afe.apis.transform

Tensor transformations that can be applied to a model’s input or output.

Classes

Transform

A transformation on a tensor.

Functions

argmax(→ Transform)

The argmax operation applied over a single axis of a tensor.

reshape_transform(→ Transform)

The reshape_transform operation

layout_transform(→ Transform)

The layout_transform operation.

tessellation_transform(→ Transform)

The tessellation_transform operation.

detessellation_transform(→ Transform)

The detessellation_transform operation.

pack_transform(→ _PackTransform)

The pack_transform operation.

unpack_transform(→ _UnpackTransform)

The unpack_transform operation.

normalization_transform(→ Transform)

The normalization_transform operation.

quantization_transform(→ Transform)

The quantization_transform operation.

dequantization_transform(→ Transform)

The dequantization_transform operation.

resize_transform(→ Transform)

The resize_transform operation.

chroma_upsample_transform(→ Transform)

The chroma_upsample_transform operation.

yuv_rgb_conversion_transform(→ Transform)

The yuv_rgb_conversion_transform operation.

bgr_rgb_conversion_transform(→ Transform)

The bgr_rgb_conversion_transform operation.

crop_transform(→ Transform)

The crop_transform operation.

slice_transform(→ Transform)

The slice_transform operation.

sigmoid_transform(→ Transform)

The sigmoid_transform operation.

nms_maxpool_transform(→ Transform)

The nms_maxpool_transform operation.

softmax_transform(→ Transform)

The softmax transform operation.

compose(→ Transform)

Make a transform that is the composition of the given transforms.

identity(→ Transform)

An identity transformation on a tensor.

Module Contents

class afe.apis.transform.Transform[source]

A transformation on a tensor.

abstract apply(t: afe.apis.compilation_job_base.Tensor) → afe.apis.compilation_job_base.Tensor[source]

Apply this transform to a tensor. Raise an exception if the tensor’s type is not suitable for this transform.

abstract get_result_type(parameter_type: afe.ir.tensor_type.TensorType) → afe.ir.tensor_type.TensorType[source]

Get the type of this transform’s output when it is applied to input of the given type. Raise an exception if the transform cannot be applied to a tensor having the given type.

abstract extract_ir(input_type: afe.ir.tensor_type.TensorType, input_name: afe.ir.defines.NodeName, output_base_name: afe.ir.defines.NodeName) → tuple[afe.ir.tensor_type.TensorType, list[afe.ir.node.AwesomeNode]][source]

Extract the internal representation of this transform.

afe.apis.transform.argmax(*, axis: int) → Transform[source]

The argmax operation applied over a single axis of a tensor.

Parameters:

axis – Axis to reduce

Returns:

The argmax transformation

afe.apis.transform.reshape_transform(*, newshape: list[int]) → Transform[source]

The reshape_transform operation

Parameters:

newshape – New shape after reshape transform

Returns:

The reshape transformation

afe.apis.transform.layout_transform(*, src_layout: str, dst_layout: str) → Transform[source]

The layout_transform operation.

Parameters:

  • src_layout – Layout of the input tensor

  • dst_layout – Layout of the output tensor

Returns:

The layout_transform transformation

afe.apis.transform.tessellation_transform(*, slice_shape: Sequence[int], align_c16: bool, cblock: bool) → Transform[source]

The tessellation_transform operation.

Parameters:

  • slice_shape – Shape of slice to tessellate.

  • align_c16 – True to align channels to 16 in a tessellated slice

  • cblock – True to interleave the channel blocks in a tessellated slice.

Returns:

The tessellation_transform transformation

afe.apis.transform.detessellation_transform(*, slice_shape: Sequence[int], align_c16: bool, cblock: bool, frame_type: afe.ir.tensor_type.TensorType) → Transform[source]

The detessellation_transform operation.

Parameters:

  • slice_shape – Shape of slice to detessellate

  • align_c16 – True to indicate that slice channels are aligned to 16.

  • cblock – True to indicate that channel blocks are interleaved in a slice.

  • frame_type – Tensor type of de-tessellated frame

Returns:

The detessellation_transform transformation

afe.apis.transform.pack_transform() → _PackTransform[source]

The pack_transform operation.

Returns:

The pack_transform transformation

afe.apis.transform.unpack_transform(*, tensor_types: list[afe.ir.tensor_type.TensorType]) → _UnpackTransform[source]

The unpack_transform operation.

Parameters:

tensor_types – List of target tensor types after unpack

Returns:

The unpack_transform transformation

afe.apis.transform.normalization_transform(*, channel_params: list[tuple[float, float, float]]) → Transform[source]

The normalization_transform operation.

Parameters:

channel_params – The list of tuples for (divisor, mean, standard deviation)

Returns:

The normalization_transform transformation

afe.apis.transform.quantization_transform(*, channel_params: list[tuple[float, int]], num_bits: int, rounding: ml_kernels.math_helpers.RoundType = _RoundType.TONEAREST) → Transform[source]

The quantization_transform operation.

Parameters:

  • channel_params – The list of tuples for (quant_scale, zero_point)

  • num_bits – The number of bits used for quantization

  • rounding – The rounding type for quantization

Returns:

The quantization_transform transformation

afe.apis.transform.dequantization_transform(*, channel_params: list[tuple[float, int]]) → Transform[source]

The dequantization_transform operation.

Parameters:

channel_params – The list of tuples for (quant_scale, zero_point)

Returns:

The dequantization_transform transformation

afe.apis.transform.resize_transform(*, target_height: int, target_width: int, keep_aspect: bool, deposit_location: afe.apis.defines.ResizeDepositLocation, method: afe.apis.defines.ResizeMethod) → Transform[source]

The resize_transform operation.

Parameters:

  • target_height – Target height of resized tensor

  • target_width – Target width of resized tensor

  • keep_aspect – Boolean flag to keep aspect ratio

  • deposit_location – Enum to indicate deposit position of resized image

  • method – Enum to indicate supported interpolation methods

Returns:

The resize_transform transformation

afe.apis.transform.chroma_upsample_transform(*, frame_height: int, frame_width: int, yuv_sampling: afe.apis.defines.ChromaSampling) → Transform[source]

The chroma_upsample_transform operation.

Parameters:

  • frame_height – Height of full sampling frame

  • frame_width – Width of full sampling frame

  • yuv_sampling – Chroma sampling Enum

Returns:

The chroma_upsample_transform transformation

afe.apis.transform.yuv_rgb_conversion_transform(*, conversion: afe.apis.defines.ColorConversion, std: afe.apis.defines.ColorSpaceStandard) → Transform[source]

The yuv_rgb_conversion_transform operation.

Parameters:

  • conversion – Direction of conversion between YUV and RGB

  • std – Standard for color space conversion

Returns:

The yuv_rgb_conversion_transform transformation

afe.apis.transform.bgr_rgb_conversion_transform(*, conversion: afe.apis.defines.ColorConversion) → Transform[source]

The bgr_rgb_conversion_transform operation.

Parameters:

conversion – Direction of conversion between BGR and RGB

Returns:

The bgr_rgb_conversion_transform transformation

afe.apis.transform.crop_transform(*, bounding_box: list[tuple[int, int]]) → Transform[source]

The crop_transform operation.

Parameters:

bounding_box – Rectangle area to crop, as a list of 2 items. bounding_box[0] is the (x,y) position of the cropped area’s origin within the input area. bounding_box[1] is the (x,y) past-the-end position of the cropped area within the input area.

Returns:

The crop_transform transformation

afe.apis.transform.slice_transform(*, begin: list[int], end: list[int]) → Transform[source]

The slice_transform operation.

Parameters:

  • begin – Begin indices for each dimension.

  • end – End indices for each dimension.

Returns:

The slice_transform transformation

afe.apis.transform.sigmoid_transform(*, save_int16: bool) → Transform[source]

The sigmoid_transform operation.

Parameters:

save_int16 – Boolean flag to save output as 16-bit fixed point

Returns:

The sigmoid_transform transformation

afe.apis.transform.nms_maxpool_transform(*, kernel: int) → Transform[source]

The nms_maxpool_transform operation.

Parameters:

kernel – Size of pooling kernel

Returns:

The nms_maxpool_transform transformation

afe.apis.transform.softmax_transform(*, axis: int) → Transform[source]

The softmax transform operation.

Parameters:

axis – Axis to sum over

Returns:

The softmax transform

afe.apis.transform.compose(transforms: list[Transform]) → Transform[source]

Make a transform that is the composition of the given transforms.

The composition performs transforms in the same order as they occur in the list:

compose([x, y]).apply(t) == y.apply(x.apply(t))

Parameters:

transforms – Transforms to compose

Returns:

Their composition

afe.apis.transform.identity() → Transform[source]

An identity transformation on a tensor.