Skip to content

Latest commit

 

History

History
562 lines (379 loc) · 10.5 KB

operator_ir.md

File metadata and controls

562 lines (379 loc) · 10.5 KB
Raw

Operator Schemas

This documentation describes the operator definitions.

BatchNorm

BatchNorm operator carries out batch normalization, only for inference phase.

Inputs:

  • input: float32

    the input 4-dimensional tensor of shape NCHW

  • gamma: float32

  • beta: float32

    the bias as 1-dimensional tensor of size C(Channel)

  • mean: float32

    the estimated mean as 1-dimensional tensor of size C(Channel)

  • var: float32

    the estimated variance as 1-dimensional tensor of size C(Channel)

Outputs:

  • output: float32

Parameters:

  • caffe_flavor: int

    if use caffe version batch_normalization. Default set to 1.

  • rescale_factor: float32

Concat

Concatenate a list of input tensors into a single output tensor.

Inputs:

  • input: float32

Outputs:

  • output: float32

Parameters:

  • axis: int

    which axis to concat on, default is set to 1.

ConstOp

A Constant tensor.

Inputs:

None

Outputs:

  • output: float32

Convolution

It computes the output of convolution of an input tensor and filter kernel.

Inputs:

  • input: float32
  • weight: float32
  • bias: float32

Outputs:

  • output: float32

Parameters:

  • kernel_h: int

    kernel size height

  • kernel_w: int

  • kernel size width

  • stride_h: int

    stride size height

  • stride_w: int

  • stride size width

  • pad_h: int

    pad size height

  • pad_w: int

    pad size width

  • dilation_h :int

  • dilation_w :int

  • output_channel: int

    number of output channel (number of kernel)

  • group: int

  • activation: int

Deconvolution

Deconvolution operator multiplies each input value by a kernel elementwise, and sums over the resulting on output windows.

Inputs:

  • input: float32
  • weight: float32
  • bias: float32

Outputs:

  • output: float32

Parameters:

  • kernel_size: int

  • stride: int

    stride size

  • pad: int

    pad size

  • num_output: int

    number of output channel (number of kernel)

  • dilation: int

Detection_output

Detection_output operator used in SSD-detection network.

Inputs:

  • input: float32

Outputs:

  • output: float32 which is coordinates of detected boxes and corresponding confidences for each class

Parameters:

  • num_classes: int

    number of classes of detection benchmark (21 for VOC and 81 for COCO)

  • confidence_threshold: float

  • nms_threshold: float

  • keep_top_k: int

    num of top_k keeping for results of nms

Dropout

Dropout operator for inference phase is Y=X.

Inputs:

  • input: float32

Outputs:

  • output: float32

Eltwise

Compute elementwise operations, such as max or sum, along multiple input tensors.

Inputs:

  • input1: float32
  • input2: float32

Outputs:

  • output: float32

Parameters:

  • type: enum (MAX, SUM)

Flatten

The Flatten operator is a utility op that flattens an input of shape [n, c, h, w] to a simple vector output of shape [n, (chw),1, 1].

Inputs:

  • input1: float32

Outputs:

  • output: float32

Parameters:

  • axis : int

Fully_connected

Fully-connnected computes the results of X*W+b with X as input,W as weight and b as bias.

Inputs:

  • input: float32
  • weight: float32
  • bias: float32

Outputs:

  • output: float32

Parameters:

  • num_output: int

    number of output, which is the size of bias

Input_op

Inpute operator to feed data into network

Inputs:

None

Outputs:

  • output: float32

LRN

Local Response Normalization normalizes over local input regions.

Inputs:

  • input: float32

Outputs:

  • output: float32

Parameters:

  • local_size: int
  • norm_region: int
  • alpha: float
  • beta: float
  • k: float

LSTM

LSTM operator.

Inputs:

  • input: float32
  • kernel: float32
  • bias: float32,(optional)
  • init_c: float32,(optional)
  • init_h: float32,(optional)

Outputs:

  • output: float32

Parameters:

  • forget_bias: float32
  • has_peephole: bool
  • has_projection: bool
  • input_size: int32
  • hidden_size: int32
  • cell_size: int32
  • output_len: int32

Normalize

Normalize operator normalizes the input alone channel axis with L2 normalization, used in SSD-detection network

Inputs:

  • input: float32
  • scale: float32

Outputs:

  • output: float32

Permute

Permute operator permutes the input with specific order, used in SSD-detection network.

Inputs:

  • input: float32

Outputs:

  • output: float32

Parameters:

  • order0: int
  • order1: int
  • order2: int
  • order3: int

Pooling

Pooling takes input tensor and applies pooling according to the kernel sizes, stride sizes, pad sizes and pooling types.

Inputs:

  • input: float32

Outputs:

  • output: float32

Parameters:

  • alg: enum

    pooling type: kPoolMax, kPoolAvg

  • global: int

    if use global pooling

  • caffe_flavor: int

  • kernel_shape: a list of int

    the size of the kernel along each axis (H, W)

  • strides: a list of int

    stride along each axis (H, W).

  • pads: a list of int

    pads zero for each axis (x1_begin, x2_begin...x1_end, x2_end,...). In case of input of shape NCHW, the pads is (pad_top,pad_left,pad_bottom,pad_right)

Priorbox

Priorbox operator computes the prior boxes for SSD (single shot detection) network. It will compute the prior boxes according to the original image size or specific image size defined by proto, as well as according to other parameters: max box size, min box size, aspect ratio for box etc.

Inputs:

  • input: float32
  • image_width:int32
  • image_height:int32

Outputs:

  • output: float32

Parameters:

  • min_size: float32
  • max_size: float32
  • variance: float32
  • aspect_ratio: float32
  • flip: int32
  • clip: int32
  • img_size: int32
  • img_h: int32
  • img_w: int32

PReLu

ReLu(Parameterized Rectified Linear Unit) takes one input data (Tensor) and produces one output data (Tensor) through yi=max(0,xi)+slope_i*min(0,xi) with slopes for negative parts.

Inputs:

  • input: float32
  • slope: float32

Outputs:

  • output: float32

Region

Region operator is used in YOLO network. It is a post process for the network output to rescale the output into [0,1] for compute the final detection out boxes.

Inputs:

  • input: float32

Outputs:

  • output: float32

Parameters:

  • num_classes: int32
  • side: int32
  • num_box: int32
  • coords: int32
  • confidence_threshold: float32
  • nms_threshold: float32
  • biases: float32

Resize

Resize layer used in several networks for upsampling. Current resize op supports two resize methods: bilinear_resize and nearest_neighbor_resize.

Inputs:

  • input: float32

Outputs:

  • output: float32

Parameters:

  • type: 0 for NEAREST_NEIGHBOR, 1 for BILINEAR_RESIZE
  • scale_w: float32, scale_w = out_w/in_w
  • scale_h: float32, scale_h = out_h/in_h

Reorg

Reorg operator is used in YOLO network. It is a process for the network to re-organize the data according the stride.

Inputs:

  • input: float32

Outputs:

  • output: float32

Parameters:

  • stride: int32

Reshape

The Reshape operator can be used to change the dimensions of its input, without changing its data.

Inputs:

  • input: float32

Outputs:

  • output: float32

Parameters:

  • dim_0: int32
  • dim_1: int32
  • dim_2: int32
  • dim_3: int32

ReLu

Relu takes one input data (Tensor) and produces one output data (Tensor) where the rectified linear function, y = max(0, x), is applied to the tensor elementwise.

Inputs:

  • input: float32

Outputs:

  • output: float32

Parameters:

  • negative_slope: float

    the relu with negative_slope = 0.1, is also called leaky-activation

RPN

Region Proposal Network(RPN) operator used in Faster-RCNN network. It generates proposal anchors.

Inputs:

  • input0: float32 scoretensor

  • input1: float32 featmap tensor

Outputs:

  • output: float32

Parameters:

  • nms_thresh: float32

  • post_nms_topn: int

    postprocess_nms_topn

  • per_nms_topn: int preprocess_nms_topn

  • min_size: int

  • basesize: int

  • feat_stride: int

  • anchor_scales:

    a list of anchor scales

  • ratios:

    a list of ratios

Roi_pooling

Roi_pooling operator used in Faster-RCNN network. It performs max pooling on regions of interest(ROI) specified by input.

Inputs:

  • input0: float32

    input0 is [N x C x H x W] feature maps on which pooling is performed.

  • input1: float32

    Input[1] [ R x 4] contains a list R ROI with each 4 coordinates.

Outputs:

  • output: float32

Parameters:

  • pooled_h: int

    The pooled output height.

  • pooled_w: int

    The pooled output width.

  • spatial_scale: float Multiplicative spatial scale factor to translate ROI coordinates from their input scale to the scale used when pooling.

Scale

Scale operator computes the output as scaling the input Y=gamma*X+(bias).

Inputs:

  • input: float32
  • gamma: float32
  • beta: float32

Outputs:

  • output: float32

Parameters:

  • axis: int

    which axis to coerce the input into 2D, default is set to 1.

  • num_axes: int

    default set to 1

  • bias_term: int

    default set to 0

Slice

Slice op takes an input and slices it along either the num or channel dimension, outputting multiple sliced tensors.

Inputs:

  • input: float32

Outputs:

  • output1: float32
  • output2: float32

Parameters:

  • axis: int

    which axis to slice along, default is set to 1 (Channel).

Softmax

Softmax computes the softmax normalized values. The output tensor has the same shape of the input shape.

Inputs:

  • input: float32

Outputs:

  • output: float32

Parameters:

  • axis: int

    which axis to coerce the input into 2D, default is set to 1.