open-standards/wai/tools/wai_int_video_sota_keyframe_encode.py

#!/usr/bin/env python3
"""Encode a REAL 256x256 int_hyper-keyframe clip: the I-frame is the integer
learned codec int_hyper (the pretrained mbt2018_mean exported to integers; its
byte-exact 256x256 reconstruction, dumped by the `int_hyper_decode_frame` Rust
example), and the P-frames pan it with integer block motion compensation +
integer-quantized residual. Needs the int_hyper vector dir (default
/tmp/sota_int) holding frame0.rgb, and exits non-zero without it.
tools/wai_sota_vectors.sh runs the whole chain.

  cargo run --release --no-default-features --features neural_int --example int_hyper_decode_frame -- <dir> <dir>/frame0.rgb
  python tools/wai_int_video_sota_keyframe_encode.py <dir>

Emits <dir>/video_sota_bitstream.json + <dir>/video_sota.json (the P-frame stream +
expected per-frame / clip hashes; frame 0's hash is int_hyper's). The Rust test
`int_video_sota_keyframe` composes int_hyper_synth::decode_conditional_checked with
int_video::decode_pframes and checks these (skipping loudly when they are absent,
failing instead under WAI_REQUIRE_SOTA_VECTORS=1).
"""
import sys, os, json, base64
import numpy as np
from compressai.ans import RansEncoder

H = W = 256  # default; main() infers the real size from frame0.rgb
BLOCK, SEARCH, Q, NF = 16, 4, 6, 8
RMAX = (255 + Q - 1) // Q


def fnv1a(buf):
    h = 14695981039346656037
    for b in buf:
        h = ((h ^ b) * 1099511628211) & 0xFFFFFFFFFFFFFFFF
    return h


def laplacian_cdf():
    n = 2 * RMAX + 1
    pmf = np.exp(-np.abs(np.arange(n) - RMAX) / 6.0)
    total = 1 << 16
    f = np.maximum((pmf / pmf.sum() * (total - 1)).round().astype(np.int64), 1)
    f[f.argmax()] += (total - 1) - int(f.sum())
    cdf = [0]
    for v in f:
        cdf.append(cdf[-1] + int(v))
    cdf.append(total)
    return cdf


def mc_predict(ref, mvs):
    pred = np.empty_like(ref)
    nbx = (W + BLOCK - 1) // BLOCK
    for by in range(0, H, BLOCK):
        for bx in range(0, W, BLOCK):
            dy, dx = mvs[(by // BLOCK) * nbx + (bx // BLOCK)]
            ys = np.clip(np.arange(by, min(by + BLOCK, H)) + dy, 0, H - 1)
            xs = np.clip(np.arange(bx, min(bx + BLOCK, W)) + dx, 0, W - 1)
            pred[by:by + len(ys), bx:bx + len(xs)] = ref[np.ix_(ys, xs)]
    return pred


def search_mvs(cur, ref):
    mvs = []
    for by in range(0, H, BLOCK):
        for bx in range(0, W, BLOCK):
            cb = cur[by:by + BLOCK, bx:bx + BLOCK]
            best, bmv = None, (0, 0)
            for dy in range(-SEARCH, SEARCH + 1):
                for dx in range(-SEARCH, SEARCH + 1):
                    ys = np.clip(np.arange(by, by + cb.shape[0]) + dy, 0, H - 1)
                    xs = np.clip(np.arange(bx, bx + cb.shape[1]) + dx, 0, W - 1)
                    sad = int(np.abs(cb - ref[np.ix_(ys, xs)]).sum())
                    if best is None or sad < best:
                        best, bmv = sad, (dy, dx)
            mvs.append([int(bmv[0]), int(bmv[1])])
    return mvs


def main(d, prefix='video_sota'):
    global H, W
    fp = f'{d}/frame0.rgb'
    if not os.path.exists(fp):
        # A producer that exits 0 without producing lets every consumer downstream
        # skip in turn, and the whole chain reports green having checked nothing.
        sys.exit(f'error: {fp} absent (run the int_hyper_decode_frame example first)')
    raw = open(fp, 'rb').read()
    H = W = int(round((len(raw) / 3) ** 0.5))   # square frame inferred from the dump
    frame0 = np.frombuffer(raw, np.uint8).reshape(H, W, 3).astype(np.int64)
    cdf, offset = laplacian_cdf(), -RMAX
    frames = [frame0] + [np.roll(frame0, (3 * f, 2 * f), axis=(0, 1)) for f in range(1, NF)]  # pan
    recon, mvs_all, syms, hashes = [frame0], [[]], [], []
    hashes.append(format(fnv1a(np.ascontiguousarray(frame0.astype(np.uint8)).tobytes()), '016x'))
    for f in range(1, NF):
        mvs = search_mvs(frames[f], recon[-1])
        pred = np.stack([mc_predict(recon[-1][..., c], mvs) for c in range(3)], -1)
        rq = np.round((frames[f] - pred) / Q).astype(np.int64).clip(-RMAX, RMAX)
        syms.extend(rq.reshape(-1).tolist())
        rec = np.clip(pred + rq * Q, 0, 255).astype(np.int64)
        recon.append(rec)
        mvs_all.append(mvs)
        hashes.append(format(fnv1a(np.ascontiguousarray(rec.astype(np.uint8)).tobytes()), '016x'))

    bits = RansEncoder().encode_with_indexes(syms, [0] * len(syms), [cdf], [len(cdf)], [offset])
    clip_hash = format(fnv1a(b''.join(bytes.fromhex(h) for h in hashes)), '016x')
    json.dump({'block': BLOCK, 'q': Q, 'rmax': RMAX, 'offset': offset, 'cdf': cdf,
               'n_frames': NF, 'h': H, 'w': W, 'mvs': mvs_all,
               'rans_b64': base64.b64encode(bytes(bits)).decode()},
              open(f'{d}/{prefix}_bitstream.json', 'w'))
    json.dump({'n_frames': NF, 'h': H, 'w': W, 'frame_hashes': hashes, 'clip_hash': clip_hash},
              open(f'{d}/{prefix}.json', 'w'))
    print(f'int_hyper-keyframe video [{prefix}]: {NF}x {H}x{W} (I-frame={hashes[0]}), {len(bits)} P rANS bytes '
          f'-> clip_hash {clip_hash}')


if __name__ == '__main__':
    main(sys.argv[1] if len(sys.argv) > 1 else '/tmp/sota_int', sys.argv[2] if len(sys.argv) > 2 else 'video_sota')