Migrate HOT3D, OakInk-v1, OakInk-v2 and FreiHAND to shared-v1.1 (#4)

10ed3aa about 1 month ago

11.4 kB

#!/usr/bin/env python3
"""Repack pinned HOT3D shared-v1.0 inputs into a non-media v1.1 staging tree.
No download, source deletion, geometry reconstruction, or HF upload. Unchanged
media/assets remain an explicit publisher reuse plan until separately assembled.
"""
import argparse
from collections import Counter
import fcntl
import hashlib
import json
from pathlib import Path
import shutil
import time

import torch

from hot3d_bundle import at, audit_view, check_field, exact, fields, new_metadata, safe_path

REVISION = "54c1bc812b903c2f2b933472163998924588681f"
PILOT = {"P0015_179e1b84", "P0004_1984bf99"}


def write_json(path, obj):
    path = Path(path); path.parent.mkdir(parents=True, exist_ok=True)
    tmp = path.with_name(path.name + ".writing")
    tmp.write_text(json.dumps(obj, ensure_ascii=False, indent=2) + "\n")
    tmp.replace(path)


def digest(path, algorithm="sha256"):
    size = path.stat().st_size
    h = hashlib.sha1(f"blob {size}\0".encode()) if algorithm == "git_blob_sha1" else hashlib.sha256()
    with path.open("rb") as stream:
        for chunk in iter(lambda: stream.read(2**20), b""):
            h.update(chunk)
    return h.hexdigest()


def verify(path, row):
    if not path.is_file() or path.stat().st_size != row['size'] or digest(path, row['algorithm']) != row['digest']:
        raise ValueError(f"Pinned source size/digest mismatch: {path}")


def atomic_copy(source, target):
    target.parent.mkdir(parents=True, exist_ok=True)
    tmp = target.with_name(target.name + '.copying')
    shutil.copy2(source, tmp)
    tmp.replace(target)


def migrate(workspace, scope, revalidate_tools=False):
    torch.set_num_threads(1)
    workspace = Path(workspace).resolve()
    source = workspace/'source/HOT3D/shared-v1.0'
    state = workspace/'tmp/hot3d-migration-20261014'/scope
    state.mkdir(parents=True, exist_ok=True)
    target = workspace/('tmp/hot3d-migration-20261014/pilot/shared-v1.1' if scope == 'pilot' else 'staging/HOT3D/shared-v1.1')
    if target.resolve() == source.resolve() or target.resolve().is_relative_to(source.resolve()):
        raise ValueError('Never migrate inside the read-only source cache')
    manifest_path = workspace/'tmp/remaining-preparation-20261014/HOT3D-manifest.json'
    manifest = json.loads(manifest_path.read_text())
    if manifest['revision'] != REVISION or manifest['dataset'] != 'HOT3D':
        raise ValueError('Wrong fixed source')
    registry_path = workspace/'spec/hand-warehouse-1.1/schema.json'
    registry = json.loads(registry_path.read_text())
    binding = {'source_revision': REVISION, 'source_manifest_sha256': digest(manifest_path),
               'schema_sha256': digest(registry_path),
               'tools': {p: digest(Path(__file__).parent/p) for p in ['migrate_hot3d.py','hot3d_bundle.py']}}
    lock = (state.parent/'migration.lock').open('a+')
    fcntl.flock(lock, fcntl.LOCK_EX | fcntl.LOCK_NB)
    binding_path = state/'binding.json'
    previous_binding = json.loads(binding_path.read_text()) if binding_path.exists() else None
    if previous_binding is not None and previous_binding != binding:
        if not revalidate_tools or any(previous_binding[k] != binding[k] for k in ['source_revision','source_manifest_sha256','schema_sha256']):
            raise ValueError('Resume source/tool binding changed; review before --revalidate-tools')
        # Explicitly reviewed tool-only upgrade: old checkpoints/hashes must
        # still match and all views are compared again. Adopt only on success.
        write_json(state/'binding-before-revalidation.json', previous_binding)
    if not binding_path.exists() and target.exists():
        raise ValueError('Existing unregistered staging tree; do not overwrite')
    if previous_binding is None:
        write_json(binding_path, binding)
    started = time.monotonic()
    entries = [r for r in manifest['files'] if scope == 'all' or not r['relative'].startswith('data/') or r['relative'].split('/')[1] in PILOT]
    selected = [r for r in entries if r['action'].startswith('download_')]
    status = {'phase': 'verifying_inputs', 'scope': scope, 'source_revision': REVISION,
              'complete_release': False, 'verified_views': 0}
    write_json(state/'status.json', status)
    for i, row in enumerate(selected,1):
        verify(source/row['relative'],row)
        if i % 2000 == 0: print(f'input hashes {i}/{len(selected)}',flush=True)
    metas = sorted(r['relative'] for r in entries if r['relative'].startswith('data/') and len(Path(r['relative']).parts)==4 and r['relative'].endswith('/metainfo.json'))
    expected_views = 6 if scope == 'pilot' else 594
    if len(metas) != expected_views:
        raise ValueError(f'Expected {expected_views} complete views, got {len(metas)}')
    expected_new, removed, plans = set(), set(), {}
    for relative in metas:
        old = json.loads((source/relative).read_text())
        new = new_metadata(old)
        v = Path(relative).parent
        fs = fields(old)
        for f in fs:
            c = at(old,f['locator'])
            removed.update((v/c[k]).as_posix() for k in ['data_ref','valid_ref'])
            expected_new.add((v/f['bundle']).as_posix())
        plans[v.as_posix()] = (old,new,fs)
    # Precisely retain RGB validity; asset validity belongs to unchanged assets.
    selected_data = {r['relative']: r for r in selected if r['relative'].startswith('data/')}
    actual_pt = {r for r in selected_data if r.endswith('.pt')}
    retained_pt = {str(Path(r).parent/'observation/rgb_valid.pt') for r in metas}
    if actual_pt != removed | retained_pt:
        raise ValueError(f'Unclassified numerical input: {sorted(actual_pt ^ (removed | retained_pt))[:10]}')
    write_json(state/'plan.json', {'binding': binding, 'views': list(plans),
        'replaced_source_files': sorted(removed), 'new_bundle_files': sorted(expected_new),
        'deferred_unchanged': [r for r in entries if r['action'].startswith('defer_')],
        'source_preserved': True, 'target_is_non_media_stage': True})
    for rel, row in selected_data.items():
        if rel in removed or rel in metas: continue
        destination = target/rel
        if destination.exists(): verify(destination,row)
        else: atomic_copy(source/rel,destination)
    rows = []
    status.update(phase='migrating', verified_input_files=len(selected), total_views=len(metas))
    write_json(state/'status.json', status)
    for v,(old,new,fs) in plans.items():
        old_view, new_view = source/v, target/v
        cp = state/'checkpoints'/f"{old['sequence_id']}__{old['camera_id']}.json"
        if cp.exists():
            checkpoint = json.loads(cp.read_text())
            for ref, sha in checkpoint['outputs'].items():
                if digest(new_view/ref) != sha: raise ValueError(f'Resume output changed: {v}/{ref}')
        else:
            bundles = {}
            for f in fs:
                c = at(old,f['locator'])
                data = torch.load(safe_path(old_view,c['data_ref']),map_location='cpu',weights_only=True)
                valid = torch.load(safe_path(old_view,c['valid_ref']),map_location='cpu',weights_only=True)
                check_field(data,valid,f,c,old['num_frames'],registry)
                bundles.setdefault(f['bundle'],{})[f['entry']] = {'data':data,'valid':valid}
            for ref,bundle in bundles.items():
                path = safe_path(new_view,ref); path.parent.mkdir(parents=True,exist_ok=True)
                tmp = path.with_name(path.name+'.writing'); torch.save(bundle,tmp)
                reread = torch.load(tmp,map_location='cpu',weights_only=True)
                if set(reread) != set(bundle) or any(not exact(bundle[k][kind],reread[k][kind]) for k in bundle for kind in ['data','valid']):
                    raise ValueError('Tensor bytes changed in serialization')
                tmp.replace(path)
            write_json(new_view/'metainfo.json',new)
        result = audit_view(new_view,registry,source=old_view)
        outputs = {ref:digest(new_view/ref) for ref in sorted({f['bundle'] for f in fs}|{'metainfo.json'})}
        write_json(cp,{'view':v,'outputs':outputs,'verification':result})
        rows.append({'view':v,**result})
        status.update(verified_views=len(rows),current_view=v,elapsed_seconds=round(time.monotonic()-started,2))
        write_json(state/'status.json',status)
        if len(rows)%20==0 or scope=='pilot':print(f'verified {len(rows)}/{len(plans)} {v}',flush=True)
    # Independently check full output path set and unchanged bytes, including frame maps.
    expected_data = (set(selected_data)-removed)|expected_new
    actual_data = {p.relative_to(target).as_posix() for p in (target/'data').rglob('*') if p.is_file()}
    if actual_data != expected_data:
        raise ValueError(f'Output file set differs: {sorted(actual_data ^ expected_data)[:10]}')
    for rel,row in selected_data.items():
        verify(source/rel,row)
        if rel not in removed and rel not in metas:verify(target/rel,row)
    counts=Counter();frames=Counter()
    for row in rows:
        counts[row['split']]+=1;frames[row['split']]+=row['frames']
    deferred = [r for r in entries if r['action'].startswith('defer_')]
    report = {'source_revision':REVISION,'scope':scope,'views':len(rows),'sequences':len({r['sequence'] for r in rows}),
              'split_views':dict(counts),'split_frames':dict(frames),'fields':sum(len(r['fields']) for r in rows),
              'old_numeric_files':len(removed),'new_bundles':len(expected_new),'source_files_verified':len(selected),
              'data_files_present':len(actual_data),'exact_data_valid_bytes':True,'unchanged_source_hashes':True,
              'unchanged_other_data_hashes':True,'metadata_only_version_storage_changed':True,
              'deferred_files':len(deferred),'deferred_bytes':sum(r['size'] for r in deferred),
              'complete_release':False,'media_visual_review':'separate_required',
              'elapsed_seconds':round(time.monotonic()-started,2),'per_view':rows}
    write_json(state/'report.json',report)
    output_hashes = [{'path':r,'size':(target/r).stat().st_size,'sha256':digest(target/r)} for r in sorted(actual_data)]
    write_json(state/'output-data-manifest.json',{'binding':binding,'files':output_hashes})
    write_json(binding_path,binding)
    status.update(phase='numerical_stage_verified',complete_release=False,total_fields=report['fields'])
    write_json(state/'status.json',status)
    print(json.dumps({k:v for k,v in report.items() if k!='per_view'},ensure_ascii=False,indent=2),flush=True)
    return target


def main():
    p=argparse.ArgumentParser(description=__doc__)
    p.add_argument('--workspace',type=Path,required=True)
    p.add_argument('--scope',choices=['pilot','all'],default='pilot')
    p.add_argument('--revalidate-tools',action='store_true',help='explicit reviewed tool-only upgrade; recheck all source/checkpoint/output bytes before adopting new hashes')
    a=p.parse_args()
    try: migrate(a.workspace,a.scope,a.revalidate_tools)
    except Exception as exc:
        status_path=a.workspace/'tmp/hot3d-migration-20261014'/a.scope/'status.json'
        if status_path.exists():
            status=json.loads(status_path.read_text());status.update(phase='failed',error=str(exc));write_json(status_path,status)
        raise


if __name__=='__main__':main()

	#!/usr/bin/env python3
	"""Repack pinned HOT3D shared-v1.0 inputs into a non-media v1.1 staging tree.

	No download, source deletion, geometry reconstruction, or HF upload. Unchanged
	media/assets remain an explicit publisher reuse plan until separately assembled.
	"""
	import argparse
	from collections import Counter
	import fcntl
	import hashlib
	import json
	from pathlib import Path
	import shutil
	import time

	import torch

	from hot3d_bundle import at, audit_view, check_field, exact, fields, new_metadata, safe_path

	REVISION = "54c1bc812b903c2f2b933472163998924588681f"
	PILOT = {"P0015_179e1b84", "P0004_1984bf99"}


	def write_json(path, obj):
	path = Path(path); path.parent.mkdir(parents=True, exist_ok=True)
	tmp = path.with_name(path.name + ".writing")
	tmp.write_text(json.dumps(obj, ensure_ascii=False, indent=2) + "\n")
	tmp.replace(path)


	def digest(path, algorithm="sha256"):
	size = path.stat().st_size
	h = hashlib.sha1(f"blob {size}\0".encode()) if algorithm == "git_blob_sha1" else hashlib.sha256()
	with path.open("rb") as stream:
	for chunk in iter(lambda: stream.read(2**20), b""):
	h.update(chunk)
	return h.hexdigest()


	def verify(path, row):
	if not path.is_file() or path.stat().st_size != row['size'] or digest(path, row['algorithm']) != row['digest']:
	raise ValueError(f"Pinned source size/digest mismatch: {path}")


	def atomic_copy(source, target):
	target.parent.mkdir(parents=True, exist_ok=True)
	tmp = target.with_name(target.name + '.copying')
	shutil.copy2(source, tmp)
	tmp.replace(target)


	def migrate(workspace, scope, revalidate_tools=False):
	torch.set_num_threads(1)
	workspace = Path(workspace).resolve()
	source = workspace/'source/HOT3D/shared-v1.0'
	state = workspace/'tmp/hot3d-migration-20261014'/scope
	state.mkdir(parents=True, exist_ok=True)
	target = workspace/('tmp/hot3d-migration-20261014/pilot/shared-v1.1' if scope == 'pilot' else 'staging/HOT3D/shared-v1.1')
	if target.resolve() == source.resolve() or target.resolve().is_relative_to(source.resolve()):
	raise ValueError('Never migrate inside the read-only source cache')
	manifest_path = workspace/'tmp/remaining-preparation-20261014/HOT3D-manifest.json'
	manifest = json.loads(manifest_path.read_text())
	if manifest['revision'] != REVISION or manifest['dataset'] != 'HOT3D':
	raise ValueError('Wrong fixed source')
	registry_path = workspace/'spec/hand-warehouse-1.1/schema.json'
	registry = json.loads(registry_path.read_text())
	binding = {'source_revision': REVISION, 'source_manifest_sha256': digest(manifest_path),
	'schema_sha256': digest(registry_path),
	'tools': {p: digest(Path(__file__).parent/p) for p in ['migrate_hot3d.py','hot3d_bundle.py']}}
	lock = (state.parent/'migration.lock').open('a+')
	fcntl.flock(lock, fcntl.LOCK_EX \| fcntl.LOCK_NB)
	binding_path = state/'binding.json'
	previous_binding = json.loads(binding_path.read_text()) if binding_path.exists() else None
	if previous_binding is not None and previous_binding != binding:
	if not revalidate_tools or any(previous_binding[k] != binding[k] for k in ['source_revision','source_manifest_sha256','schema_sha256']):
	raise ValueError('Resume source/tool binding changed; review before --revalidate-tools')
	# Explicitly reviewed tool-only upgrade: old checkpoints/hashes must
	# still match and all views are compared again. Adopt only on success.
	write_json(state/'binding-before-revalidation.json', previous_binding)
	if not binding_path.exists() and target.exists():
	raise ValueError('Existing unregistered staging tree; do not overwrite')
	if previous_binding is None:
	write_json(binding_path, binding)
	started = time.monotonic()
	entries = [r for r in manifest['files'] if scope == 'all' or not r['relative'].startswith('data/') or r['relative'].split('/')[1] in PILOT]
	selected = [r for r in entries if r['action'].startswith('download_')]
	status = {'phase': 'verifying_inputs', 'scope': scope, 'source_revision': REVISION,
	'complete_release': False, 'verified_views': 0}
	write_json(state/'status.json', status)
	for i, row in enumerate(selected,1):
	verify(source/row['relative'],row)
	if i % 2000 == 0: print(f'input hashes {i}/{len(selected)}',flush=True)
	metas = sorted(r['relative'] for r in entries if r['relative'].startswith('data/') and len(Path(r['relative']).parts)==4 and r['relative'].endswith('/metainfo.json'))
	expected_views = 6 if scope == 'pilot' else 594
	if len(metas) != expected_views:
	raise ValueError(f'Expected {expected_views} complete views, got {len(metas)}')
	expected_new, removed, plans = set(), set(), {}
	for relative in metas:
	old = json.loads((source/relative).read_text())
	new = new_metadata(old)
	v = Path(relative).parent
	fs = fields(old)
	for f in fs:
	c = at(old,f['locator'])
	removed.update((v/c[k]).as_posix() for k in ['data_ref','valid_ref'])
	expected_new.add((v/f['bundle']).as_posix())
	plans[v.as_posix()] = (old,new,fs)
	# Precisely retain RGB validity; asset validity belongs to unchanged assets.
	selected_data = {r['relative']: r for r in selected if r['relative'].startswith('data/')}
	actual_pt = {r for r in selected_data if r.endswith('.pt')}
	retained_pt = {str(Path(r).parent/'observation/rgb_valid.pt') for r in metas}
	if actual_pt != removed \| retained_pt:
	raise ValueError(f'Unclassified numerical input: {sorted(actual_pt ^ (removed \| retained_pt))[:10]}')
	write_json(state/'plan.json', {'binding': binding, 'views': list(plans),
	'replaced_source_files': sorted(removed), 'new_bundle_files': sorted(expected_new),
	'deferred_unchanged': [r for r in entries if r['action'].startswith('defer_')],
	'source_preserved': True, 'target_is_non_media_stage': True})
	for rel, row in selected_data.items():
	if rel in removed or rel in metas: continue
	destination = target/rel
	if destination.exists(): verify(destination,row)
	else: atomic_copy(source/rel,destination)
	rows = []
	status.update(phase='migrating', verified_input_files=len(selected), total_views=len(metas))
	write_json(state/'status.json', status)
	for v,(old,new,fs) in plans.items():
	old_view, new_view = source/v, target/v
	cp = state/'checkpoints'/f"{old['sequence_id']}__{old['camera_id']}.json"
	if cp.exists():
	checkpoint = json.loads(cp.read_text())
	for ref, sha in checkpoint['outputs'].items():
	if digest(new_view/ref) != sha: raise ValueError(f'Resume output changed: {v}/{ref}')
	else:
	bundles = {}
	for f in fs:
	c = at(old,f['locator'])
	data = torch.load(safe_path(old_view,c['data_ref']),map_location='cpu',weights_only=True)
	valid = torch.load(safe_path(old_view,c['valid_ref']),map_location='cpu',weights_only=True)
	check_field(data,valid,f,c,old['num_frames'],registry)
	bundles.setdefault(f['bundle'],{})[f['entry']] = {'data':data,'valid':valid}
	for ref,bundle in bundles.items():
	path = safe_path(new_view,ref); path.parent.mkdir(parents=True,exist_ok=True)
	tmp = path.with_name(path.name+'.writing'); torch.save(bundle,tmp)
	reread = torch.load(tmp,map_location='cpu',weights_only=True)
	if set(reread) != set(bundle) or any(not exact(bundle[k][kind],reread[k][kind]) for k in bundle for kind in ['data','valid']):
	raise ValueError('Tensor bytes changed in serialization')
	tmp.replace(path)
	write_json(new_view/'metainfo.json',new)
	result = audit_view(new_view,registry,source=old_view)
	outputs = {ref:digest(new_view/ref) for ref in sorted({f['bundle'] for f in fs}\|{'metainfo.json'})}
	write_json(cp,{'view':v,'outputs':outputs,'verification':result})
	rows.append({'view':v,**result})
	status.update(verified_views=len(rows),current_view=v,elapsed_seconds=round(time.monotonic()-started,2))
	write_json(state/'status.json',status)
	if len(rows)%20==0 or scope=='pilot':print(f'verified {len(rows)}/{len(plans)} {v}',flush=True)
	# Independently check full output path set and unchanged bytes, including frame maps.
	expected_data = (set(selected_data)-removed)\|expected_new
	actual_data = {p.relative_to(target).as_posix() for p in (target/'data').rglob('*') if p.is_file()}
	if actual_data != expected_data:
	raise ValueError(f'Output file set differs: {sorted(actual_data ^ expected_data)[:10]}')
	for rel,row in selected_data.items():
	verify(source/rel,row)
	if rel not in removed and rel not in metas:verify(target/rel,row)
	counts=Counter();frames=Counter()
	for row in rows:
	counts[row['split']]+=1;frames[row['split']]+=row['frames']
	deferred = [r for r in entries if r['action'].startswith('defer_')]
	report = {'source_revision':REVISION,'scope':scope,'views':len(rows),'sequences':len({r['sequence'] for r in rows}),
	'split_views':dict(counts),'split_frames':dict(frames),'fields':sum(len(r['fields']) for r in rows),
	'old_numeric_files':len(removed),'new_bundles':len(expected_new),'source_files_verified':len(selected),
	'data_files_present':len(actual_data),'exact_data_valid_bytes':True,'unchanged_source_hashes':True,
	'unchanged_other_data_hashes':True,'metadata_only_version_storage_changed':True,
	'deferred_files':len(deferred),'deferred_bytes':sum(r['size'] for r in deferred),
	'complete_release':False,'media_visual_review':'separate_required',
	'elapsed_seconds':round(time.monotonic()-started,2),'per_view':rows}
	write_json(state/'report.json',report)
	output_hashes = [{'path':r,'size':(target/r).stat().st_size,'sha256':digest(target/r)} for r in sorted(actual_data)]
	write_json(state/'output-data-manifest.json',{'binding':binding,'files':output_hashes})
	write_json(binding_path,binding)
	status.update(phase='numerical_stage_verified',complete_release=False,total_fields=report['fields'])
	write_json(state/'status.json',status)
	print(json.dumps({k:v for k,v in report.items() if k!='per_view'},ensure_ascii=False,indent=2),flush=True)
	return target


	def main():
	p=argparse.ArgumentParser(description=__doc__)
	p.add_argument('--workspace',type=Path,required=True)
	p.add_argument('--scope',choices=['pilot','all'],default='pilot')
	p.add_argument('--revalidate-tools',action='store_true',help='explicit reviewed tool-only upgrade; recheck all source/checkpoint/output bytes before adopting new hashes')
	a=p.parse_args()
	try: migrate(a.workspace,a.scope,a.revalidate_tools)
	except Exception as exc:
	status_path=a.workspace/'tmp/hot3d-migration-20261014'/a.scope/'status.json'
	if status_path.exists():
	status=json.loads(status_path.read_text());status.update(phase='failed',error=str(exc));write_json(status_path,status)
	raise


	if __name__=='__main__':main()