Reorganize layout: put ONNX variants under onnx/ and metadata under metadata/

.gitattributes

@@ -1,36 +1,3 @@
-*.7z filter=lfs diff=lfs merge=lfs -text
-*.arrow filter=lfs diff=lfs merge=lfs -text
-*.bin filter=lfs diff=lfs merge=lfs -text
-*.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
-*.ftz filter=lfs diff=lfs merge=lfs -text
-*.gz filter=lfs diff=lfs merge=lfs -text
-*.h5 filter=lfs diff=lfs merge=lfs -text
-*.joblib filter=lfs diff=lfs merge=lfs -text
-*.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
-*.model filter=lfs diff=lfs merge=lfs -text
-*.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
 *.onnx filter=lfs diff=lfs merge=lfs -text
-*.ot filter=lfs diff=lfs merge=lfs -text
-*.parquet filter=lfs diff=lfs merge=lfs -text
-*.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
+*.onnx.data filter=lfs diff=lfs merge=lfs -text
 *.pt filter=lfs diff=lfs merge=lfs -text
-*.pth filter=lfs diff=lfs merge=lfs -text
-*.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
-saved_model/**/* filter=lfs diff=lfs merge=lfs -text
-*.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
-*.tflite filter=lfs diff=lfs merge=lfs -text
-*.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
-*.xz filter=lfs diff=lfs merge=lfs -text
-*.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
-*tfevents* filter=lfs diff=lfs merge=lfs -text
-sharp_ndc_opset18.onnx.data filter=lfs diff=lfs merge=lfs -text

README.md

@@ -2,73 +2,37 @@
 library_name: onnxruntime
 tags:
 - onnx
-- apple
 - sharp
 - view-synthesis
 - 3d-gaussian-splatting
 base_model: apple/Sharp
 ---
 
-# SHARP (NDC) — ONNX FP32 (opset 18)
+# SHARP (NDC) — ONNX exports (FP32 + INT8)
 
-This repository contains an **ONNX FP32** export of Apple's **SHARP** model ("Single-image view synthesis via 3D Gaussians"), exported from the official checkpoint **sharp_2572gikvuh.pt**.
+This repo contains ONNX exports of Apple’s SHARP model **predictor** that outputs **NDC-space 3D Gaussians**.
 
-## What this ONNX contains (important)
-This ONNX exports **only the core predictor network**:
+## What’s included
+- **FP32 (opset18, external data)**: `onnx/fp32/sharp_ndc_opset18.onnx` (+ `.onnx.data`)
+- **INT8 Dynamic (QDQ)**: `onnx/int8_dynamic/sharp_ndc_opset18_int8_dynamic.onnx` (+ optional `.onnx.data`)
 
-`gaussians_ndc = predictor(image_resized_pt, disparity_factor)`
+> **External data note:** if a model has a `.onnx.data` file, it **must** stay in the same folder as its `.onnx`.
 
-It outputs **3D Gaussians in NDC space**.  
-It does **NOT** include the NDC→metric unprojection step because Apple's unprojection path uses SVD + CPU fp64 conversions, which is not a good fit for ONNX portability.
+## I/O contract (all variants)
+Inputs:
+- `image_resized_pt`: float32 `[1, 3, 1536, 1536]` (NCHW, RGB, normalized to [0,1], resized bilinear with align_corners=True)
+- `disparity_factor`: float32 `[1]` computed as `f_px / width_original`
 
-If you want metric-space Gaussians and/or `.ply` export, do that step outside ONNX (e.g., reuse Apple’s `sharp.utils.gaussians.unproject_gaussians` in Python).
-
-## Files
-- `sharp_ndc_opset18.onnx` (graph)
-- `sharp_ndc_opset18.onnx.data` (external weights — required!)
-- `export_config.json`, `run_manifest.json`, `parity_metrics.json` for reproducibility
-
-## Input contract
-**Inputs**
-- `image_resized_pt`: float32 tensor of shape **[1, 3, 1536, 1536]**
-  - RGB
-  - normalized to **[0, 1]** (divide by 255)
-  - resized with **bilinear** and **align_corners=True**
-  - layout **NCHW**
-- `disparity_factor`: float32 tensor of shape **[1]**
-  - computed as: `disparity_factor = f_px / width_original`
-
-Where:
-- `width_original` is the input image width before resizing
-- `f_px` is focal length in pixels (Apple code defaults if EXIF is missing)
-
-**Outputs**
-- `mean_vectors`: [1, 1179648, 3]
-- `singular_values`: [1, 1179648, 3]
-- `quaternions`: [1, 1179648, 4]
-- `colors`: [1, 1179648, 3]
-- `opacities`: [1, 1179648]
-
-All outputs are float32.
-
-## Minimal ONNX Runtime example
-
-```python
-import numpy as np
-import onnxruntime as ort
-
-sess = ort.InferenceSession(
-    "sharp_ndc_opset18.onnx",
-    providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
-)
-
-# Provide:
-# - image_resized_pt as np.float32 [1,3,1536,1536]
-# - disparity_factor as np.float32 [1]
-outputs = sess.run(
-    ["mean_vectors","singular_values","quaternions","colors","opacities"],
-    {"image_resized_pt": image_resized_pt, "disparity_factor": disparity_factor},
-)
-```
+Outputs (float32):
+- `mean_vectors`: `[1, 1179648, 3]`
+- `singular_values`: `[1, 1179648, 3]`
+- `quaternions`: `[1, 1179648, 4]`
+- `colors`: `[1, 1179648, 3]`
+- `opacities`: `[1, 1179648]`
 
+## Notes
+- These ONNX models output **NDC Gaussians** only. Converting to metric space / writing `.ply` can be done outside ONNX.
+- See `metadata/` for manifests and parity metrics.
 
+## License
+This repository contains model derivatives (ONNX conversions). See `LICENSE`.

examples/onnx_to_ply_cpu.py

@@ -0,0 +1 @@
+# TODO: add your CPU-only ONNX->PLY script here

examples/preview_ply_plotly.py

@@ -0,0 +1 @@
+# TODO: add your Plotly PLY preview script here

metadata/parity_metrics.json

@@ -0,0 +1,72 @@
+{
+  "providers": [
+    "CUDAExecutionProvider",
+    "CPUExecutionProvider"
+  ],
+  "outputs": {
+    "mean_vectors": {
+      "shape": [
+        1,
+        1179648,
+        3
+      ],
+      "dtype": "float32",
+      "max_abs": 0.13911914825439453,
+      "mean_abs": 0.00018130234093405306,
+      "ref_max_abs": 6.372216701507568,
+      "nan_out": false,
+      "inf_out": false
+    },
+    "singular_values": {
+      "shape": [
+        1,
+        1179648,
+        3
+      ],
+      "dtype": "float32",
+      "max_abs": 0.035487107932567596,
+      "mean_abs": 4.6182478399714455e-05,
+      "ref_max_abs": 0.07875056564807892,
+      "nan_out": false,
+      "inf_out": false
+    },
+    "quaternions": {
+      "shape": [
+        1,
+        1179648,
+        4
+      ],
+      "dtype": "float32",
+      "max_abs": 13.2420015335083,
+      "mean_abs": 0.008916228078305721,
+      "ref_max_abs": 20.611085891723633,
+      "nan_out": false,
+      "inf_out": false
+    },
+    "colors": {
+      "shape": [
+        1,
+        1179648,
+        3
+      ],
+      "dtype": "float32",
+      "max_abs": 0.06479707360267639,
+      "mean_abs": 0.0002580047002993524,
+      "ref_max_abs": 0.9927035570144653,
+      "nan_out": false,
+      "inf_out": false
+    },
+    "opacities": {
+      "shape": [
+        1,
+        1179648
+      ],
+      "dtype": "float32",
+      "max_abs": 0.6220522522926331,
+      "mean_abs": 0.008647891692817211,
+      "ref_max_abs": 1.0,
+      "nan_out": false,
+      "inf_out": false
+    }
+  }
+}

metadata/run_manifest.json

@@ -0,0 +1,10 @@
+{
+  "python": "3.12.12 (main, Oct 10 2025, 08:52:57) [GCC 11.4.0]",
+  "platform": "Linux-6.6.105+-x86_64-with-glibc2.35",
+  "torch": "2.9.0+cu126",
+  "torch_cuda": "12.6",
+  "cuda_available": true,
+  "gpu": "NVIDIA L4",
+  "checkpoint_path": "/content/sharp_work/weights/sharp_2572gikvuh.pt",
+  "checkpoint_sha256": "94211a75198c47f61fca7d739ba08a215418d8d398d48fddf023baccc24f073d"
+}

metadata/sha256sums.txt

@@ -0,0 +1,10 @@
+cf9fd1ebe78d3effd9f782d053ffbed7f69b033e5260e706c0f216e0bd4675fa  LICENSE
+805ef0cac7c06d877615f80ff86da399f30fb1d73bd3fb6c522605c3400f8eb5  README.md
+438d4881b4ece849eaea46065aeaa4243a260648a5619b330f31c30b730fbcf4  metadata/parity_metrics.json
+5613ce0464dfbd2bc0902e2f874a836ea800dc6e2ec1a8e9213d98360e15bab7  metadata/run_manifest.json
+d5e13c449dd15f8538c26566bbbffa10172278b9d5f6e500a27c0c7c837f3d4f  onnx/fp32/export_config.json
+97c7c35a1e5ff1c1d0762556952cfeb5cc0ff3915cd0118b8c1c0359829b59de  onnx/fp32/sharp_ndc_opset18.onnx
+23caa148af9590d4880c47fb69ba1dbd56dde91eab50e2c3d9254ddc1d001604  onnx/fp32/sharp_ndc_opset18.onnx.data
+9174fe456d17806fc9c37d843f24d928d5a7c11836bd63bc39eef4427ba8b8ca  onnx/int8_dynamic/quant_int8_dynamic_config.json
+a4dafd5f2b66caa375c277cce0ee6e9203a729f24edba1a07a9bd3b7cc0cfd88  onnx/int8_dynamic/sharp_ndc_opset18_int8_dynamic.onnx
+42c516e4c09adb18941cc84ca2459fb8907b4ad3613fc9753d9b77f1b1820432  onnx/int8_dynamic/sharp_ndc_opset18_int8_dynamic.onnx.data

onnx/fp32/export_config.json

@@ -0,0 +1,24 @@
+{
+  "onnx_path": "/content/sharp_work/phase3/onnx/sharp_ndc_opset18.onnx",
+  "opset_version": 18,
+  "inputs": {
+    "image_resized_pt": [
+      1,
+      3,
+      1536,
+      1536
+    ],
+    "disparity_factor": [
+      1
+    ]
+  },
+  "outputs": [
+    "mean_vectors",
+    "singular_values",
+    "quaternions",
+    "colors",
+    "opacities"
+  ],
+  "external_data": true,
+  "dynamo": true
+}

onnx/fp32/sharp_ndc_opset18.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97c7c35a1e5ff1c1d0762556952cfeb5cc0ff3915cd0118b8c1c0359829b59de
+size 7279677

onnx/fp32/sharp_ndc_opset18.onnx.data

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23caa148af9590d4880c47fb69ba1dbd56dde91eab50e2c3d9254ddc1d001604
+size 2616066048

onnx/int8_dynamic/quant_int8_dynamic_config.json

@@ -0,0 +1,12 @@
+{
+  "source_model": "sharp_ndc_opset18.onnx",
+  "output_model": "sharp_ndc_opset18_int8_dynamic.onnx",
+  "method": "dynamic",
+  "quant_format": "QDQ",
+  "weight_type": "QInt8",
+  "op_types_to_quantize": [
+    "MatMul",
+    "Gemm"
+  ],
+  "onnxruntime_version": "1.23.2"
+}

onnx/int8_dynamic/sharp_ndc_opset18_int8_dynamic.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a4dafd5f2b66caa375c277cce0ee6e9203a729f24edba1a07a9bd3b7cc0cfd88
+size 6820060

onnx/int8_dynamic/sharp_ndc_opset18_int8_dynamic.onnx.data

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42c516e4c09adb18941cc84ca2459fb8907b4ad3613fc9753d9b77f1b1820432
+size 803940096