contrib.ndarray¶

NDArray namespace used to register contrib functions

Functions

mxnet.contrib.ndarray.AdaptiveAvgPooling2D(data=None, output_size=_Null, out=None, name=None, **kwargs)¶

Applies a 2D adaptive average pooling over a 4D input with the shape of (NCHW). The pooling kernel and stride sizes are automatically chosen for desired output sizes.

• If a single integer is provided for output_size, the output size is (N x C x output_size x output_size) for any input (NCHW).

• If a tuple of integers (height, width) are provided for output_size, the output size is (N x C x height x width) for any input (NCHW).

Defined in src/operator/contrib/adaptive_avg_pooling.cc:L214

Parameters

• data (NDArray) – Input data

• output_size (Shape(tuple), optional, default=[]) – int (output size) or a tuple of int for output (height, width).

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.BilinearResize2D(data=None, like=None, height=_Null, width=_Null, scale_height=_Null, scale_width=_Null, mode=_Null, align_corners=_Null, out=None, name=None, **kwargs)¶

Perform 2D resizing (upsampling or downsampling) for 4D input using bilinear interpolation.

Expected input is a 4 dimensional NDArray (NCHW) and the output with the shape of (N x C x height x width). The key idea of bilinear interpolation is to perform linear interpolation first in one direction, and then again in the other direction. See the wikipedia of Bilinear interpolation for more details.

Defined in src/operator/contrib/bilinear_resize.cc:L220

Parameters

• data (NDArray) – Input data

• like (NDArray) – Resize data to it’s shape

• height (int, optional, default='1') – output height (required, but ignored if scale_height is defined or mode is not “size”)

• width (int, optional, default='1') – output width (required, but ignored if scale_width is defined or mode is not “size”)

• scale_height (float or None, optional, default=None) – sampling scale of the height (optional, used in modes “scale” and “odd_scale”)

• scale_width (float or None, optional, default=None) – sampling scale of the width (optional, used in modes “scale” and “odd_scale”)

• mode ({'like', 'odd_scale', 'size', 'to_even_down', 'to_even_up', 'to_odd_down', 'to_odd_up'},optional, default='size') – resizing mode. “simple” - output height equals parameter “height” if “scale_height” parameter is not defined or input height multiplied by “scale_height” otherwise. Same for width;”odd_scale” - if original height or width is odd, then result height is calculated like result_h = (original_h - 1) * scale + 1; for scale > 1 the result shape would be like if we did deconvolution with kernel = (1, 1) and stride = (height_scale, width_scale); and for scale < 1 shape would be like we did convolution with kernel = (1, 1) and stride = (int(1 / height_scale), int( 1/ width_scale);”like” - resize first input to the height and width of second input; “to_even_down” - resize input to nearest lower even height and width (if original height is odd then result height = original height - 1);”to_even_up” - resize input to nearest bigger even height and width (if original height is odd then result height = original height + 1);”to_odd_down” - resize input to nearest odd height and width (if original height is odd then result height = original height - 1);”to_odd_up” - resize input to nearest odd height and width (if original height is odd then result height = original height + 1);

• align_corners (boolean, optional, default=1) – With align_corners = True, the interpolating doesn’t proportionally align theoutput and input pixels, and thus the output values can depend on the input size.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.CTCLoss(data=None, label=None, data_lengths=None, label_lengths=None, use_data_lengths=_Null, use_label_lengths=_Null, blank_label=_Null, out=None, name=None, **kwargs)¶

Connectionist Temporal Classification Loss.

Note

The existing alias contrib_CTCLoss is deprecated.

The shapes of the inputs and outputs:

• data: (sequence_length, batch_size, alphabet_size)

• label: (batch_size, label_sequence_length)

• out: (batch_size)

The data tensor consists of sequences of activation vectors (without applying softmax), with i-th channel in the last dimension corresponding to i-th label for i between 0 and alphabet_size-1 (i.e always 0-indexed). Alphabet size should include one additional value reserved for blank label. When blank_label is "first", the 0-th channel is be reserved for activation of blank label, or otherwise if it is “last”, (alphabet_size-1)-th channel should be reserved for blank label.

label is an index matrix of integers. When blank_label is "first", the value 0 is then reserved for blank label, and should not be passed in this matrix. Otherwise, when blank_label is "last", the value (alphabet_size-1) is reserved for blank label.

If a sequence of labels is shorter than label_sequence_length, use the special padding value at the end of the sequence to conform it to the correct length. The padding value is 0 when blank_label is "first", and -1 otherwise.

For example, suppose the vocabulary is [a, b, c], and in one batch we have three sequences ‘ba’, ‘cbb’, and ‘abac’. When blank_label is "first", we can index the labels as {‘a’: 1, ‘b’: 2, ‘c’: 3}, and we reserve the 0-th channel for blank label in data tensor. The resulting label tensor should be padded to be:

[[2, 1, 0, 0], [3, 2, 2, 0], [1, 2, 1, 3]]

When blank_label is "last", we can index the labels as {‘a’: 0, ‘b’: 1, ‘c’: 2}, and we reserve the channel index 3 for blank label in data tensor. The resulting label tensor should be padded to be:

[[1, 0, -1, -1], [2, 1, 1, -1], [0, 1, 0, 2]]

out is a list of CTC loss values, one per example in the batch.

See Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks, A. Graves et al. for more information on the definition and the algorithm.

Defined in src/operator/nn/ctc_loss.cc:L100

Parameters

• data (NDArray) – Input ndarray

• label (NDArray) – Ground-truth labels for the loss.

• data_lengths (NDArray) – Lengths of data for each of the samples. Only required when use_data_lengths is true.

• label_lengths (NDArray) – Lengths of labels for each of the samples. Only required when use_label_lengths is true.

• use_data_lengths (boolean, optional, default=0) – Whether the data lenghts are decided by data_lengths. If false, the lengths are equal to the max sequence length.

• use_label_lengths (boolean, optional, default=0) – Whether the label lenghts are decided by label_lengths, or derived from padding_mask. If false, the lengths are derived from the first occurrence of the value of padding_mask. The value of padding_mask is 0 when first CTC label is reserved for blank, and -1 when last label is reserved for blank. See blank_label.

• blank_label ({'first', 'last'},optional, default='first') – Set the label that is reserved for blank label.If “first”, 0-th label is reserved, and label values for tokens in the vocabulary are between 1 and alphabet_size-1, and the padding mask is -1. If “last”, last label value alphabet_size-1 is reserved for blank label instead, and label values for tokens in the vocabulary are between 0 and alphabet_size-2, and the padding mask is 0.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.DeformableConvolution(data=None, offset=None, weight=None, bias=None, kernel=_Null, stride=_Null, dilate=_Null, pad=_Null, num_filter=_Null, num_group=_Null, num_deformable_group=_Null, workspace=_Null, no_bias=_Null, layout=_Null, out=None, name=None, **kwargs)¶

Compute 2-D deformable convolution on 4-D input.

The deformable convolution operation is described in https://arxiv.org/abs/1703.06211

For 2-D deformable convolution, the shapes are

• data: (batch_size, channel, height, width)

• offset: (batch_size, num_deformable_group * kernel[0] * kernel[1] * 2, height, width)

• weight: (num_filter, channel, kernel[0], kernel[1])

• bias: (num_filter,)

• out: (batch_size, num_filter, out_height, out_width).

Define:

f(x,k,p,s,d) = floor((x+2*p-d*(k-1)-1)/s)+1

then we have:

out_height=f(height, kernel[0], pad[0], stride[0], dilate[0])
out_width=f(width, kernel[1], pad[1], stride[1], dilate[1])

If no_bias is set to be true, then the bias term is ignored.

The default data layout is NCHW, namely (batch_size, channle, height, width).

If num_group is larger than 1, denoted by g, then split the input data evenly into g parts along the channel axis, and also evenly split weight along the first dimension. Next compute the convolution on the i-th part of the data with the i-th weight part. The output is obtained by concating all the g results.

If num_deformable_group is larger than 1, denoted by dg, then split the input offset evenly into dg parts along the channel axis, and also evenly split data into dg parts along the channel axis. Next compute the deformable convolution, apply the i-th part of the offset on the i-th part of the data.

Both weight and bias are learnable parameters.

Defined in src/operator/contrib/deformable_convolution.cc:L100

Parameters

• data (NDArray) – Input data to the DeformableConvolutionOp.

• offset (NDArray) – Input offset to the DeformableConvolutionOp.

• weight (NDArray) – Weight matrix.

• bias (NDArray) – Bias parameter.

• kernel (Shape(tuple), required) – Convolution kernel size: (h, w) or (d, h, w)

• stride (Shape(tuple), optional, default=[]) – Convolution stride: (h, w) or (d, h, w). Defaults to 1 for each dimension.

• dilate (Shape(tuple), optional, default=[]) – Convolution dilate: (h, w) or (d, h, w). Defaults to 1 for each dimension.

• pad (Shape(tuple), optional, default=[]) – Zero pad for convolution: (h, w) or (d, h, w). Defaults to no padding.

• num_filter (int, required) – Convolution filter(channel) number

• num_group (int, optional, default='1') – Number of group partitions.

• num_deformable_group (int, optional, default='1') – Number of deformable group partitions.

• workspace (long (non-negative), optional, default=1024) – Maximum temperal workspace allowed for convolution (MB).

• no_bias (boolean, optional, default=0) – Whether to disable bias parameter.

• layout ({None, 'NCDHW', 'NCHW', 'NCW'},optional, default='None') – Set layout for input, output and weight. Empty for default layout: NCW for 1d, NCHW for 2d and NCDHW for 3d.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.DeformablePSROIPooling(data=None, rois=None, trans=None, spatial_scale=_Null, output_dim=_Null, group_size=_Null, pooled_size=_Null, part_size=_Null, sample_per_part=_Null, trans_std=_Null, no_trans=_Null, out=None, name=None, **kwargs)¶

Performs deformable position-sensitive region-of-interest pooling on inputs. The DeformablePSROIPooling operation is described in https://arxiv.org/abs/1703.06211 .batch_size will change to the number of region bounding boxes after DeformablePSROIPooling

Parameters

• data (Symbol) – Input data to the pooling operator, a 4D Feature maps

• rois (Symbol) – Bounding box coordinates, a 2D array of [[batch_index, x1, y1, x2, y2]]. (x1, y1) and (x2, y2) are top left and down right corners of designated region of interest. batch_index indicates the index of corresponding image in the input data

• trans (Symbol) – transition parameter

• spatial_scale (float, required) – Ratio of input feature map height (or w) to raw image height (or w). Equals the reciprocal of total stride in convolutional layers

• output_dim (int, required) – fix output dim

• group_size (int, required) – fix group size

• pooled_size (int, required) – fix pooled size

• part_size (int, optional, default='0') – fix part size

• sample_per_part (int, optional, default='1') – fix samples per part

• trans_std (float, optional, default=0) – fix transition std

• no_trans (boolean, optional, default=0) – Whether to disable trans parameter.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.ModulatedDeformableConvolution(data=None, offset=None, mask=None, weight=None, bias=None, kernel=_Null, stride=_Null, dilate=_Null, pad=_Null, num_filter=_Null, num_group=_Null, num_deformable_group=_Null, workspace=_Null, no_bias=_Null, im2col_step=_Null, layout=_Null, out=None, name=None, **kwargs)¶

Compute 2-D modulated deformable convolution on 4-D input.

The modulated deformable convolution operation is described in https://arxiv.org/abs/1811.11168

For 2-D modulated deformable convolution, the shapes are

• data: (batch_size, channel, height, width)

• offset: (batch_size, num_deformable_group * kernel[0] * kernel[1] * 2, height, width)

• mask: (batch_size, num_deformable_group * kernel[0] * kernel[1], height, width)

• weight: (num_filter, channel, kernel[0], kernel[1])

• bias: (num_filter,)

• out: (batch_size, num_filter, out_height, out_width).

Define:

f(x,k,p,s,d) = floor((x+2*p-d*(k-1)-1)/s)+1

then we have:

out_height=f(height, kernel[0], pad[0], stride[0], dilate[0])
out_width=f(width, kernel[1], pad[1], stride[1], dilate[1])

If no_bias is set to be true, then the bias term is ignored.

The default data layout is NCHW, namely (batch_size, channle, height, width).

If num_group is larger than 1, denoted by g, then split the input data evenly into g parts along the channel axis, and also evenly split weight along the first dimension. Next compute the convolution on the i-th part of the data with the i-th weight part. The output is obtained by concating all the g results.

If num_deformable_group is larger than 1, denoted by dg, then split the input offset evenly into dg parts along the channel axis, and also evenly split out evenly into dg parts along the channel axis. Next compute the deformable convolution, apply the i-th part of the offset part on the i-th out.

Both weight and bias are learnable parameters.

Defined in src/operator/contrib/modulated_deformable_convolution.cc:L102

Parameters

• data (NDArray) – Input data to the ModulatedDeformableConvolutionOp.

• offset (NDArray) – Input offset to ModulatedDeformableConvolutionOp.

• mask (NDArray) – Input mask to the ModulatedDeformableConvolutionOp.

• weight (NDArray) – Weight matrix.

• bias (NDArray) – Bias parameter.

• kernel (Shape(tuple), required) – Convolution kernel size: (h, w) or (d, h, w)

• stride (Shape(tuple), optional, default=[]) – Convolution stride: (h, w) or (d, h, w). Defaults to 1 for each dimension.

• dilate (Shape(tuple), optional, default=[]) – Convolution dilate: (h, w) or (d, h, w). Defaults to 1 for each dimension.

• pad (Shape(tuple), optional, default=[]) – Zero pad for convolution: (h, w) or (d, h, w). Defaults to no padding.

• num_filter (int (non-negative), required) – Convolution filter(channel) number

• num_group (int (non-negative), optional, default=1) – Number of group partitions.

• num_deformable_group (int (non-negative), optional, default=1) – Number of deformable group partitions.

• workspace (long (non-negative), optional, default=1024) – Maximum temperal workspace allowed for convolution (MB).

• no_bias (boolean, optional, default=0) – Whether to disable bias parameter.

• im2col_step (int (non-negative), optional, default=64) – Maximum number of images per im2col computation; The total batch size should be divisable by this value or smaller than this value; if you face out of memory problem, you can try to use a smaller value here.

• layout ({None, 'NCDHW', 'NCHW', 'NCW'},optional, default='None') – Set layout for input, output and weight. Empty for default layout: NCW for 1d, NCHW for 2d and NCDHW for 3d.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.MultiBoxDetection(cls_prob=None, loc_pred=None, anchor=None, clip=_Null, threshold=_Null, background_id=_Null, nms_threshold=_Null, force_suppress=_Null, variances=_Null, nms_topk=_Null, out=None, name=None, **kwargs)¶

Convert multibox detection predictions.

Parameters

• cls_prob (NDArray) – Class probabilities.

• loc_pred (NDArray) – Location regression predictions.

• anchor (NDArray) – Multibox prior anchor boxes

• clip (boolean, optional, default=1) – Clip out-of-boundary boxes.

• threshold (float, optional, default=0.00999999978) – Threshold to be a positive prediction.

• background_id (int, optional, default='0') – Background id.

• nms_threshold (float, optional, default=0.5) – Non-maximum suppression threshold.

• force_suppress (boolean, optional, default=0) – Suppress all detections regardless of class_id.

• variances (tuple of <float>, optional, default=[0.1,0.1,0.2,0.2]) – Variances to be decoded from box regression output.

• nms_topk (int, optional, default='-1') – Keep maximum top k detections before nms, -1 for no limit.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.MultiBoxPrior(data=None, sizes=_Null, ratios=_Null, clip=_Null, steps=_Null, offsets=_Null, out=None, name=None, **kwargs)¶

Generate prior(anchor) boxes from data, sizes and ratios.

Parameters

• data (NDArray) – Input data.

• sizes (tuple of <float>, optional, default=[1]) – List of sizes of generated MultiBoxPriores.

• ratios (tuple of <float>, optional, default=[1]) – List of aspect ratios of generated MultiBoxPriores.

• clip (boolean, optional, default=0) – Whether to clip out-of-boundary boxes.

• steps (tuple of <float>, optional, default=[-1,-1]) – Priorbox step across y and x, -1 for auto calculation.

• offsets (tuple of <float>, optional, default=[0.5,0.5]) – Priorbox center offsets, y and x respectively

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.MultiBoxTarget(anchor=None, label=None, cls_pred=None, overlap_threshold=_Null, ignore_label=_Null, negative_mining_ratio=_Null, negative_mining_thresh=_Null, minimum_negative_samples=_Null, variances=_Null, out=None, name=None, **kwargs)¶

Compute Multibox training targets

Parameters

• anchor (NDArray) – Generated anchor boxes.

• label (NDArray) – Object detection labels.

• cls_pred (NDArray) – Class predictions.

• overlap_threshold (float, optional, default=0.5) – Anchor-GT overlap threshold to be regarded as a positive match.

• ignore_label (float, optional, default=-1) – Label for ignored anchors.

• negative_mining_ratio (float, optional, default=-1) – Max negative to positive samples ratio, use -1 to disable mining

• negative_mining_thresh (float, optional, default=0.5) – Threshold used for negative mining.

• minimum_negative_samples (int, optional, default='0') – Minimum number of negative samples.

• variances (tuple of <float>, optional, default=[0.1,0.1,0.2,0.2]) – Variances to be encoded in box regression target.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.MultiProposal(cls_prob=None, bbox_pred=None, im_info=None, rpn_pre_nms_top_n=_Null, rpn_post_nms_top_n=_Null, threshold=_Null, rpn_min_size=_Null, scales=_Null, ratios=_Null, feature_stride=_Null, output_score=_Null, iou_loss=_Null, out=None, name=None, **kwargs)¶

Generate region proposals via RPN

Parameters

• cls_prob (NDArray) – Score of how likely proposal is object.

• bbox_pred (NDArray) – BBox Predicted deltas from anchors for proposals

• im_info (NDArray) – Image size and scale.

• rpn_pre_nms_top_n (int, optional, default='6000') – Number of top scoring boxes to keep before applying NMS to RPN proposals

• rpn_post_nms_top_n (int, optional, default='300') – Number of top scoring boxes to keep after applying NMS to RPN proposals

• threshold (float, optional, default=0.699999988) – NMS value, below which to suppress.

• rpn_min_size (int, optional, default='16') – Minimum height or width in proposal

• scales (tuple of <float>, optional, default=[4,8,16,32]) – Used to generate anchor windows by enumerating scales

• ratios (tuple of <float>, optional, default=[0.5,1,2]) – Used to generate anchor windows by enumerating ratios

• feature_stride (int, optional, default='16') – The size of the receptive field each unit in the convolution layer of the rpn,for example the product of all stride’s prior to this layer.

• output_score (boolean, optional, default=0) – Add score to outputs

• iou_loss (boolean, optional, default=0) – Usage of IoU Loss

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.PSROIPooling(data=None, rois=None, spatial_scale=_Null, output_dim=_Null, pooled_size=_Null, group_size=_Null, out=None, name=None, **kwargs)¶

Performs region-of-interest pooling on inputs. Resize bounding box coordinates by spatial_scale and crop input feature maps accordingly. The cropped feature maps are pooled by max pooling to a fixed size output indicated by pooled_size. batch_size will change to the number of region bounding boxes after PSROIPooling

Parameters

• data (Symbol) – Input data to the pooling operator, a 4D Feature maps

• rois (Symbol) – Bounding box coordinates, a 2D array of [[batch_index, x1, y1, x2, y2]]. (x1, y1) and (x2, y2) are top left and down right corners of designated region of interest. batch_index indicates the index of corresponding image in the input data

• spatial_scale (float, required) – Ratio of input feature map height (or w) to raw image height (or w). Equals the reciprocal of total stride in convolutional layers

• output_dim (int, required) – fix output dim

• pooled_size (int, required) – fix pooled size

• group_size (int, optional, default='0') – fix group size

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.Proposal(cls_prob=None, bbox_pred=None, im_info=None, rpn_pre_nms_top_n=_Null, rpn_post_nms_top_n=_Null, threshold=_Null, rpn_min_size=_Null, scales=_Null, ratios=_Null, feature_stride=_Null, output_score=_Null, iou_loss=_Null, out=None, name=None, **kwargs)¶

Generate region proposals via RPN

Parameters

• cls_prob (NDArray) – Score of how likely proposal is object.

• bbox_pred (NDArray) – BBox Predicted deltas from anchors for proposals

• im_info (NDArray) – Image size and scale.

• rpn_pre_nms_top_n (int, optional, default='6000') – Number of top scoring boxes to keep before applying NMS to RPN proposals

• rpn_post_nms_top_n (int, optional, default='300') – Number of top scoring boxes to keep after applying NMS to RPN proposals

• threshold (float, optional, default=0.699999988) – NMS value, below which to suppress.

• rpn_min_size (int, optional, default='16') – Minimum height or width in proposal

• scales (tuple of <float>, optional, default=[4,8,16,32]) – Used to generate anchor windows by enumerating scales

• ratios (tuple of <float>, optional, default=[0.5,1,2]) – Used to generate anchor windows by enumerating ratios

• feature_stride (int, optional, default='16') – The size of the receptive field each unit in the convolution layer of the rpn,for example the product of all stride’s prior to this layer.

• output_score (boolean, optional, default=0) – Add score to outputs

• iou_loss (boolean, optional, default=0) – Usage of IoU Loss

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.ROIAlign(data=None, rois=None, pooled_size=_Null, spatial_scale=_Null, sample_ratio=_Null, position_sensitive=_Null, aligned=_Null, out=None, name=None, **kwargs)¶

This operator takes a 4D feature map as an input array and region proposals as rois, then align the feature map over sub-regions of input and produces a fixed-sized output array. This operator is typically used in Faster R-CNN & Mask R-CNN networks. If roi batchid is less than 0, it will be ignored, and the corresponding output will be set to 0.

Different from ROI pooling, ROI Align removes the harsh quantization, properly aligning the extracted features with the input. RoIAlign computes the value of each sampling point by bilinear interpolation from the nearby grid points on the feature map. No quantization is performed on any coordinates involved in the RoI, its bins, or the sampling points. Bilinear interpolation is used to compute the exact values of the input features at four regularly sampled locations in each RoI bin. Then the feature map can be aggregated by avgpooling.

References

He, Kaiming, et al. “Mask R-CNN.” ICCV, 2017

Defined in src/operator/contrib/roi_align.cc:L559

Parameters

• data (NDArray) – Input data to the pooling operator, a 4D Feature maps

• rois (NDArray) – Bounding box coordinates, a 2D array, if batchid is less than 0, it will be ignored.

• pooled_size (Shape(tuple), required) – ROI Align output roi feature map height and width: (h, w)

• spatial_scale (float, required) – Ratio of input feature map height (or w) to raw image height (or w). Equals the reciprocal of total stride in convolutional layers

• sample_ratio (int, optional, default='-1') – Optional sampling ratio of ROI align, using adaptive size by default.

• position_sensitive (boolean, optional, default=0) – Whether to perform position-sensitive RoI pooling. PSRoIPooling is first proposaled by R-FCN and it can reduce the input channels by ph*pw times, where (ph, pw) is the pooled_size

• aligned (boolean, optional, default=0) – Center-aligned ROIAlign introduced in Detectron2. To enable, set aligned to True.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.RROIAlign(data=None, rois=None, pooled_size=_Null, spatial_scale=_Null, sampling_ratio=_Null, out=None, name=None, **kwargs)¶

Performs Rotated ROI Align on the input array.

This operator takes a 4D feature map as an input array and region proposals as rois, then align the feature map over sub-regions of input and produces a fixed-sized output array.

Different from ROI Align, RROI Align uses rotated rois, which is suitable for text detection. RRoIAlign computes the value of each sampling point by bilinear interpolation from the nearby grid points on the rotated feature map. No quantization is performed on any coordinates involved in the RoI, its bins, or the sampling points. Bilinear interpolation is used to compute the exact values of the input features at four regularly sampled locations in each RoI bin. Then the feature map can be aggregated by avgpooling.

References

Ma, Jianqi, et al. “Arbitrary-Oriented Scene Text Detection via Rotation Proposals.” IEEE Transactions on Multimedia, 2018.

Defined in src/operator/contrib/rroi_align.cc:L274

Parameters

• data (NDArray) – Input data to the pooling operator, a 4D Feature maps

• rois (NDArray) – Bounding box coordinates, a 2D array

• pooled_size (Shape(tuple), required) – RROI align output shape (h,w)

• spatial_scale (float, required) – Ratio of input feature map height (or width) to raw image height (or width). Equals the reciprocal of total stride in convolutional layers

• sampling_ratio (int, optional, default='-1') – Optional sampling ratio of RROI align, using adaptive size by default.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.SparseEmbedding(data=None, weight=None, input_dim=_Null, output_dim=_Null, dtype=_Null, sparse_grad=_Null, out=None, name=None, **kwargs)¶

Maps integer indices to vector representations (embeddings).

note:: contrib.SparseEmbedding is deprecated, use Embedding instead.

This operator maps words to real-valued vectors in a high-dimensional space, called word embeddings. These embeddings can capture semantic and syntactic properties of the words. For example, it has been noted that in the learned embedding spaces, similar words tend to be close to each other and dissimilar words far apart.

For an input array of shape (d1, …, dK), the shape of an output array is (d1, …, dK, output_dim). All the input values should be integers in the range [0, input_dim).

If the input_dim is ip0 and output_dim is op0, then shape of the embedding weight matrix must be (ip0, op0).

The storage type of the gradient will be row_sparse.

Note

SparseEmbedding is designed for the use case where input_dim is very large (e.g. 100k). The operator is available on both CPU and GPU. When deterministic is set to True, the accumulation of gradients follows a deterministic order if a feature appears multiple times in the input. However, the accumulation is usually slower when the order is enforced on GPU. When the operator is used on the GPU, the recommended value for deterministic is True.

Examples:

input_dim = 4
output_dim = 5

// Each row in weight matrix y represents a word. So, y = (w0,w1,w2,w3)
y = [[  0.,   1.,   2.,   3.,   4.],
     [  5.,   6.,   7.,   8.,   9.],
     [ 10.,  11.,  12.,  13.,  14.],
     [ 15.,  16.,  17.,  18.,  19.]]

// Input array x represents n-grams(2-gram). So, x = [(w1,w3), (w0,w2)]
x = [[ 1.,  3.],
     [ 0.,  2.]]

// Mapped input x to its vector representation y.
SparseEmbedding(x, y, 4, 5) = [[[  5.,   6.,   7.,   8.,   9.],
                               [ 15.,  16.,  17.,  18.,  19.]],

                              [[  0.,   1.,   2.,   3.,   4.],
                               [ 10.,  11.,  12.,  13.,  14.]]]

Defined in src/operator/tensor/indexing_op.cc:L675

Parameters

• data (NDArray) – The input array to the embedding operator.

• weight (NDArray) – The embedding weight matrix.

• input_dim (int, required) – Vocabulary size of the input indices.

• output_dim (int, required) – Dimension of the embedding vectors.

• dtype ({'bfloat16', 'float16', 'float32', 'float64', 'int32', 'int64', 'int8', 'uint8'},optional, default='float32') – Data type of weight.

• sparse_grad (boolean, optional, default=0) – Compute row sparse gradient in the backward calculation. If set to True, the grad’s storage type is row_sparse.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.SyncBatchNorm(data=None, gamma=None, beta=None, moving_mean=None, moving_var=None, eps=_Null, momentum=_Null, fix_gamma=_Null, use_global_stats=_Null, output_mean_var=_Null, ndev=_Null, key=_Null, out=None, name=None, **kwargs)¶

Batch normalization.

Normalizes a data batch by mean and variance, and applies a scale gamma as well as offset beta. Standard BN 1 implementation only normalize the data within each device. SyncBN normalizes the input within the whole mini-batch. We follow the sync-onece implmentation described in the paper 2.

Assume the input has more than one dimension and we normalize along axis 1. We first compute the mean and variance along this axis:

\[\begin{split}data\_mean[i] = mean(data[:,i,:,...]) \\ data\_var[i] = var(data[:,i,:,...])\end{split}\]

Then compute the normalized output, which has the same shape as input, as following:

\[out[:,i,:,...] = \frac{data[:,i,:,...] - data\_mean[i]}{\sqrt{data\_var[i]+\epsilon}} * gamma[i] + beta[i]\]

Both mean and var returns a scalar by treating the input as a vector.

Assume the input has size k on axis 1, then both gamma and beta have shape (k,). If output_mean_var is set to be true, then outputs both data_mean and data_var as well, which are needed for the backward pass.

Besides the inputs and the outputs, this operator accepts two auxiliary states, moving_mean and moving_var, which are k-length vectors. They are global statistics for the whole dataset, which are updated by:

moving_mean = moving_mean * momentum + data_mean * (1 - momentum)
moving_var = moving_var * momentum + data_var * (1 - momentum)

If use_global_stats is set to be true, then moving_mean and moving_var are used instead of data_mean and data_var to compute the output. It is often used during inference.

Both gamma and beta are learnable parameters. But if fix_gamma is true, then set gamma to 1 and its gradient to 0.

Reference:

1

Ioffe, Sergey, and Christian Szegedy. “Batch normalization: Accelerating deep network training by reducing internal covariate shift.” ICML 2015

2

Hang Zhang, Kristin Dana, Jianping Shi, Zhongyue Zhang, Xiaogang Wang, Ambrish Tyagi, and Amit Agrawal. “Context Encoding for Semantic Segmentation.” CVPR 2018

Defined in src/operator/contrib/sync_batch_norm.cc:L97

Parameters

• data (NDArray) – Input data to batch normalization

• gamma (NDArray) – gamma array

• beta (NDArray) – beta array

• moving_mean (NDArray) – running mean of input

• moving_var (NDArray) – running variance of input

• eps (float, optional, default=0.00100000005) – Epsilon to prevent div 0

• momentum (float, optional, default=0.899999976) – Momentum for moving average

• fix_gamma (boolean, optional, default=1) – Fix gamma while training

• use_global_stats (boolean, optional, default=0) – Whether use global moving statistics instead of local batch-norm. This will force change batch-norm into a scale shift operator.

• output_mean_var (boolean, optional, default=0) – Output All,normal mean and var

• ndev (int, optional, default='1') – The count of GPU devices

• key (string, required) – Hash key for synchronization, please set the same hash key for same layer, Block.prefix is typically used as in gluon.nn.contrib.SyncBatchNorm.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.allclose(a=None, b=None, rtol=_Null, atol=_Null, equal_nan=_Null, out=None, name=None, **kwargs)¶

This operators implements the numpy.allclose(a, b, rtol=1e-05, atol=1e-08, equal_nan=False)

\[f(x) = |a−b|≤atol+rtol|b|\]

where \(a, b\) are the input tensors of equal types an shapes \(atol, rtol\) the values of absolute and relative tolerance (by default, rtol=1e-05, atol=1e-08)

Examples:

a = [1e10, 1e-7],
b = [1.00001e10, 1e-8]
y = allclose(a, b)
y = False

a = [1e10, 1e-8],
b = [1.00001e10, 1e-9]
y = allclose(a, b)
y = True

Defined in src/operator/contrib/allclose_op.cc:L55

Parameters

• a (NDArray) – Input array a

• b (NDArray) – Input array b

• rtol (float, optional, default=9.99999975e-06) – Relative tolerance.

• atol (float, optional, default=9.99999994e-09) – Absolute tolerance.

• equal_nan (boolean, optional, default=1) – Whether to compare NaN’s as equal. If True, NaN’s in A will be considered equal to NaN’s in B in the output array.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.arange_like(data=None, start=_Null, step=_Null, repeat=_Null, ctx=_Null, axis=_Null, out=None, name=None, **kwargs)¶

Return an array with evenly spaced values. If axis is not given, the output will have the same shape as the input array. Otherwise, the output will be a 1-D array with size of the specified axis in input shape.

Examples:

x = [[0.14883883 0.7772398  0.94865847 0.7225052 ]
     [0.23729339 0.6112595  0.66538996 0.5132841 ]
     [0.30822644 0.9912457  0.15502319 0.7043658 ]]
     <NDArray 3x4 @cpu(0)>

out = mx.nd.contrib.arange_like(x, start=0)

  [[ 0.  1.  2.  3.]
   [ 4.  5.  6.  7.]
   [ 8.  9. 10. 11.]]
   <NDArray 3x4 @cpu(0)>

out = mx.nd.contrib.arange_like(x, start=0, axis=-1)

  [0. 1. 2. 3.]
  <NDArray 4 @cpu(0)>

Parameters

• data (NDArray) – The input

• start (double, optional, default=0) – Start of interval. The interval includes this value. The default start value is 0.

• step (double, optional, default=1) – Spacing between values.

• repeat (int, optional, default='1') – The repeating time of all elements. E.g repeat=3, the element a will be repeated three times –> a, a, a.

• ctx (string, optional, default='') – Context of output, in format [cpu|gpu|cpu_pinned](n).Only used for imperative calls.

• axis (int or None, optional, default='None') – Arange elements according to the size of a certain axis of input array. The negative numbers are interpreted counting from the backward. If not provided, will arange elements according to the input shape.

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.backward_gradientmultiplier(data=None, scalar=_Null, is_int=_Null, out=None, name=None, **kwargs)¶

Parameters

• data (NDArray) – source input

• scalar (double, optional, default=1) – Scalar input value

• is_int (boolean, optional, default=1) – Indicate whether scalar input is int type

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.backward_hawkesll(out=None, name=None, **kwargs)¶

Parameters

out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.backward_index_copy(out=None, name=None, **kwargs)¶

Parameters

out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.backward_quadratic(out=None, name=None, **kwargs)¶

Parameters

out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays

mxnet.contrib.ndarray.bipartite_matching(data=None, is_ascend=_Null, threshold=_Null, topk=_Null, out=None, name=None, **kwargs)¶

Compute bipartite matching.

The matching is performed on score matrix with shape [B, N, M] - B: batch_size - N: number of rows to match - M: number of columns as reference to be matched against.

Returns: x : matched column indices. -1 indicating non-matched elements in rows. y : matched row indices.

Note:

Zero gradients are back-propagated in this op for now.

Example:

s = [[0.5, 0.6], [0.1, 0.2], [0.3, 0.4]]
x, y = bipartite_matching(x, threshold=1e-12, is_ascend=False)
x = [1, -1, 0]
y = [2, 0]

Defined in src/operator/contrib/bounding_box.cc:L183

Parameters

• data (NDArray) – The input

• is_ascend (boolean, optional, default=0) – Use ascend order for scores instead of descending. Please set threshold accordingly.

• threshold (float, required) – Ignore matching when score < thresh, if is_ascend=false, or ignore score > thresh, if is_ascend=true.

• topk (int, optional, default='-1') – Limit the number of matches to topk, set -1 for no limit

• out (NDArray, optional) – The output NDArray to hold the result.

Returns

out – The output of this function.

Return type

NDArray or list of NDArrays