File size: 42,058 Bytes
20d5a17 7f9c062 20d5a17 c5811fd 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 c5811fd 20d5a17 c5811fd 7f9c062 20d5a17 db2d954 20d5a17 db2d954 c5811fd 20d5a17 8338745 db2d954 8338745 db2d954 8338745 20d5a17 8338745 20d5a17 7f9c062 8338745 7f9c062 8338745 7f9c062 8338745 7f9c062 8338745 7f9c062 8338745 7f9c062 8338745 7f9c062 8338745 7f9c062 8338745 7f9c062 20d5a17 c5811fd 20d5a17 c5811fd 20d5a17 c5811fd 20d5a17 c5811fd 20d5a17 c5811fd 20d5a17 c5811fd 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 7f9c062 c5811fd 20d5a17 c5811fd 20d5a17 c5811fd 7f9c062 c5811fd 7f9c062 c5811fd 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 c5811fd 7f9c062 c5811fd 7f9c062 c5811fd 20d5a17 db2d954 20d5a17 c5811fd 20d5a17 c5811fd 20d5a17 db2d954 c5811fd db2d954 20d5a17 3b60855 20d5a17 8338745 3b60855 8338745 3b60855 8338745 3b60855 8338745 db2d954 8338745 db2d954 8338745 db2d954 8338745 db2d954 8338745 20d5a17 db2d954 8338745 db2d954 8338745 db2d954 8338745 db2d954 8338745 20d5a17 c5811fd 20d5a17 7f9c062 20d5a17 7f9c062 20d5a17 c5811fd 7f9c062 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 | """Convert SHARP PyTorch model to Core ML . mlmodel format.
This script converts the SHARP (Sharp Monocular View Synthesis) model
from PyTorch (.pt) to Core ML (.mlmodel) format for deployment on Apple devices.
"""
from __future__ import annotations
import argparse
import logging
from pathlib import Path
from typing import Any
import coremltools as ct
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
try:
import onnx
import onnxruntime as ort
ONNX_AVAILABLE = True
except ImportError:
ONNX_AVAILABLE = False
print("Warning: ONNX not available, will use direct tracing")
# Import SHARP model components
from sharp. models import PredictorParams, create_predictor
from sharp.models.predictor import RGBGaussianPredictor
LOGGER = logging.getLogger(__name__)
DEFAULT_MODEL_URL = "https://ml-site.cdn-apple.com/models/sharp/sharp_2572gikvuh.pt"
class SharpModelWrapper(nn.Module):
"""Wrapper around RGBGaussianPredictor for Core ML export.
This wrapper simplifies the model interface for Core ML conversion by:
1. Removing optional depth input (inference mode only)
2. Flattening the Gaussians3D NamedTuple output to individual tensors
3. Handling internal preprocessing
"""
def __init__(self, predictor: RGBGaussianPredictor):
"""Initialize the wrapper.
Args:
predictor: The SHARP RGBGaussianPredictor model.
"""
super().__init__()
self.predictor = predictor
# Remove depth alignment for inference-only export
# (depth alignment requires ground truth depth which is not available at inference)
self.predictor.depth_alignment. scale_map_estimator = None
def forward(
self,
image: torch.Tensor,
disparity_factor: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
"""Run inference and return flattened Gaussian parameters.
Args:
image: Input image tensor of shape (1, 3, H, W) in range [0, 1].
disparity_factor: Disparity factor tensor of shape (1,).
Returns:
Tuple of tensors representing 3D Gaussians:
- mean_vectors: (1, N, 3) - 3D positions
- singular_values: (1, N, 3) - scales
- quaternions: (1, N, 4) - rotations
- colors: (1, N, 3) - RGB colors
- opacities: (1, N) - opacity values
"""
# Run the predictor (depth=None for inference)
gaussians = self.predictor(image, disparity_factor, depth=None)
# Return as individual tensors (Core ML doesn't support NamedTuple)
return (
gaussians.mean_vectors,
gaussians.singular_values,
gaussians.quaternions,
gaussians.colors,
gaussians.opacities,
)
class SafeClamp(nn.Module):
"""Safe clamp operation that avoids tracing issues."""
def forward(self, x, min_val=1e-4, max_val=1e4):
return torch.clamp(x, min=min_val, max=max_val)
class SafeDivision(nn.Module):
"""Safe division that avoids division by zero."""
def forward(self, numerator, denominator):
return numerator / torch.clamp(denominator, min=1e-8)
class SharpModelTraceable(nn.Module):
"""Fully traceable version of SHARP for Core ML conversion.
This version removes all dynamic control flow and makes the model
fully traceable with torch.jit.trace.
"""
def __init__(self, predictor: RGBGaussianPredictor):
"""Initialize the traceable wrapper.
Args:
predictor: The SHARP RGBGaussianPredictor model.
"""
super().__init__()
# Copy all submodules
self.init_model = predictor.init_model
self. feature_model = predictor.feature_model
self.monodepth_model = predictor.monodepth_model
self.prediction_head = predictor.prediction_head
self.gaussian_composer = predictor.gaussian_composer
self.depth_alignment = predictor.depth_alignment
# Replace problematic operations with custom modules
self.safe_clamp = SafeClamp()
self.safe_div = SafeDivision()
def forward(
self,
image: torch. Tensor,
disparity_factor: torch.Tensor
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
"""Run inference with traceable forward pass.
Args:
image: Input image tensor of shape (1, 3, H, W) in range [0, 1].
disparity_factor: Disparity factor tensor of shape (1,).
Returns:
Tuple of 5 tensors representing 3D Gaussians.
"""
# Estimate depth using monodepth
monodepth_output = self.monodepth_model(image)
monodepth_disparity = monodepth_output. disparity
# Convert disparity to depth with higher precision
# Use tighter clamp bounds and higher precision intermediate computation
disparity_factor_expanded = disparity_factor[: , None, None, None]
# Cast to float64 for more precise division, then back to float32
disparity_clamped = monodepth_disparity.clamp(min=1e-6, max=1e4)
monodepth = disparity_factor_expanded.double() / disparity_clamped.double()
monodepth = monodepth.float()
# Apply depth alignment (inference mode)
monodepth, _ = self.depth_alignment(monodepth, None, monodepth_output.decoder_features)
# Initialize gaussians
init_output = self.init_model(image, monodepth)
# Extract features
image_features = self. feature_model(
init_output.feature_input,
encodings=monodepth_output.output_features
)
# Predict deltas
delta_values = self.prediction_head(image_features)
# Compose final gaussians
gaussians = self.gaussian_composer(
delta=delta_values,
base_values=init_output. gaussian_base_values,
global_scale=init_output.global_scale,
)
# Normalize quaternions for consistent validation and inference
# This is critical for CoreML conversion accuracy
quaternions = gaussians.quaternions
# Use double precision for quaternion normalization to reduce numerical errors
quaternions_fp64 = quaternions.double()
quat_norm_sq = torch.sum(quaternions_fp64 * quaternions_fp64, dim=-1, keepdim=True)
quat_norm = torch.sqrt(torch.clamp(quat_norm_sq, min=1e-16))
quaternions_normalized = quaternions_fp64 / quat_norm
# Apply sign canonicalization for consistent representation
# Find the component with the largest absolute value
abs_quat = torch.abs(quaternions_normalized)
max_idx = torch.argmax(abs_quat, dim=-1, keepdim=True)
# Create one-hot selector for the max component
one_hot = torch.zeros_like(quaternions_normalized)
one_hot.scatter_(-1, max_idx, 1.0)
# Get the sign of the max component
max_component_sign = torch.sum(quaternions_normalized * one_hot, dim=-1, keepdim=True)
# Flip if negative (multiply by sign to canonicalize)
sign = torch.sign(max_component_sign)
sign = torch.where(sign == 0.0, torch.ones_like(sign), sign) # Handle zero case
quaternions = (quaternions_normalized * sign).float()
return (
gaussians.mean_vectors,
gaussians.singular_values,
quaternions,
gaussians.colors,
gaussians. opacities,
)
def load_sharp_model(checkpoint_path: Path | None = None) -> RGBGaussianPredictor:
"""Load SHARP model from checkpoint.
Args:
checkpoint_path: Path to the . pt checkpoint file.
If None, downloads the default model.
Returns:
The loaded RGBGaussianPredictor model in eval mode.
"""
if checkpoint_path is None:
LOGGER.info("Downloading default model from %s", DEFAULT_MODEL_URL)
state_dict = torch.hub.load_state_dict_from_url(DEFAULT_MODEL_URL, progress=True)
else:
LOGGER.info("Loading checkpoint from %s", checkpoint_path)
state_dict = torch.load(checkpoint_path, weights_only=True, map_location="cpu")
# Create model with default parameters
predictor = create_predictor(PredictorParams())
predictor. load_state_dict(state_dict)
predictor. eval()
return predictor
def convert_via_onnx(
predictor: RGBGaussianPredictor,
output_path: Path,
input_shape: tuple[int, int] = (1536, 1536),
compute_precision: ct.precision = ct.precision.FLOAT32,
compute_units: ct.ComputeUnit = ct.ComputeUnit.ALL,
minimum_deployment_target: ct.target | None = None,
) -> ct.models.MLModel:
"""Convert SHARP model via ONNX intermediate format for better tracing.
Args:
predictor: The SHARP RGBGaussianPredictor model.
output_path: Path to save the .mlmodel file.
input_shape: Input image shape (height, width).
compute_precision: Precision for compute (FLOAT16 or FLOAT32).
compute_units: Target compute units.
minimum_deployment_target: Minimum iOS/macOS deployment target.
Returns:
The converted Core ML model.
"""
if not ONNX_AVAILABLE:
raise ImportError("ONNX libraries not available. Install with: pip install onnx onnxruntime")
LOGGER.info("Attempting ONNX-based conversion for better accuracy...")
# Ensure depth alignment is disabled for inference
predictor.depth_alignment.scale_map_estimator = None
# Create traceable wrapper
model_wrapper = SharpModelTraceable(predictor)
model_wrapper.eval()
# Pre-warm the model
LOGGER.info("Pre-warming model...")
with torch.no_grad():
for _ in range(3):
warm_image = torch.randn(1, 3, input_shape[0], input_shape[1])
warm_disparity = torch.tensor([1.0])
_ = model_wrapper(warm_image, warm_disparity)
# Create example inputs
height, width = input_shape
torch.manual_seed(42)
example_image = torch.randn(1, 3, height, width)
example_disparity_factor = torch.tensor([1.0])
# Export to ONNX with external data to handle large models
onnx_path = output_path.with_suffix('.onnx')
LOGGER.info(f"Exporting to ONNX: {onnx_path}")
try:
# Export with external data format to handle large models (>2GB)
torch.onnx.export(
model_wrapper,
(example_image, example_disparity_factor),
str(onnx_path), # Must be string path for external data format
export_params=True,
verbose=False,
input_names=['image', 'disparity_factor'],
output_names=[
'mean_vectors_3d_positions',
'singular_values_scales',
'quaternions_rotations',
'colors_rgb_linear',
'opacities_alpha_channel'
],
dynamic_axes={
'mean_vectors_3d_positions': {1: 'num_gaussians'},
'singular_values_scales': {1: 'num_gaussians'},
'quaternions_rotations': {1: 'num_gaussians'},
'colors_rgb_linear': {1: 'num_gaussians'},
'opacities_alpha_channel': {1: 'num_gaussians'}
},
opset_version=17, # Use latest stable opset
)
# For models >2GB, save with external data format
try:
import onnx
model_proto = onnx.load(str(onnx_path))
# Check if model is larger than 2GB (need external data)
model_size = model_proto.ByteSize()
if model_size > 2e9: # 2GB
LOGGER.info(f"Model size {model_size/1e9:.2f}GB > 2GB, converting to external data format...")
onnx.save_model(
model_proto,
str(onnx_path),
save_as_external_data=True,
all_tensors_to_one_file=True,
location=f"{onnx_path.stem}.onnx.data",
size_threshold=1024,
convert_attribute=False,
)
LOGGER.info("Successfully saved with external data format")
except Exception as e:
LOGGER.warning(f"Could not check/convert to external data format: {e}")
LOGGER.info("ONNX export successful")
except Exception as e:
LOGGER.error(f"ONNX export failed: {e}")
raise
# Verify ONNX model (skip for large models >2GB)
try:
# For large models, check without loading external data
onnx.checker.check_model(str(onnx_path))
LOGGER.info("ONNX model validation passed")
except Exception as e:
LOGGER.warning(f"ONNX model validation skipped for large model: {e}")
# Convert ONNX to Core ML
LOGGER.info("Converting ONNX to Core ML...")
# Define input types for Core ML
inputs = [
ct.TensorType(
name="image",
shape=(1, 3, height, width),
dtype=np.float32,
),
ct.TensorType(
name="disparity_factor",
shape=(1,),
dtype=np.float32,
),
]
# Convert from ONNX - must specify source="pytorch" for proper conversion
mlmodel = ct.convert(
str(onnx_path),
source="pytorch", # Specify PyTorch as source framework
inputs=inputs,
convert_to="mlprogram",
compute_precision=compute_precision,
compute_units=compute_units,
minimum_deployment_target=minimum_deployment_target,
)
# Add metadata
mlmodel.author = "Apple Inc. (via ONNX)"
mlmodel.license = "See LICENSE_MODEL in ml-sharp repository"
mlmodel.short_description = (
"SHARP: Sharp Monocular View Synthesis - Predicts 3D Gaussian splats from a single image (ONNX conversion)"
)
mlmodel.version = "1.0.0"
# Add output descriptions
spec = mlmodel.get_spec()
output_descriptions = {
"mean_vectors_3d_positions": (
"3D positions of Gaussian splats in normalized device coordinates (NDC). "
"Shape: (1, N, 3), where N is the number of Gaussians."
),
"singular_values_scales": (
"Scale factors for each Gaussian along its principal axes. "
"Represents size and anisotropy. Shape: (1, N, 3)."
),
"quaternions_rotations": (
"Rotation of each Gaussian as a unit quaternion [w, x, y, z]. "
"Used to orient the ellipsoid. Shape: (1, N, 4)."
),
"colors_rgb_linear": (
"RGB color values in linear RGB space (not gamma-corrected). "
"Shape: (1, N, 3), with range [0, 1]."
),
"opacities_alpha_channel": (
"Opacity value per Gaussian (alpha channel), used for blending. "
"Shape: (1, N), where values are in [0, 1]."
),
}
# Update output names and descriptions
for i, name in enumerate([
'mean_vectors_3d_positions',
'singular_values_scales',
'quaternions_rotations',
'colors_rgb_linear',
'opacities_alpha_channel'
]):
if i < len(spec.description.output):
output = spec.description.output[i]
output.name = name
output.shortDescription = output_descriptions[name]
# Validate output names are set correctly
LOGGER.info("Output names after update: %s", [o.name for o in spec.description.output])
# Save the model
LOGGER.info(f"Saving Core ML model to {output_path}")
mlmodel.save(str(output_path))
# Clean up ONNX file
try:
onnx_path.unlink()
LOGGER.info("Cleaned up temporary ONNX file")
except Exception as e:
LOGGER.warning(f"Could not clean up ONNX file: {e}")
return mlmodel
def convert_to_coreml(
predictor: RGBGaussianPredictor,
output_path: Path,
input_shape: tuple[int, int] = (1536, 1536),
compute_precision: ct.precision = ct.precision.FLOAT16,
compute_units: ct.ComputeUnit = ct.ComputeUnit.ALL,
minimum_deployment_target: ct.target | None = None,
) -> ct.models.MLModel:
"""Convert SHARP model to Core ML format.
Args:
predictor: The SHARP RGBGaussianPredictor model.
output_path: Path to save the . mlmodel file.
input_shape: Input image shape (height, width). Default is (1536, 1536).
compute_precision: Precision for compute (FLOAT16 or FLOAT32).
compute_units: Target compute units (ALL, CPU_AND_GPU, CPU_ONLY, etc.).
minimum_deployment_target: Minimum iOS/macOS deployment target.
Returns:
The converted Core ML model.
"""
LOGGER.info("Preparing model for Core ML conversion...")
# Ensure depth alignment is disabled for inference
predictor.depth_alignment.scale_map_estimator = None
# Create traceable wrapper
model_wrapper = SharpModelTraceable(predictor)
model_wrapper.eval()
# Pre-warm the model with a few forward passes for better tracing
LOGGER.info("Pre-warming model for better tracing...")
with torch.no_grad():
for _ in range(3):
warm_image = torch.randn(1, 3, input_shape[0], input_shape[1])
warm_disparity = torch.tensor([1.0])
_ = model_wrapper(warm_image, warm_disparity)
# Create deterministic example inputs for tracing (same as validation)
height, width = input_shape
torch.manual_seed(42) # Use same seed as validation for consistency
example_image = torch.randn(1, 3, height, width)
example_disparity_factor = torch.tensor([1.0])
LOGGER.info("Attempting torch.jit.script for better tracing...")
try:
with torch.no_grad():
scripted_model = torch.jit.script(model_wrapper)
LOGGER.info("torch.jit.script succeeded, using scripted model")
traced_model = scripted_model
except Exception as e:
LOGGER.warning(f"torch.jit.script failed: {e}")
LOGGER.info("Falling back to torch.jit.trace...")
with torch.no_grad():
traced_model = torch.jit.trace(
model_wrapper,
(example_image, example_disparity_factor),
strict=False, # Allow some flexibility for complex models
check_trace=False, # Skip trace checking to allow more flexibility
)
LOGGER.info("Converting traced model to Core ML...")
# Define input types for Core ML
inputs = [
ct.TensorType(
name="image",
shape=(1, 3, height, width),
dtype=np.float32,
),
ct.TensorType(
name="disparity_factor",
shape=(1,),
dtype=np.float32,
),
]
# Define output names with clear, descriptive labels
output_names = [
"mean_vectors_3d_positions", # 3D positions (NDC space)
"singular_values_scales", # Scale parameters (diagonal of covariance)
"quaternions_rotations", # Rotation as quaternions
"colors_rgb_linear", # RGB colors in linear color space
"opacities_alpha_channel", # Opacity values (alpha)
]
# Define outputs with proper names for Core ML conversion
outputs = [
ct.TensorType(name=output_names[0], dtype=np.float32),
ct.TensorType(name=output_names[1], dtype=np.float32),
ct.TensorType(name=output_names[2], dtype=np.float32),
ct.TensorType(name=output_names[3], dtype=np.float32),
ct.TensorType(name=output_names[4], dtype=np.float32),
]
# Set up conversion config
conversion_kwargs: dict[str, Any] = {
"inputs": inputs,
"outputs": outputs, # Specify output names during conversion
"convert_to": "mlprogram", # Use ML Program format for better performance
"compute_precision": compute_precision,
"compute_units": compute_units,
}
if minimum_deployment_target is not None:
conversion_kwargs["minimum_deployment_target"] = minimum_deployment_target
# Convert to Core ML
mlmodel = ct.convert(
traced_model,
**conversion_kwargs,
)
# Add metadata
mlmodel.author = "Apple Inc."
mlmodel.license = "See LICENSE_MODEL in ml-sharp repository"
mlmodel.short_description = (
"SHARP: Sharp Monocular View Synthesis - Predicts 3D Gaussian splats from a single image"
)
mlmodel.version = "1.0.0"
# Update output names and descriptions via spec BEFORE saving
spec = mlmodel.get_spec()
# Input descriptions
input_descriptions = {
"image": "RGB image normalized to [0, 1], shape (1, 3, H, W)",
"disparity_factor": "Focal length / image width ratio, shape (1,)",
}
# Output descriptions with clear intent and units
output_descriptions = {
"mean_vectors_3d_positions": (
"3D positions of Gaussian splats in normalized device coordinates (NDC). "
"Shape: (1, N, 3), where N is the number of Gaussians."
),
"singular_values_scales": (
"Scale factors for each Gaussian along its principal axes. "
"Represents size and anisotropy. Shape: (1, N, 3)."
),
"quaternions_rotations": (
"Rotation of each Gaussian as a unit quaternion [w, x, y, z]. "
"Used to orient the ellipsoid. Shape: (1, N, 4)."
),
"colors_rgb_linear": (
"RGB color values in linear RGB space (not gamma-corrected). "
"Shape: (1, N, 3), with range [0, 1]."
),
"opacities_alpha_channel": (
"Opacity value per Gaussian (alpha channel), used for blending. "
"Shape: (1, N), where values are in [0, 1]."
),
}
# Update output names and descriptions
for i, name in enumerate(output_names):
if i < len(spec.description.output):
output = spec.description.output[i]
output.name = name # Update name
output.shortDescription = output_descriptions[name] # Add description
# Validate output names are set correctly
LOGGER.info("Output names after update: %s", [o.name for o in spec.description.output])
# Save the model with correct names
LOGGER.info("Saving Core ML model to %s", output_path)
mlmodel.save(str(output_path))
return mlmodel
def convert_to_coreml_with_preprocessing(
predictor: RGBGaussianPredictor,
output_path: Path,
input_shape: tuple[int, int] = (1536, 1536),
) -> ct.models. MLModel:
"""Convert SHARP model to Core ML with built-in image preprocessing.
This version includes image normalization as part of the model,
accepting uint8 images as input.
Args:
predictor: The SHARP RGBGaussianPredictor model.
output_path: Path to save the .mlmodel file.
input_shape: Input image shape (height, width).
Returns:
The converted Core ML model.
"""
class SharpWithPreprocessing(nn.Module):
"""SHARP model with integrated preprocessing."""
def __init__(self, base_model: SharpModelTraceable):
super().__init__()
self.base_model = base_model
def forward(
self,
image: torch. Tensor,
disparity_factor: torch. Tensor
) -> tuple[torch. Tensor, torch. Tensor, torch. Tensor, torch. Tensor, torch. Tensor]:
# Normalize image from [0, 255] to [0, 1]
image_normalized = image / 255.0
return self.base_model(image_normalized, disparity_factor)
model_wrapper = SharpWithPreprocessing(SharpModelTraceable(predictor))
model_wrapper.eval()
height, width = input_shape
example_image = torch.randint(0, 256, (1, 3, height, width), dtype=torch.float32)
example_disparity_factor = torch.tensor([1.0])
LOGGER.info("Tracing model with preprocessing...")
with torch.no_grad():
traced_model = torch.jit.trace(
model_wrapper,
(example_image, example_disparity_factor),
strict=False,
)
inputs = [
ct.ImageType(
name="image",
shape=(1, 3, height, width),
scale=1.0, # Will be normalized in the model
color_layout=ct.colorlayout.RGB,
),
ct.TensorType(
name="disparity_factor",
shape=(1,),
dtype=np. float32,
),
]
# Define output names with clear, descriptive labels
output_names = [
"mean_vectors_3d_positions", # 3D positions (NDC space)
"singular_values_scales", # Scale parameters (diagonal of covariance)
"quaternions_rotations", # Rotation as quaternions
"colors_rgb_linear", # RGB colors in linear color space
"opacities_alpha_channel", # Opacity values (alpha)
]
# Define outputs with proper names for Core ML conversion
outputs = [
ct.TensorType(name=output_names[0], dtype=np.float32),
ct.TensorType(name=output_names[1], dtype=np.float32),
ct.TensorType(name=output_names[2], dtype=np.float32),
ct.TensorType(name=output_names[3], dtype=np.float32),
ct.TensorType(name=output_names[4], dtype=np.float32),
]
mlmodel = ct.convert(
traced_model,
inputs=inputs,
outputs=outputs, # Specify output names during conversion
convert_to="mlprogram",
compute_precision=ct.precision.FLOAT16,
)
mlmodel.author = "Apple Inc."
mlmodel.short_description = "SHARP model with integrated image preprocessing"
mlmodel.version = "1.0.0"
# Define output names with clear, descriptive labels
output_names = [
"mean_vectors_3d_positions", # 3D positions (NDC space)
"singular_values_scales", # Scale parameters (diagonal of covariance)
"quaternions_rotations", # Rotation as quaternions
"colors_rgb_linear", # RGB colors in linear color space
"opacities_alpha_channel", # Opacity values (alpha)
]
# Output descriptions with clear intent and units
output_descriptions = {
"mean_vectors_3d_positions": (
"3D positions of Gaussian splats in normalized device coordinates (NDC). "
"Shape: (1, N, 3), where N is the number of Gaussians."
),
"singular_values_scales": (
"Scale factors for each Gaussian along its principal axes. "
"Represents size and anisotropy. Shape: (1, N, 3)."
),
"quaternions_rotations": (
"Rotation of each Gaussian as a unit quaternion [w, x, y, z]. "
"Used to orient the ellipsoid. Shape: (1, N, 4)."
),
"colors_rgb_linear": (
"RGB color values in linear RGB space (not gamma-corrected). "
"Shape: (1, N, 3), with range [0, 1]."
),
"opacities_alpha_channel": (
"Opacity value per Gaussian (alpha channel), used for blending. "
"Shape: (1, N), where values are in [0, 1]."
),
}
# Update output names and descriptions via spec BEFORE saving
spec = mlmodel.get_spec()
for i, name in enumerate(output_names):
if i < len(spec.description.output):
output = spec.description.output[i]
output.name = name # Update name
output.shortDescription = output_descriptions[name] # Add description
# Validate output names are set correctly
LOGGER.info("Output names after update: %s", [o.name for o in spec.description.output])
# Save the model with correct names
mlmodel.save(str(output_path))
return mlmodel
def validate_coreml_model(
mlmodel: ct.models.MLModel,
pytorch_model: RGBGaussianPredictor,
input_shape: tuple[int, int] = (1536, 1536),
tolerance: float = 0.01, # Much tighter tolerance
) -> bool:
"""Validate Core ML model outputs against PyTorch model.
Args:
mlmodel: The Core ML model to validate.
pytorch_model: The original PyTorch model.
input_shape: Input image shape (height, width).
tolerance: Maximum allowed difference between outputs.
Returns:
True if validation passes, False otherwise.
"""
LOGGER.info("Validating Core ML model against PyTorch...")
height, width = input_shape
# Create test input
np. random.seed(42)
test_image_np = np.random.rand(1, 3, height, width).astype(np.float32)
test_disparity = np.array([1.0], dtype=np.float32)
# Run PyTorch model
test_image_pt = torch.from_numpy(test_image_np)
test_disparity_pt = torch.from_numpy(test_disparity)
traceable_wrapper = SharpModelTraceable(pytorch_model)
traceable_wrapper.eval()
with torch.no_grad():
pt_outputs = traceable_wrapper(test_image_pt, test_disparity_pt)
# Run Core ML model
coreml_inputs = {
"image": test_image_np,
"disparity_factor": test_disparity,
}
coreml_outputs = mlmodel.predict(coreml_inputs)
# Debug: Print shapes and keys
LOGGER.info(f"PyTorch outputs shapes: {[o.shape for o in pt_outputs]}")
LOGGER.info(f"Core ML outputs keys: {list(coreml_outputs.keys())}")
# Compare outputs with per-output tolerances
output_names = ["mean_vectors_3d_positions", "singular_values_scales", "quaternions_rotations", "colors_rgb_linear", "opacities_alpha_channel"]
# Define tighter tolerances per output type
tolerances = {
"mean_vectors_3d_positions": 0.001, # 1mm precision
"singular_values_scales": 0.0001, # 0.01% scale precision
"quaternions_rotations": 2.0, # Raw component diff (use angular for actual check)
"colors_rgb_linear": 0.002, # ~0.5/255 in 8-bit
"opacities_alpha_channel": 0.005, # ~1/255 in 8-bit
}
# Angular tolerances for quaternions (in degrees)
angular_tolerances = {
"mean": 0.01, # Mean angular error should be < 0.01°
"p99": 0.5, # 99th percentile should be < 0.5°
"max": 10.0, # Max angular error should be < 10° (allows a few outliers)
}
all_passed = True
# Additional diagnostics for depth/position analysis
LOGGER.info("=== Depth/Position Statistics ===")
pt_positions = pt_outputs[0].numpy()
coreml_positions = coreml_outputs[list(coreml_outputs.keys())[0] if "mean_vectors" not in list(coreml_outputs.keys())[0] else [k for k in coreml_outputs.keys() if "mean_vectors" in k][0]]
LOGGER.info(f"PyTorch positions - X range: [{pt_positions[..., 0].min():.4f}, {pt_positions[..., 0].max():.4f}], mean: {pt_positions[..., 0].mean():.4f}")
LOGGER.info(f"PyTorch positions - Y range: [{pt_positions[..., 1].min():.4f}, {pt_positions[..., 1].max():.4f}], mean: {pt_positions[..., 1].mean():.4f}")
LOGGER.info(f"PyTorch positions - Z range: [{pt_positions[..., 2].min():.4f}, {pt_positions[..., 2].max():.4f}], mean: {pt_positions[..., 2].mean():.4f}, std: {pt_positions[..., 2].std():.4f}")
LOGGER.info(f"CoreML positions - X range: [{coreml_positions[..., 0].min():.4f}, {coreml_positions[..., 0].max():.4f}], mean: {coreml_positions[..., 0].mean():.4f}")
LOGGER.info(f"CoreML positions - Y range: [{coreml_positions[..., 1].min():.4f}, {coreml_positions[..., 1].max():.4f}], mean: {coreml_positions[..., 1].mean():.4f}")
LOGGER.info(f"CoreML positions - Z range: [{coreml_positions[..., 2].min():.4f}, {coreml_positions[..., 2].max():.4f}], mean: {coreml_positions[..., 2].mean():.4f}, std: {coreml_positions[..., 2].std():.4f}")
z_diff = np.abs(pt_positions[..., 2] - coreml_positions[..., 2])
LOGGER.info(f"Z-coordinate difference - max: {z_diff.max():.6f}, mean: {z_diff.mean():.6f}, std: {z_diff.std():.6f}")
LOGGER.info("=================================")
for i, name in enumerate(output_names):
pt_output = pt_outputs[i]. numpy()
# Find matching Core ML output (names may have suffixes)
coreml_key = None
for key in coreml_outputs:
if name in key. lower() or f"var_{i}" in key:
coreml_key = key
break
if coreml_key is None:
# Try by index
coreml_key = list(coreml_outputs.keys())[i]
coreml_output = coreml_outputs[coreml_key]
# Special handling for quaternions - account for sign ambiguity
if name == "quaternions_rotations":
# Normalize both quaternion outputs to ensure they're unit quaternions
pt_quat_norm = np.linalg.norm(pt_output, axis=-1, keepdims=True)
pt_output_normalized = pt_output / np.clip(pt_quat_norm, 1e-12, None)
coreml_quat_norm = np.linalg.norm(coreml_output, axis=-1, keepdims=True)
coreml_output_normalized = coreml_output / np.clip(coreml_quat_norm, 1e-12, None)
# Canonicalize sign: handle edge cases where w ≈ 0
# When w is near zero, use the largest magnitude component for canonicalization
def canonicalize_quaternion(q):
"""Canonicalize quaternion to ensure unique representation."""
# Find the component with the largest absolute value
abs_q = np.abs(q)
max_component_idx = np.argmax(abs_q, axis=-1, keepdims=True)
# Create a selector for the max component
selector = np.zeros_like(q)
np.put_along_axis(selector, max_component_idx, 1, axis=-1)
# Get the sign of the max component
max_component_sign = np.sum(q * selector, axis=-1, keepdims=True)
# Flip quaternion if the max component is negative
return np.where(max_component_sign < 0, -q, q)
pt_output_canonical = canonicalize_quaternion(pt_output_normalized)
coreml_output_canonical = canonicalize_quaternion(coreml_output_normalized)
# Now compute differences with canonicalized quaternions
diff = np.abs(pt_output_canonical - coreml_output_canonical)
max_diff = np.max(diff)
mean_diff = np.mean(diff)
# Also compute angular difference for better insight
# Angular distance = 2 * arccos(|dot(q1, q2)|)
dot_products = np.sum(pt_output_canonical * coreml_output_canonical, axis=-1)
dot_products = np.clip(np.abs(dot_products), 0.0, 1.0)
angular_diff_rad = 2 * np.arccos(dot_products)
angular_diff_deg = np.degrees(angular_diff_rad)
max_angular = np.max(angular_diff_deg)
mean_angular = np.mean(angular_diff_deg)
p99_angular = np.percentile(angular_diff_deg, 99)
# Check angular tolerances (more meaningful for rotations)
quat_passed = True
failure_reasons = []
if mean_angular > angular_tolerances["mean"]:
quat_passed = False
failure_reasons.append(f"mean angular {mean_angular:.4f}° > {angular_tolerances['mean']:.4f}°")
if p99_angular > angular_tolerances["p99"]:
quat_passed = False
failure_reasons.append(f"p99 angular {p99_angular:.4f}° > {angular_tolerances['p99']:.4f}°")
if max_angular > angular_tolerances["max"]:
quat_passed = False
failure_reasons.append(f"max angular {max_angular:.4f}° > {angular_tolerances['max']:.4f}°")
if not quat_passed:
LOGGER.warning(
"Output %s: max diff %.6f, mean diff %.6f, p99 diff %.6f, max angular diff %.4f°, mean angular diff %.4f°, p99 angular %.4f° (FAILED: %s)",
name, max_diff, mean_diff, np.percentile(diff, 99), max_angular, mean_angular, p99_angular, "; ".join(failure_reasons)
)
all_passed = False
else:
LOGGER.info(
"Output %s: max diff %.6f, mean diff %.6f, p99 diff %.6f, max angular diff %.4f°, mean angular diff %.4f°, p99 angular %.4f° (PASSED)",
name, max_diff, mean_diff, np.percentile(diff, 99), max_angular, mean_angular, p99_angular
)
else:
diff = np.abs(pt_output - coreml_output)
max_diff = np.max(diff)
mean_diff = np.mean(diff)
p99_diff = np.percentile(diff, 99)
# Compute relative error for better insight
relative_diff = diff / (np.abs(pt_output) + 1e-8)
max_relative = np.max(relative_diff)
mean_relative = np.mean(relative_diff)
output_tolerance = tolerances.get(name, tolerance)
if max_diff > output_tolerance:
LOGGER.warning(
"Output %s: max diff %.6f, mean diff %.6f, p99 diff %.6f, max rel %.4f%%, mean rel %.4f%% (FAILED, tolerance=%.6f)",
name, max_diff, mean_diff, p99_diff, max_relative * 100, mean_relative * 100, output_tolerance
)
all_passed = False
else:
LOGGER.info(
"Output %s: max diff %.6f, mean diff %.6f, p99 diff %.6f, max rel %.4f%%, mean rel %.4f%% (PASSED)",
name, max_diff, mean_diff, p99_diff, max_relative * 100, mean_relative * 100
)
return all_passed
def main():
"""Main conversion script."""
parser = argparse.ArgumentParser(
description="Convert SHARP PyTorch model to Core ML format"
)
parser.add_argument(
"-c", "--checkpoint",
type=Path,
default=None,
help="Path to PyTorch checkpoint. Downloads default if not provided.",
)
parser.add_argument(
"-o", "--output",
type=Path,
default=Path("sharp.mlpackage"),
help="Output path for Core ML model (default: sharp.mlpackage)",
)
parser.add_argument(
"--height",
type=int,
default=1536,
help="Input image height (default: 1536)",
)
parser.add_argument(
"--width",
type=int,
default=1536,
help="Input image width (default: 1536)",
)
parser.add_argument(
"--precision",
choices=["float16", "float32"],
default="float32",
help="Compute precision (default: float32)",
)
parser.add_argument(
"--validate",
action="store_true",
help="Validate Core ML model against PyTorch",
)
parser.add_argument(
"--with-preprocessing",
action="store_true",
help="Include image preprocessing (uint8 -> float normalization)",
)
parser.add_argument(
"-v", "--verbose",
action="store_true",
help="Enable verbose logging",
)
args = parser.parse_args()
# Configure logging
logging.basicConfig(
level=logging. DEBUG if args.verbose else logging.INFO,
format="%(asctime)s - %(name)s - %(levelname)s - %(message)s",
)
# Load PyTorch model
LOGGER.info("Loading SHARP model...")
predictor = load_sharp_model(args. checkpoint)
# Convert to Core ML
input_shape = (args.height, args. width)
precision = ct.precision.FLOAT16 if args. precision == "float16" else ct.precision.FLOAT32
if args.with_preprocessing:
LOGGER.info("Converting with integrated preprocessing...")
mlmodel = convert_to_coreml_with_preprocessing(
predictor,
args.output,
input_shape=input_shape,
)
else:
# Try ONNX conversion first for better accuracy
if ONNX_AVAILABLE:
try:
LOGGER.info("Attempting ONNX-based conversion...")
mlmodel = convert_via_onnx(
predictor,
args.output,
input_shape=input_shape,
compute_precision=precision,
)
LOGGER.info("ONNX conversion successful!")
except Exception as e:
LOGGER.warning(f"ONNX conversion failed: {e}")
LOGGER.info("Falling back to direct tracing...")
mlmodel = convert_to_coreml(
predictor,
args.output,
input_shape=input_shape,
compute_precision=precision,
)
else:
LOGGER.info("ONNX not available, using direct tracing...")
mlmodel = convert_to_coreml(
predictor,
args.output,
input_shape=input_shape,
compute_precision=precision,
)
LOGGER.info("Core ML model saved to %s", args.output)
# Validate if requested
if args.validate:
if validate_coreml_model(mlmodel, predictor, input_shape):
LOGGER.info("✓ Validation passed!")
else:
LOGGER.error("✗ Validation failed!")
return 1
LOGGER.info("Conversion complete!")
return 0
if __name__ == "__main__":
exit(main())
|