yobx.tensorflow.ops.pooling#

Converters for TF 2-D pooling ops: MaxPool, AvgPool.

TensorFlow uses NHWC (batch, height, width, channels) data format by default. ONNX MaxPool and AveragePool expect NCHW input. The converters therefore insert Transpose nodes:

  • Before pooling: NHWC → NCHW (perm=[0, 3, 1, 2])

  • After pooling: NCHW → NHWC (perm=[0, 2, 3, 1])

Padding is always expressed as explicit pads rather than auto_pad so that ONNX shape inference can propagate output shapes even when the batch dimension is dynamic.

yobx.tensorflow.ops.pooling.convert_avg_pool(g: GraphBuilderExtendedProtocol, sts: Dict[str, Any], outputs: List[str], op: Operation) → str[source]#

Converts TF AvgPool (NHWC) → ONNX AveragePool (NCHW).

TF ksize / strides attributes are 4-element lists in NHWC order [batch, H, W, C]; only the spatial dimensions [H, W] are forwarded to the ONNX node. Padding is represented as explicit pads to enable ONNX shape inference regardless of whether the batch dimension is dynamic.

yobx.tensorflow.ops.pooling.convert_max_pool(g: GraphBuilderExtendedProtocol, sts: Dict[str, Any], outputs: List[str], op: Operation) → str[source]#

Converts TF MaxPool (NHWC) → ONNX MaxPool (NCHW).

TF ksize / strides attributes are 4-element lists in NHWC order [batch, H, W, C]; only the spatial dimensions [H, W] are forwarded to the ONNX node. Padding is represented as explicit pads to enable ONNX shape inference regardless of whether the batch dimension is dynamic.