TensorConverter

A lightweight C header for safe and efficient tensor data layout and type conversion between ONNX and TFLite formats.

Author: Jiacheng.Du

Features

• Supports conversion between ONNX (NCHW) and TFLite (NHWC) tensor layouts
• Handles multiple data types: float32, int32, uint8, int64, int16, int8, float16
• Provides memory safety checks (overflow, null pointer, allocation failure)
• All API and types use snake_case naming convention
• Pure C99, header-only, cross-platform

Data Types

// Data type enum
typedef enum {
    TENSOR_FLOAT32 = 0,
    TENSOR_INT32 = 1,
    TENSOR_UINT8 = 2,
    TENSOR_INT64 = 3,
    TENSOR_INT16 = 6,
    TENSOR_INT8 = 8,
    TENSOR_FLOAT16 = 9
} tensor_data_type_t;

Layout Types

typedef enum {
    LAYOUT_UNKNOWN = 0,
    LAYOUT_NCHW = 1,    // ONNX: [N, C, H, W]
    LAYOUT_NHWC = 2,    // TFLite: [N, H, W, C]
    LAYOUT_GENERIC = 3
} tensor_layout_t;

Main Structures

typedef struct {
    int32_t* dims;
    size_t num_dims;
    tensor_data_type_t data_type;
    size_t total_elements;
    tensor_layout_t layout;
} tensor_shape_t;

typedef struct {
    void* data;
    size_t data_size;
    tensor_shape_t shape;
    bool success;
    char error_msg[256];
} conversion_result_t;

Main API

Conversion with Layout Support

conversion_result_t onnx_to_tflite_with_layout(
    const void* onnx_data,
    const int32_t* dims,
    size_t num_dims,
    tensor_data_type_t data_type,
    tensor_layout_t src_layout,
    tensor_layout_t dst_layout);

conversion_result_t tflite_to_onnx_with_layout(
    const void* tflite_data,
    const int32_t* dims,
    size_t num_dims,
    tensor_data_type_t data_type,
    tensor_layout_t src_layout,
    tensor_layout_t dst_layout);

Utilities

void free_conversion_result(conversion_result_t* result);
void print_tensor_info(const tensor_shape_t* shape);
size_t get_data_type_size(tensor_data_type_t data_type);
size_t calculate_total_elements(const int32_t* dims, size_t num_dims);
bool validate_tensor_shape(const int32_t* dims, size_t num_dims);

Usage Example

#include "tensor_converter.h"

// Example: Convert ONNX NCHW float32 tensor to TFLite NHWC
float onnx_data[1*3*224*224];
int32_t dims[4] = {1, 3, 224, 224};

// Convert with explicit layout conversion: NCHW -> NHWC
conversion_result_t result = onnx_to_tflite_with_layout(
    onnx_data, dims, 4, TENSOR_FLOAT32, LAYOUT_NCHW, LAYOUT_NHWC);

if (result.success) {
    printf("Conversion successful!\n");
    print_tensor_info(&result.shape);
    // Use result.data ...
    free_conversion_result(&result);
} else {
    printf("Conversion failed: %s\n", result.error_msg);
}

Safety Notes

• All memory allocations are checked for failure
• All index and size calculations are checked for overflow
• All pointer arguments are checked for null
• Use free_conversion_result to release memory in conversion_result_t
• The caller is responsible for providing valid input data and dimensions

License

This project is licensed under the MIT License - see the LICENSE file for details.

MIT License allows you to:

• ✅ Use the code commercially
• ✅ Modify and distribute
• ✅ Include in proprietary software
• ✅ Private use

Requirements:

• 📄 Include the original copyright notice
• 📄 Include the license text