| """Serialize frozen benchmark tensors into Core ML. No model execution or weight authoring.""" |
| from pathlib import Path |
| from datetime import datetime, timezone |
| import hashlib, json |
| import numpy as np |
| import coremltools as ct |
| from coremltools.converters.mil import Builder as mb |
| from coremltools.libmilstoragepython import _BlobStorageReader |
|
|
| ROOT = Path(__file__).resolve().parent |
| SOURCE = ROOT.parent / 'creddata-confirmation-2026-11-06/results/model' |
| OUT = ROOT / 'm-package' |
| OUT.mkdir(exist_ok=True) |
| sha = lambda p: hashlib.sha256(p.read_bytes()).hexdigest() |
| freeze = json.loads((SOURCE / 'frozen.json').read_text()) |
| assert sha(SOURCE / 'model.py') == freeze['source_sha256'] |
| assert sha(SOURCE / 'weights.npz') == freeze['weights_sha256'] |
| with np.load(SOURCE / 'weights.npz', allow_pickle=False) as archive: |
| W, b, o = (archive[k] for k in ['W', 'b', 'o']) |
| assert W.shape == (1189, 5153) and b.shape == o.shape == (1189,) |
| assert all(a.dtype == np.float32 for a in [W, b, o]) |
| assert W.size + b.size + o.size + 1 == freeze['parameters'] |
|
|
| @mb.program(input_specs=[mb.TensorSpec(shape=(128, 5153))], opset_version=ct.target.iOS18) |
| def network(features): |
| hidden = mb.relu(x=mb.linear(x=features, weight=W, bias=b)) |
| return mb.linear(x=hidden, weight=o[None, :], bias=np.array([-0.5], np.float32), name='scores') |
|
|
| |
| model = ct.convert(network, minimum_deployment_target=ct.target.iOS18, |
| compute_precision=ct.precision.FLOAT32, skip_model_load=True) |
| package = OUT / 'bundle/BlotHandwritten.mlpackage' |
| package.parent.mkdir(exist_ok=True) |
| model.save(str(package)) |
| spec = ct.utils.load_spec(str(package / 'Data/com.apple.CoreML/model.mlmodel')) |
| block = spec.mlProgram.functions['main'].block_specializations['CoreML8'] |
| reader = _BlobStorageReader(str(package / 'Data/com.apple.CoreML/weights/weight.bin')) |
| expected = {'linear_0_weight_0': W, 'linear_0_bias_0': b, 'scores_weight_0': o[None, :]} |
| verified = [] |
| for op in block.operations: |
| name = op.outputs[0].name |
| if op.type == 'const' and name in expected: |
| val = op.attributes['val'] |
| shape = tuple(d.constant.size for d in val.type.tensorType.dimensions) |
| actual = np.asarray(reader.read_float_data(val.blobFileValue.offset), dtype=np.float32) |
| assert shape == expected[name].shape |
| assert actual.tobytes() == expected[name].tobytes(), name |
| verified.append(name) |
| if op.type == 'const' and name == 'scores_bias_0': |
| assert list(op.attributes['val'].immediateValue.tensor.floats.values) == [-0.5] |
| assert set(verified) == set(expected) |
| assert [op.type for op in block.operations if op.type != 'const'] == ['linear', 'relu', 'linear'] |
| assert list(spec.description.input[0].type.multiArrayType.shape) == [128, 5153] |
| assert list(spec.description.output[0].type.multiArrayType.shape) == [128, 1] |
| assert sha(SOURCE / 'model.py') == freeze['source_sha256'] |
| assert sha(SOURCE / 'weights.npz') == freeze['weights_sha256'] |
| files = {str(p.relative_to(OUT)): sha(p) for p in sorted(OUT.rglob('*')) if p.is_file()} |
| size = sum((OUT / p).stat().st_size for p in files) + (SOURCE / 'model.py').stat().st_size + (SOURCE / 'weights.npz').stat().st_size |
| assert size < 30_000_000 |
| receipt = {'checked_utc': datetime.now(timezone.utc).isoformat(), 'coremltools': ct.__version__, |
| 'source_model_sha256': freeze['source_sha256'], 'source_weights_sha256': freeze['weights_sha256'], |
| 'parameters': freeze['parameters'], 'inference_asset_bytes': size, |
| 'graph': 'linear -> ReLU -> linear', 'features': 5153, 'hidden_units': 1189, |
| 'input_shape': [128, 5153], 'output_shape': [128, 1], |
| 'all_serialized_weight_bytes_match_frozen_npz': True, 'readout_bias': -0.5, |
| 'native_model_loaded': False, 'inference_executed': False, 'new_weights_authored': False, |
| 'native_runtime_parity_tested': False, 'benchmark_rerun': False, |
| 'files': files} |
| (ROOT / 'coreml-packaging.json').write_text(json.dumps(receipt, indent=2) + '\n') |
| print(json.dumps(receipt, indent=2)) |
|
|
