ailia SDK

Bundled with ailia SDK: the ONNX inference API (ailia) plus high-level helpers for common tasks — Classification for image classification, Detector for object detection, PoseEstimation for skeletal pose, and ailia Audio for audio pre/post-processing.

Supplemental libraries (separate packages): ailia Tokenizer, ailia Speech, ailia Voice, and ailia Tracker ship as separate libraries on top of the SDK.

The evaluation license is downloaded automatically at runtime for Python, Unity, Flutter, and JNI, and is valid for one month for C++. It is intended for development and trial use only.

For commercial deployment, redistribution, or longer-term use, request a production license. See the ailia license terms for details.

Pass an env_id to ailia.Net(). List available environments (CPU, CUDA, Metal, Vulkan, MKL) with ailia.get_environment_list(), then select the one you want.

By default, ailia chooses the fastest available environment for your platform.

To use CUDA, install the CUDA Toolkit and cuDNN. See the CUDA Toolkit / cuDNN Installation Guide for details.

To use Vulkan, see the Vulkan Setup Guide.

Sample scripts in ailia-models download .onnx and .onnx.prototxt files into the model's own directory the first time you run them. Subsequent runs reuse the cached files.

For ailia Speech and ailia Voice, models are downloaded into ./models/ by default, configurable via initialize_model(model_path=...).

Yes, after the first run. The evaluation license and any auto-downloaded model files require an internet connection on first use; once cached, subsequent inference works offline.

For C++, the license is fetched once via download_license.py from the binding repository.

Export your model to ONNX with the framework's standard tooling (torch.onnx.export, tf2onnx, etc.), then generate the matching .onnx.prototxt using the script bundled with ailia SDK.

The model conversion tutorial walks through the process.

No, prototxt is optional when loading an ONNX model directly (e.g. ailia.Net(stream=None, weight="model.onnx")). The ailia-models repo ships prototxt files alongside ONNX so that Netron can visualize them quickly, but your own models work without one.

The Python API accepts a name-to-array dict as the argument to net.run(), so multi-input models work out of the box and multi-output results come back as a list.

The C API's ailiaPredict only supports one input and one output. For multi-IO models, write each input via ailiaSetInputBlobData, run inference with ailiaUpdate, then read each output via ailiaGetBlobData. Use ailiaFindBlobIndexByName to look up blob indices by name.

Input and output tensor buffers are always passed as float (FP32) regardless of the underlying ONNX datatype. Internally ailia executes the model using whatever datatype the ONNX defines (FP16, INT8 quantization, etc.). You can query the actual datatype of any blob with ailiaGetBlobDataType.

AILIAShape represents up to 4 dimensions through x / y / z / w + dim. x is the innermost (memory-contiguous) axis and w is the outermost. For a numpy-style (batch, channel, height, width) 4-D tensor, that maps to w = batch, z = channel, y = height, x = width.

The dim field tells you how many of those axes are valid:

• dim = 0: scalar (ONNX rank-0 tensor)
• dim = 1: x only
• dim = 2: x and y
• dim = 3: x / y / z
• dim = 4: x / y / z / w

Tensors with rank ≥ 5 cannot be expressed in AILIAShape. Use the ND variants instead — ailiaSetInputShapeND / ailiaGetOutputShapeND (with ailiaGetInputDim / ailiaGetOutputDim) — which take a flat unsigned int* shape array.

When the ONNX was exported with dynamic shape, the engine can't resolve the shape until you tell it the actual input dimensions. Calling inference or ailiaGetOutputShape before ailiaSetInputShape (or ailiaSetInputShapeND for rank ≥ 5) returns AILIA_STATUS_UNSETTLED_SHAPE (-18).

Python: net.run() inspects the input array's shape and calls ailiaSetInputShape for you, so no extra step is needed.

C / C# (Unity) / Kotlin (JNI) / Dart (Flutter): set the input shape explicitly via the equivalent ailiaSetInputShape call before running inference.

A built-in profiler reports per-layer timing.

Python: call net.set_profile_mode(True), run inference, and read the per-layer summary with net.get_summary().

C: enable profiling with ailiaSetProfileMode, run inference, then call ailiaGetSummaryLength to get the buffer size and ailiaSummary to fill it with the summary string.

The default mode is speed-first and keeps every intermediate tensor in memory. Enabling intermediate reuse and constant reduction lowers peak memory.

Python: pass a bit-flag to set_memory_mode:

memory_mode = ailia.get_memory_mode(
    reduce_constant=True,
    ignore_input_with_initializer=True,
    reduce_interstage=False,
    reuse_interstage=True,
)
net.set_memory_mode(memory_mode)

C: the equivalent control is ailiaSetMemoryMode. OR together flags such as AILIA_MEMORY_REDUCE_CONSTANT and AILIA_MEMORY_REUSE_INTERSTAGE.

For bug reports and questions on the sample repositories, open an issue on the relevant GitHub repo. For SDK licensing and commercial inquiries, contact ailia Inc. directly.