Datasets:

cp500
/

hyperkd-rs

"""End-to-end codec demo on the WORLD corpus: one exemplar track per concept.
For each exemplar: build the codec track from its shard (polygon from the
teacher mask, 7-channel spectral fill from the raw multispec, exact-date
clock), hide a third of its observations with CLOUD mask tokens, let codec v3
impute them, and render ground truth vs decode side by side.
Writes renders/world_codec_<tag>.png + world_codec_report.json and the
Space asset spaces/hrm-space/assets/worldtour/*.png.
Usage: python world_codec.py   (needs shards in experiments/recall/)
"""
import json, os, pathlib, sys
import numpy as np
import cv2
import torch

HERE = pathlib.Path(__file__).parent
sys.path.insert(0, str(HERE))
os.environ["PLAYBACK_ONLY"] = "1"
import codec_v3 as C3                      # loads codec_v3.pt, no training
from codec_mlm import encode_shape, decode_shape, iou_poly
from codec_time import timefeat, F_MU, F_SD, L_MU, L_SD, NV, CC

SH = HERE.parent / "experiments/recall"
from huggingface_hub import hf_hub_download

EXEMPLARS = [  # (shard tag, concept substring, display name)
    ("kansas",      "center pivot",        "center-pivot field · Kansas"),
    ("guayas",      "aquaculture pond",    "shrimp pond · Guayas"),
    ("parkfire_ev", "burned",              "burn scar · Park Fire 2024"),
    ("indus_ev",    "flood",               "flooded farmland · Indus 2010/22"),
    ("atacama",     "mine",                "open-pit mine · Atacama"),
    ("oregon",      "clearcut",            "clearcut · Oregon"),
    ("aralsea",     "lake",                "shrinking lake · Aral Sea"),
]

def months_since(datestr):
    y, mo = int(str(datestr)[:4]), int(str(datestr)[5:7])
    return (y - 2023) * 12 + (mo - 1)

def fill7(px):                              # px: (N,4) r,g,b,nir reflectance
    r, g, b, nir = px[:, 0].mean(), px[:, 1].mean(), px[:, 2].mean(), px[:, 3].mean()
    s = r + g + b + 1e-6
    return np.array([r, g, b, (nir - r) / (nir + r + 1e-6),
                     (2 * g - r - b) / s, 1 - np.clip(s / 3 / 0.12, 0, 1), s / 3],
                    np.float32)

def build_track(tag, concept_sub):
    sh = np.load(SH / f"shard_{tag}.npz")
    lf = hf_hub_download("cp500/hyperkd-rs", f"labels/labels_{tag}.npz", repo_type="dataset")
    lb = np.load(lf)
    names = lb["concepts"]; sl = lb["sample_labels"]
    samples, masks, images = sh["samples"], sh["masks"], sh["images"]
    dates = [str(d) for d in sh["dates"]]
    ms = sh["multispec"].astype(np.float32)
    want = {int(r[0]) for r in sl if concept_sub in str(names[int(r[1])])}
    # longest track containing a wanted sample
    by_track = {}
    for i, r in enumerate(samples):
        by_track.setdefault(int(r[7]), []).append(i)
    cands = [(len(v), t) for t, v in by_track.items() if any(i in want for i in v)]
    if not cands: return None
    cands.sort(reverse=True)
    _, tid = cands[0]
    steps = []
    for i in sorted(by_track[tid], key=lambda i: int(samples[i][0])):
        r = samples[i]; si = int(r[0]); mi = int(r[3 + int(r[6])])
        if mi < 0: continue
        m8 = (masks[mi] > 0).astype(np.uint8)
        cs, _ = cv2.findContours(m8, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
        if not cs: continue
        cnt = max(cs, key=cv2.contourArea)
        if cv2.contourArea(cnt) < 20 or len(cnt) < 8: continue
        P = cnt[:, 0, :].astype(float) * (800 / 256)
        try: vec, pose = encode_shape(P)
        except Exception: continue
        m160 = cv2.resize(m8, (160, 160), interpolation=cv2.INTER_NEAREST) > 0
        px = ms[si][m160]
        if len(px) < 4: continue
        steps.append(dict(vec=vec, pose=pose, logs=pose[1], fill=fill7(px),
                          tm=months_since(dates[si]), period=dates[si][:7],
                          si=si, m8=m8))
    if len(steps) < 5: return None
    s = dict(tag=tag, steps=steps)
    s["X"] = np.stack([np.concatenate([st["vec"], [(st["logs"] - L_MU) / L_SD],
                                       (st["fill"] - F_MU) / F_SD]).astype(np.float32)
                       for st in steps])
    s["T"] = np.stack([timefeat(st["tm"]) for st in steps])
    return s

def stretch_rgb(fill):
    return tuple(int(np.clip((max(float(v), 0) / 0.35) ** (1 / 2.2), 0, 1) * 255)
                 for v in fill[:3][::-1])   # BGR for cv2

OUTD = HERE / "renders"; OUTD.mkdir(exist_ok=True)
SPACED = HERE.parent / "spaces/hrm-space/assets/worldtour"; SPACED.mkdir(parents=True, exist_ok=True)
report = []
TILE = 128
for tag, sub, title in EXEMPLARS:
    s = build_track(tag, sub)
    if s is None:
        print(f"{tag}: no track for '{sub}'"); continue
    T = len(s["steps"])
    hide = list(range(1, T, 3))             # hide every 3rd obs (CLOUD)
    pred = C3.impute(s, hide)               # rows for hidden idxs
    # metrics vs persistence
    ious, pious, fmae, pfmae = [], [], [], []
    sh_img = np.load(SH / f"shard_{tag}.npz")["images"]
    for k, t in enumerate(hide):
        st, ref = s["steps"][t], s["steps"][t - 1]
        e = pred[k]
        vec = e[:NV]; logs = e[NV] * L_SD + L_MU; fill = e[NV + 1:] * F_SD + F_MU
        pose = (st["pose"][0], logs, ref["pose"][2], ref["pose"][3])
        ious.append(iou_poly(decode_shape(vec, pose), decode_shape(st["vec"], st["pose"])))
        pious.append(iou_poly(decode_shape(ref["vec"], ref["pose"]),
                              decode_shape(st["vec"], st["pose"])))
        fmae.append(np.abs(fill - st["fill"]).mean())
        pfmae.append(np.abs(ref["fill"] - st["fill"]).mean())
        s["steps"][t]["dec"] = (vec, pose, fill)
    # render strip: up to 10 columns, GT on top, decode below
    cols = [t for t in range(T)][:: max(1, T // 10)][:10]
    tiles_top, tiles_bot = [], []
    # common crop around the track's union bbox (256 grid)
    u = np.zeros((256, 256), np.uint8)
    for st in s["steps"]: u |= st["m8"]
    ys, xs = np.where(u > 0)
    cx, cy = int(xs.mean()), int(ys.mean())
    r = max(int(max(xs.max()-xs.min(), ys.max()-ys.min()) * 0.8), 40)
    x0, y0 = np.clip(cx - r, 0, 255), np.clip(cy - r, 0, 255)
    x1, y1 = np.clip(cx + r, 1, 256), np.clip(cy + r, 1, 256)
    for t in cols:
        st = s["steps"][t]
        img = sh_img[st["si"]][y0:y1, x0:x1].copy()
        cs, _ = cv2.findContours(st["m8"][y0:y1, x0:x1], cv2.RETR_EXTERNAL,
                                 cv2.CHAIN_APPROX_SIMPLE)
        cv2.drawContours(img, cs, -1, (90, 220, 90), 1)
        gt = cv2.resize(img, (TILE, TILE), interpolation=cv2.INTER_NEAREST)
        gt = cv2.copyMakeBorder(gt, 16, 0, 0, 0, cv2.BORDER_CONSTANT, value=(15, 15, 15))
        cv2.putText(gt, st["period"], (3, 12), cv2.FONT_HERSHEY_SIMPLEX, 0.33,
                    (230, 230, 230), 1)
        tiles_top.append(gt[..., ::-1])     # to BGR
        bot = np.full((TILE, TILE, 3), 24, np.uint8)
        if t in hide and "dec" in st:
            vec, pose, fill = st["dec"]
            poly = decode_shape(vec, pose) * (256 / 800)
            poly = (poly - [x0, y0]) * (TILE / max(x1 - x0, 1))
            cv2.fillPoly(bot, [poly.astype(np.int32)], stretch_rgb(fill))
            cv2.polylines(bot, [poly.astype(np.int32)], True, (80, 160, 255), 1)
            tagtxt, col = "IMPUTED", (80, 160, 255)
        else:
            poly = st["m8"][y0:y1, x0:x1]
            cs, _ = cv2.findContours(poly, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
            scale = TILE / max(x1 - x0, 1)
            for c in cs:
                cv2.fillPoly(bot, [(c[:, 0, :] * scale).astype(np.int32)],
                             stretch_rgb(st["fill"]))
            tagtxt, col = "OBSERVED", (120, 200, 120)
        bot = cv2.copyMakeBorder(bot, 16, 0, 0, 0, cv2.BORDER_CONSTANT, value=(15, 15, 15))
        cv2.putText(bot, tagtxt, (3, 12), cv2.FONT_HERSHEY_SIMPLEX, 0.33, col, 1)
        tiles_bot.append(bot)
    strip = np.vstack([np.hstack(tiles_top), np.hstack(tiles_bot)])
    fn = f"world_codec_{tag}.png"
    cv2.imwrite(str(OUTD / fn), strip)
    cv2.imwrite(str(SPACED / fn), strip)
    rec = dict(tag=tag, title=title, obs=T, hidden=len(hide),
               iou=round(float(np.mean(ious)), 3),
               pers_iou=round(float(np.mean(pious)), 3),
               fmae=round(float(np.mean(fmae)), 4),
               pers_fmae=round(float(np.mean(pfmae)), 4))
    report.append(rec); print(rec)
json.dump(report, open(HERE / "world_codec_report.json", "w"), indent=1)
json.dump(report, open(SPACED / "report.json", "w"), indent=1)
print(f"-> {len(report)} concept tracks rendered")

	"""End-to-end codec demo on the WORLD corpus: one exemplar track per concept.

	For each exemplar: build the codec track from its shard (polygon from the
	teacher mask, 7-channel spectral fill from the raw multispec, exact-date
	clock), hide a third of its observations with CLOUD mask tokens, let codec v3
	impute them, and render ground truth vs decode side by side.

	Writes renders/world_codec_<tag>.png + world_codec_report.json and the
	Space asset spaces/hrm-space/assets/worldtour/*.png.

	Usage: python world_codec.py (needs shards in experiments/recall/)
	"""
	import json, os, pathlib, sys
	import numpy as np
	import cv2
	import torch

	HERE = pathlib.Path(__file__).parent
	sys.path.insert(0, str(HERE))
	os.environ["PLAYBACK_ONLY"] = "1"
	import codec_v3 as C3 # loads codec_v3.pt, no training
	from codec_mlm import encode_shape, decode_shape, iou_poly
	from codec_time import timefeat, F_MU, F_SD, L_MU, L_SD, NV, CC

	SH = HERE.parent / "experiments/recall"
	from huggingface_hub import hf_hub_download

	EXEMPLARS = [ # (shard tag, concept substring, display name)
	("kansas", "center pivot", "center-pivot field · Kansas"),
	("guayas", "aquaculture pond", "shrimp pond · Guayas"),
	("parkfire_ev", "burned", "burn scar · Park Fire 2024"),
	("indus_ev", "flood", "flooded farmland · Indus 2010/22"),
	("atacama", "mine", "open-pit mine · Atacama"),
	("oregon", "clearcut", "clearcut · Oregon"),
	("aralsea", "lake", "shrinking lake · Aral Sea"),
	]

	def months_since(datestr):
	y, mo = int(str(datestr)[:4]), int(str(datestr)[5:7])
	return (y - 2023) * 12 + (mo - 1)

	def fill7(px): # px: (N,4) r,g,b,nir reflectance
	r, g, b, nir = px[:, 0].mean(), px[:, 1].mean(), px[:, 2].mean(), px[:, 3].mean()
	s = r + g + b + 1e-6
	return np.array([r, g, b, (nir - r) / (nir + r + 1e-6),
	(2 * g - r - b) / s, 1 - np.clip(s / 3 / 0.12, 0, 1), s / 3],
	np.float32)

	def build_track(tag, concept_sub):
	sh = np.load(SH / f"shard_{tag}.npz")
	lf = hf_hub_download("cp500/hyperkd-rs", f"labels/labels_{tag}.npz", repo_type="dataset")
	lb = np.load(lf)
	names = lb["concepts"]; sl = lb["sample_labels"]
	samples, masks, images = sh["samples"], sh["masks"], sh["images"]
	dates = [str(d) for d in sh["dates"]]
	ms = sh["multispec"].astype(np.float32)
	want = {int(r[0]) for r in sl if concept_sub in str(names[int(r[1])])}
	# longest track containing a wanted sample
	by_track = {}
	for i, r in enumerate(samples):
	by_track.setdefault(int(r[7]), []).append(i)
	cands = [(len(v), t) for t, v in by_track.items() if any(i in want for i in v)]
	if not cands: return None
	cands.sort(reverse=True)
	_, tid = cands[0]
	steps = []
	for i in sorted(by_track[tid], key=lambda i: int(samples[i][0])):
	r = samples[i]; si = int(r[0]); mi = int(r[3 + int(r[6])])
	if mi < 0: continue
	m8 = (masks[mi] > 0).astype(np.uint8)
	cs, _ = cv2.findContours(m8, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
	if not cs: continue
	cnt = max(cs, key=cv2.contourArea)
	if cv2.contourArea(cnt) < 20 or len(cnt) < 8: continue
	P = cnt[:, 0, :].astype(float) * (800 / 256)
	try: vec, pose = encode_shape(P)
	except Exception: continue
	m160 = cv2.resize(m8, (160, 160), interpolation=cv2.INTER_NEAREST) > 0
	px = ms[si][m160]
	if len(px) < 4: continue
	steps.append(dict(vec=vec, pose=pose, logs=pose[1], fill=fill7(px),
	tm=months_since(dates[si]), period=dates[si][:7],
	si=si, m8=m8))
	if len(steps) < 5: return None
	s = dict(tag=tag, steps=steps)
	s["X"] = np.stack([np.concatenate([st["vec"], [(st["logs"] - L_MU) / L_SD],
	(st["fill"] - F_MU) / F_SD]).astype(np.float32)
	for st in steps])
	s["T"] = np.stack([timefeat(st["tm"]) for st in steps])
	return s

	def stretch_rgb(fill):
	return tuple(int(np.clip((max(float(v), 0) / 0.35) ** (1 / 2.2), 0, 1) * 255)
	for v in fill[:3][::-1]) # BGR for cv2

	OUTD = HERE / "renders"; OUTD.mkdir(exist_ok=True)
	SPACED = HERE.parent / "spaces/hrm-space/assets/worldtour"; SPACED.mkdir(parents=True, exist_ok=True)
	report = []
	TILE = 128
	for tag, sub, title in EXEMPLARS:
	s = build_track(tag, sub)
	if s is None:
	print(f"{tag}: no track for '{sub}'"); continue
	T = len(s["steps"])
	hide = list(range(1, T, 3)) # hide every 3rd obs (CLOUD)
	pred = C3.impute(s, hide) # rows for hidden idxs
	# metrics vs persistence
	ious, pious, fmae, pfmae = [], [], [], []
	sh_img = np.load(SH / f"shard_{tag}.npz")["images"]
	for k, t in enumerate(hide):
	st, ref = s["steps"][t], s["steps"][t - 1]
	e = pred[k]
	vec = e[:NV]; logs = e[NV] * L_SD + L_MU; fill = e[NV + 1:] * F_SD + F_MU
	pose = (st["pose"][0], logs, ref["pose"][2], ref["pose"][3])
	ious.append(iou_poly(decode_shape(vec, pose), decode_shape(st["vec"], st["pose"])))
	pious.append(iou_poly(decode_shape(ref["vec"], ref["pose"]),
	decode_shape(st["vec"], st["pose"])))
	fmae.append(np.abs(fill - st["fill"]).mean())
	pfmae.append(np.abs(ref["fill"] - st["fill"]).mean())
	s["steps"][t]["dec"] = (vec, pose, fill)
	# render strip: up to 10 columns, GT on top, decode below
	cols = [t for t in range(T)][:: max(1, T // 10)][:10]
	tiles_top, tiles_bot = [], []
	# common crop around the track's union bbox (256 grid)
	u = np.zeros((256, 256), np.uint8)
	for st in s["steps"]: u \|= st["m8"]
	ys, xs = np.where(u > 0)
	cx, cy = int(xs.mean()), int(ys.mean())
	r = max(int(max(xs.max()-xs.min(), ys.max()-ys.min()) * 0.8), 40)
	x0, y0 = np.clip(cx - r, 0, 255), np.clip(cy - r, 0, 255)
	x1, y1 = np.clip(cx + r, 1, 256), np.clip(cy + r, 1, 256)
	for t in cols:
	st = s["steps"][t]
	img = sh_img[st["si"]][y0:y1, x0:x1].copy()
	cs, _ = cv2.findContours(st["m8"][y0:y1, x0:x1], cv2.RETR_EXTERNAL,
	cv2.CHAIN_APPROX_SIMPLE)
	cv2.drawContours(img, cs, -1, (90, 220, 90), 1)
	gt = cv2.resize(img, (TILE, TILE), interpolation=cv2.INTER_NEAREST)
	gt = cv2.copyMakeBorder(gt, 16, 0, 0, 0, cv2.BORDER_CONSTANT, value=(15, 15, 15))
	cv2.putText(gt, st["period"], (3, 12), cv2.FONT_HERSHEY_SIMPLEX, 0.33,
	(230, 230, 230), 1)
	tiles_top.append(gt[..., ::-1]) # to BGR
	bot = np.full((TILE, TILE, 3), 24, np.uint8)
	if t in hide and "dec" in st:
	vec, pose, fill = st["dec"]
	poly = decode_shape(vec, pose) * (256 / 800)
	poly = (poly - [x0, y0]) * (TILE / max(x1 - x0, 1))
	cv2.fillPoly(bot, [poly.astype(np.int32)], stretch_rgb(fill))
	cv2.polylines(bot, [poly.astype(np.int32)], True, (80, 160, 255), 1)
	tagtxt, col = "IMPUTED", (80, 160, 255)
	else:
	poly = st["m8"][y0:y1, x0:x1]
	cs, _ = cv2.findContours(poly, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
	scale = TILE / max(x1 - x0, 1)
	for c in cs:
	cv2.fillPoly(bot, [(c[:, 0, :] * scale).astype(np.int32)],
	stretch_rgb(st["fill"]))
	tagtxt, col = "OBSERVED", (120, 200, 120)
	bot = cv2.copyMakeBorder(bot, 16, 0, 0, 0, cv2.BORDER_CONSTANT, value=(15, 15, 15))
	cv2.putText(bot, tagtxt, (3, 12), cv2.FONT_HERSHEY_SIMPLEX, 0.33, col, 1)
	tiles_bot.append(bot)
	strip = np.vstack([np.hstack(tiles_top), np.hstack(tiles_bot)])
	fn = f"world_codec_{tag}.png"
	cv2.imwrite(str(OUTD / fn), strip)
	cv2.imwrite(str(SPACED / fn), strip)
	rec = dict(tag=tag, title=title, obs=T, hidden=len(hide),
	iou=round(float(np.mean(ious)), 3),
	pers_iou=round(float(np.mean(pious)), 3),
	fmae=round(float(np.mean(fmae)), 4),
	pers_fmae=round(float(np.mean(pfmae)), 4))
	report.append(rec); print(rec)
	json.dump(report, open(HERE / "world_codec_report.json", "w"), indent=1)
	json.dump(report, open(SPACED / "report.json", "w"), indent=1)
	print(f"-> {len(report)} concept tracks rendered")