# WIA-AI-014: Neural Network Format - Phase 1 Data Format Specification **Version:** 1.0.0 **Status:** Draft **Last Updated:** 2025-01-15 **Philosophy:** 弘益人間 (Hongik Ingan) - Benefit All Humanity ## 1. Introduction This specification defines the data format requirements for WIA-AI-014 Neural Network Format standard, enabling universal model interchange and deployment across frameworks and platforms. ### 1.1 Scope This phase covers: - Binary serialization format - Tensor representation - Graph structure encoding - Metadata schema - Compression and optimization ### 1.2 Goals - **Interoperability:** Seamless conversion between major ML frameworks - **Efficiency:** Optimized storage and loading performance - **Completeness:** Preserve all necessary model information - **Extensibility:** Support future architectures and operations ## 2. File Format Structure ### 2.1 Container Format WIA-AI-014 uses a hierarchical container structure: ``` WIA-AI-014 Model Container ├── Header (256 bytes) ├── Metadata Block ├── Graph Definition Block ├── Tensor Data Block └── Signature Block ``` ### 2.2 Header Format ```c struct WIAHeader { uint32_t magic; // 0x57494131 ('WIA1') uint16_t version_major; // 1 uint16_t version_minor; // 0 uint64_t metadata_offset; uint64_t metadata_size; uint64_t graph_offset; uint64_t graph_size; uint64_t tensor_offset; uint64_t tensor_size; uint64_t signature_offset; uint64_t signature_size; uint8_t compression; // 0=none, 1=gzip, 2=zstd uint8_t encryption; // 0=none, 1=AES-256 uint8_t reserved[214]; }; ``` ## 3. Tensor Representation ### 3.1 Tensor Data Types Supported data types with enum values: | Type | Code | Bytes | Description | |------|------|-------|-------------| | FLOAT32 | 1 | 4 | IEEE 754 single precision | | FLOAT16 | 2 | 2 | IEEE 754 half precision | | BFLOAT16 | 3 | 2 | Brain floating point | | INT64 | 4 | 8 | 64-bit signed integer | | INT32 | 5 | 4 | 32-bit signed integer | | INT16 | 6 | 2 | 16-bit signed integer | | INT8 | 7 | 1 | 8-bit signed integer | | UINT8 | 8 | 1 | 8-bit unsigned integer | | BOOL | 9 | 1 | Boolean (0 or 1) | | COMPLEX64 | 10 | 8 | Complex number (2x float32) | ### 3.2 Tensor Structure ```protobuf message Tensor { string name = 1; repeated int64 dims = 2; DataType dtype = 3; oneof data { bytes raw_data = 4; ExternalData external_data = 5; } TensorMetadata metadata = 6; } message ExternalData { string location = 1; int64 offset = 2; int64 length = 3; string checksum = 4; // SHA-256 } message TensorMetadata { string doc_string = 1; repeated KeyValue properties = 2; QuantizationParams quantization = 3; } ``` ### 3.3 Tensor Layout Tensors use row-major (C-style) layout by default: ```python # For tensor with shape [N, C, H, W] # Index calculation: index = n*C*H*W + c*H*W + h*W + w # Memory layout: [ N0C0H0W0, N0C0H0W1, ..., N0C0H0Wn, N0C0H1W0, N0C0H1W1, ..., N0C0H1Wn, ... ] ``` ## 4. Graph Representation ### 4.1 Computational Graph Structure ```protobuf message Graph { string name = 1; repeated Node nodes = 2; repeated ValueInfo inputs = 3; repeated ValueInfo outputs = 4; repeated Tensor initializers = 5; GraphMetadata metadata = 6; } message Node { string name = 1; string op_type = 2; repeated string inputs = 3; repeated string outputs = 4; repeated Attribute attributes = 5; string doc_string = 6; } message ValueInfo { string name = 1; TypeProto type = 2; string doc_string = 3; } ``` ### 4.2 Operator Definitions Operators follow ONNX operator set conventions: ```yaml operators: - name: Conv domain: "" since_version: 1 inputs: - name: X type: T - name: W type: T - name: B type: T optional: true outputs: - name: Y type: T attributes: - name: kernel_shape type: ints required: true - name: strides type: ints default: [1, 1] ``` ## 5. Metadata Schema ### 5.1 Model Metadata ```json { "model_info": { "name": "string", "version": "semver", "description": "string", "author": "string", "license": "string", "created_at": "iso8601", "framework": "string", "framework_version": "string" }, "training_info": { "dataset": "string", "hyperparameters": {}, "metrics": {} }, "deployment_info": { "target_platforms": ["string"], "input_preprocessing": {}, "output_postprocessing": {} } } ``` ## 6. Compression and Optimization ### 6.1 Supported Compression - **None (0):** No compression - **GZIP (1):** Standard gzip compression - **ZSTD (2):** Zstandard compression (recommended) ### 6.2 Tensor Compression Individual tensors may be compressed: ```protobuf message CompressedTensor { Tensor tensor = 1; CompressionType compression = 2; bytes compressed_data = 3; int64 uncompressed_size = 4; } ``` ## 7. Validation and Integrity ### 7.1 Checksums Each major block includes SHA-256 checksum: ```protobuf message BlockChecksum { string block_type = 1; bytes sha256 = 2; } ``` ### 7.2 Model Signature Optional cryptographic signature: ```protobuf message Signature { string algorithm = 1; // "RSA-SHA256", "ECDSA" bytes public_key = 2; bytes signature = 3; int64 timestamp = 4; } ``` ## 8. Quantization Support ### 8.1 Quantization Parameters ```protobuf message QuantizationParams { QuantizationType type = 1; // LINEAR, LOGARITHMIC float scale = 2; int32 zero_point = 3; DataType quantized_type = 4; } enum QuantizationType { NONE = 0; LINEAR = 1; LOGARITHMIC = 2; } ``` ### 8.2 Per-Channel Quantization ```protobuf message PerChannelQuantization { repeated float scales = 1; repeated int32 zero_points = 2; int32 channel_dim = 3; } ``` ## 9. External Data References ### 9.1 Large Tensor Storage For models >2GB, tensors can be stored externally: ```yaml model_structure: - model.wia: - header - metadata - graph_definition - weights/: - layer1.bin - layer2.bin - layer3.bin ``` ### 9.2 Reference Format ```json { "external_data": { "location": "weights/layer1.bin", "offset": 0, "length": 1048576, "checksum": "sha256:abc123..." } } ``` ## 10. Versioning and Compatibility ### 10.1 Version Numbering - **Major version:** Breaking changes to format structure - **Minor version:** Backward-compatible additions - **Patch version:** Bug fixes and clarifications ### 10.2 Compatibility Rules - Readers MUST support all minor versions within same major version - Readers MAY support older major versions - Writers SHOULD write latest stable version - Unknown extensions MUST be ignored (forward compatibility) ## 11. Example: Complete Model ```python # Pseudo-code for WIA-AI-014 model structure model = WIAModel( header=WIAHeader( magic=0x57494131, version=(1, 0), compression=ZSTD ), metadata=ModelMetadata( name="ResNet50", version="2.1.0", framework="pytorch", created_at="2025-01-15T10:00:00Z" ), graph=Graph( name="main", nodes=[ Node(name="conv1", op_type="Conv", ...), Node(name="relu1", op_type="Relu", ...) ], inputs=[ValueInfo(name="input", type=TensorType(...))], outputs=[ValueInfo(name="output", type=TensorType(...))] ), tensors={ "conv1.weight": Tensor(dims=[64,3,7,7], dtype=FLOAT32, ...), "conv1.bias": Tensor(dims=[64], dtype=FLOAT32, ...) } ) ``` ## 12. Compliance Requirements ### 12.1 Minimum Requirements A WIA-AI-014 compliant implementation MUST: - Support all data types in Section 3.1 - Implement tensor serialization per Section 3.2 - Support graph representation per Section 4.1 - Validate checksums per Section 7.1 ### 12.2 Optional Features Implementations MAY support: - Encryption (Section 2.2) - Compression beyond GZIP (Section 6.1) - External tensor storage (Section 9) - Cryptographic signatures (Section 7.2) ## 13. References - ONNX Format Specification: https://github.com/onnx/onnx/blob/main/docs/IR.md - Protocol Buffers: https://developers.google.com/protocol-buffers - IEEE 754 Floating Point: https://ieeexplore.ieee.org/document/8766229 --- **Copyright © 2025 WIA (World Certification Industry Association)** **弘益人間 (Hongik Ingan) · Benefit All Humanity**