hailo_model_zoo: Skill for Claude Code

Hailo Parse

First stage of the DFC flow: take an ONNX (.onnx) or TensorFlow (.tflite / SavedModel) file and produce a Hailo Archive (.har) containing the network in Hailo's internal representation. The HAR is then consumed by /hailo-optimize.

This skill is shipped with the Hailo Model Zoo. It assumes the DFC wheel (hailo_sdk_client) is installed in the active virtualenv.

Inputs

Arg	Required	Notes
<path-to-model>	yes	Path to an .onnx or .tflite file.
[hw_arch]	no	One of hailo10h (default), hailo15h, hailo15l

Optional via follow-up questions: model_name (defaults to filename stem), explicit start_node_names / end_node_names / net_input_shapes overrides.

TF input preconditions (verify before parsing — these are common failure modes that surface as confusing parser errors):

• TF1 frozen graphs / checkpoints: deprecated by the DFC since v5.1.0. Convert to TF2 SavedModel and then to TFLite (see next bullet) before parsing.
• TF2 SavedModel: not directly accepted by the parser. Convert to TFLite first with tf.lite.TFLiteConverter.from_saved_model(...).convert() and pass the resulting .tflite.
• Int8-quantized TFLite: not supported by the parser. Use a 32-bit or 16-bit TFLite — quantization is applied later by /hailo-optimize, not at parse time.

ONNX export checklist (PyTorch users — the most common parse-failure sources):

When the user is exporting from PyTorch with torch.onnx.export(...), the following arguments are required for clean DFC parsing:

• training=torch.onnx.TrainingMode.PRESERVE (or EVAL) — otherwise BatchNorm folds incorrectly and the parsed model silently diverges from the source.
• do_constant_folding=False.
• opset_version in 15–21 (DFC supported range).
• dynamo=False — the new TorchDynamo exporter is not supported.

If the user can't change the export, suggest re-exporting before parsing.

• AdaptiveAvgPool2d: not exportable to a fixed ONNX graph when its output size does not evenly divide the input size — this is exactly the situation in pyramid-pooling modules (common in segmentation networks). Replace it before export with a fixed AvgPool2d(kernel_size=ceil(in/out), stride=floor(in/out)). On a real pyramid-pooling segmentation network this swap changed the metric by under 0.01, so it's safe to apply by default. Verify it by comparing ONNX Runtime against the original torch model on the same inputs before parsing.