.argsort(-sim, axis=1, kind="stable")
	out = {}
	for k in ks:
	hits = 0
	for i in range(n_query):
	top = ranks[i, :k]
	if any(int(j) in relevance[i] for j in top):
	hits += 1
	out[f"r@{k}"] = hits / n_query if n_query else 0.0

	aps: list[float] = []
	first_ranks: list[int] = []
	for i in range(n_query):
	rel = relevance[i]
	if not rel:
	continue
	found = 0
	prec_sum = 0.0
	first = 0
	for pos, j in enumerate(ranks[i], start=1):
	if int(j) in rel:
	found += 1
	if first == 0:
	first = pos
	prec_sum += found / pos
	if found == len(rel):
	break
	aps.append(prec_sum / len(rel))
	first_ranks.append(first if first else len(ranks[i]) + 1)
	out["map"] = float(np.mean(aps)) if aps else 0.0
	out["n_queries"] = n_query
	if per_query:
	out["first_hit_rank"] = first_ranks
	out["ap"] = aps
	# Queries inside one clip are not independent (five captions describe the
	# same audio), so a bootstrap has to resample clips, not queries.
	out["cluster"] = list(clusters) if clusters is not None else list(range(n_query))
	return out


	def build_problem(clips: list[dict], audio_emb: np.ndarray, text_emb: np.ndarray,
	per_query: bool = False):
	"""Return (t2a_metrics, a2t_metrics) from per-clip and per-caption embeddings."""
	clip_index = {c["clip_id"]: i for i, c in enumerate(clips)}
	cap_owner: list[int] = []
	for i, c in enumerate(clips):
	cap_owner.extend([i] * len(c["captions"]))
	assert len(cap_owner) == text_emb.shape[0], (len(cap_owner), text_emb.shape)

	a = l2norm(audio_emb)
	t = l2norm(text_emb)

	# T2A: each caption queries for its own clip
	sim_t2a = t @ a.T
	rel_t2a = [{owner} for owner in cap_owner]
	t2a = retrieval_metrics(sim_t2a, rel_t2a, per_query=per_query, clusters=cap_owner)

	# A2T: each clip queries for any of its captions
	sim_a2t = a @ t.T
	rel_a2t = [set(j for j, o in enumerate(cap_owner) if o == i) for i in range(len(clips))]
	a2t = retrieval_metrics(sim_a2t, rel_a2t, per_query=per_query)

	return t2a, a2t, clip_index


	def write_result(path: Path, payload: dict) -> None:
	path.parent.mkdir(parents=True, exist_ok=True)
	# newline="\n" keeps artifacts byte-identical across platforms; the default
	# would emit CRLF on Windows and break byte-level review of the release.
	path.write_text(json.dumps(payload, indent=2, sort_keys=True) + "\n",
	encoding="utf-8", newline="\n")
	print(f"[harness] wrote {path}")


	def now() -> str:
	return time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())