diff --git a/third_party/fwkacllib/inc/ops/linalg_ops.h b/third_party/fwkacllib/inc/ops/linalg_ops.h index 1a59d88e..a334570c 100644 --- a/third_party/fwkacllib/inc/ops/linalg_ops.h +++ b/third_party/fwkacllib/inc/ops/linalg_ops.h @@ -82,6 +82,25 @@ REG_OP(Cholesky) DT_FLOAT16, DT_COMPLEX64, DT_COMPLEX128})) .OP_END_FACTORY_REG(Cholesky) +/** +*@brief Computes the outer product of two 1D vectors . \n + +*@par Inputs: +*The input x and vec2 has to be a 1D vector.Inputs include: +*@li x:A Tensor. Must be one of the following types: float16, float32. +Shape is [N] . \n +*@li vec2:A Tensor. Must have the same type as x. Shape is [M] . \n + +*@par Outputs: +*y:A Tensor. Has the same type as x . \n +*/ + +REG_OP(Ger) + .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(vec2, TensorType({DT_FLOAT16, DT_FLOAT})) + .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) + .OP_END_FACTORY_REG(Ger) + /** *@brief Computes the sign and the log of the absolute value of the determinant of one or more square matrices . \n diff --git a/third_party/fwkacllib/inc/ops/nn_detect_ops.h b/third_party/fwkacllib/inc/ops/nn_detect_ops.h index af59b4e2..33148e62 100644 --- a/third_party/fwkacllib/inc/ops/nn_detect_ops.h +++ b/third_party/fwkacllib/inc/ops/nn_detect_ops.h @@ -254,22 +254,22 @@ is min_size/sqrt(aspect_ratio), the width is min_size*sqrt(aspect_ratio). Defaul *@par Third-party framework compatibility * It is a custom operator. It has no corresponding operator in Caffe. */ - REG_OP(PriorBox) - .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) - .INPUT(img, TensorType({DT_FLOAT16, DT_FLOAT})) - .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) - .REQUIRED_ATTR(min_size, ListFloat) - .REQUIRED_ATTR(max_size, ListFloat) - .REQUIRED_ATTR(aspect_ratio, ListFloat) - .ATTR(img_h, Int, 0) - .ATTR(img_w, Int, 0) - .ATTR(step_h, Float, 0.0) - .ATTR(step_w, Float, 0.0) - .ATTR(flip, Bool, true) - .ATTR(clip, Bool, false) - .ATTR(offset, Float, 0.5) - .ATTR(variance, ListFloat, {0.1}) - .OP_END_FACTORY_REG(PriorBox); +REG_OP(PriorBox) + .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(img, TensorType({DT_FLOAT16, DT_FLOAT})) + .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) + .REQUIRED_ATTR(min_size, ListFloat) + .REQUIRED_ATTR(max_size, ListFloat) + .REQUIRED_ATTR(aspect_ratio, ListFloat) + .ATTR(img_h, Int, 0) + .ATTR(img_w, Int, 0) + .ATTR(step_h, Float, 0.0) + .ATTR(step_w, Float, 0.0) + .ATTR(flip, Bool, true) + .ATTR(clip, Bool, false) + .ATTR(offset, Float, 0.5) + .ATTR(variance, ListFloat, {0.1}) + .OP_END_FACTORY_REG(PriorBox); /** *@brief Performs SSD prior box detection, with four additional matrices and the "aspect_ratio" attribute deleted compared to PriorBox . \n @@ -306,25 +306,25 @@ is min_size/sqrt(aspect_ratio), the width is min_size*sqrt(aspect_ratio). Defaul *@par Restrictions: *Warning: THIS FUNCTION IS DEPRECATED. Please use PriorBox instead. */ - REG_OP(PriorBoxD) - .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) - .INPUT(img, TensorType({DT_FLOAT16, DT_FLOAT})) - .INPUT(data_h, TensorType({DT_FLOAT16, DT_FLOAT})) - .INPUT(data_w, TensorType({DT_FLOAT16, DT_FLOAT})) - .INPUT(box_height, TensorType({DT_FLOAT16, DT_FLOAT})) - .INPUT(box_width, TensorType({DT_FLOAT16, DT_FLOAT})) - .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) - .REQUIRED_ATTR(min_size, ListFloat) - .REQUIRED_ATTR(max_size, ListFloat) - .ATTR(img_h, Int, 0) - .ATTR(img_w, Int, 0) - .ATTR(step_h, Float, 0.0) - .ATTR(step_w, Float, 0.0) - .ATTR(flip, Bool, true) - .ATTR(clip, Bool, false) - .ATTR(offset, Float, 0.5) - .ATTR(variance, ListFloat, {0.1}) - .OP_END_FACTORY_REG(PriorBoxD); +REG_OP(PriorBoxD) + .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(img, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(data_h, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(data_w, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(box_height, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(box_width, TensorType({DT_FLOAT16, DT_FLOAT})) + .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) + .REQUIRED_ATTR(min_size, ListFloat) + .REQUIRED_ATTR(max_size, ListFloat) + .ATTR(img_h, Int, 0) + .ATTR(img_w, Int, 0) + .ATTR(step_h, Float, 0.0) + .ATTR(step_w, Float, 0.0) + .ATTR(flip, Bool, true) + .ATTR(clip, Bool, false) + .ATTR(offset, Float, 0.5) + .ATTR(variance, ListFloat, {0.1}) + .OP_END_FACTORY_REG(PriorBoxD); /** *@brief Performs SSD prior box detection, with four additional matrices and the "aspect_ratio" attribute deleted compared to PriorBox . \n @@ -358,22 +358,22 @@ is min_size/sqrt(aspect_ratio), the width is min_size*sqrt(aspect_ratio). Defaul *@par Restrictions: *Warning: THIS FUNCTION IS DEPRECATED. Please use PriorBox instead. */ - REG_OP(PriorBoxDV2) - .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) - .INPUT(img, TensorType({DT_FLOAT16, DT_FLOAT})) - .INPUT(boxes, TensorType({DT_FLOAT16, DT_FLOAT})) - .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) - .REQUIRED_ATTR(min_size, ListFloat) - .REQUIRED_ATTR(max_size, ListFloat) - .ATTR(img_h, Int, 0) - .ATTR(img_w, Int, 0) - .ATTR(step_h, Float, 0.0) - .ATTR(step_w, Float, 0.0) - .ATTR(flip, Bool, true) - .ATTR(clip, Bool, false) - .ATTR(offset, Float, 0.5) - .ATTR(variance, ListFloat, {0.1}) - .OP_END_FACTORY_REG(PriorBoxDV2); +REG_OP(PriorBoxDV2) + .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(img, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(boxes, TensorType({DT_FLOAT16, DT_FLOAT})) + .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) + .REQUIRED_ATTR(min_size, ListFloat) + .REQUIRED_ATTR(max_size, ListFloat) + .ATTR(img_h, Int, 0) + .ATTR(img_w, Int, 0) + .ATTR(step_h, Float, 0.0) + .ATTR(step_w, Float, 0.0) + .ATTR(flip, Bool, true) + .ATTR(clip, Bool, false) + .ATTR(offset, Float, 0.5) + .ATTR(variance, ListFloat, {0.1}) + .OP_END_FACTORY_REG(PriorBoxDV2); /** *@brief Performs Position Sensitive ROI Pooling . \n @@ -531,10 +531,10 @@ as xx...xyy...yww...whh...hbb...bc0c0..c0c1c1...c1......cncn...cn . \n * It is a custom operator. It has no corresponding operator in Caffe. */ REG_OP(Yolo) - .INPUT(x, TensorType({DT_FLOAT16,DT_FLOAT})) - .OUTPUT(coord_data, TensorType({DT_FLOAT16,DT_FLOAT})) - .OUTPUT(obj_prob, TensorType({DT_FLOAT16,DT_FLOAT})) - .OUTPUT(classes_prob, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) + .OUTPUT(coord_data, TensorType({DT_FLOAT16, DT_FLOAT})) + .OUTPUT(obj_prob, TensorType({DT_FLOAT16, DT_FLOAT})) + .OUTPUT(classes_prob, TensorType({DT_FLOAT16, DT_FLOAT})) .ATTR(boxes, Int, 3) .ATTR(coords, Int, 4) .ATTR(classes, Int, 80) @@ -584,10 +584,10 @@ REG_OP(Yolo) * It is a custom operator. It has no corresponding operator in Caffe. */ REG_OP(YoloV2DetectionOutput) - .INPUT(coord_data, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(obj_prob, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(classes_prob, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(img_info, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(coord_data, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(obj_prob, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(classes_prob, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(img_info, TensorType({DT_FLOAT16, DT_FLOAT})) .REQUIRED_ATTR(biases, ListFloat) .ATTR(boxes, Int, 5) .ATTR(coords, Int, 4) @@ -598,7 +598,7 @@ REG_OP(YoloV2DetectionOutput) .ATTR(score_threshold, Float, 0.5) .ATTR(iou_threshold, Float, 0.45) .ATTR(pre_nms_topn, Int, 512) - .OUTPUT(box_out, TensorType({DT_FLOAT16,DT_FLOAT})) + .OUTPUT(box_out, TensorType({DT_FLOAT16, DT_FLOAT})) .OUTPUT(box_out_num, TensorType({DT_INT32})) .OP_END_FACTORY_REG(YoloV2DetectionOutput) @@ -647,12 +647,12 @@ REG_OP(YoloV2DetectionOutput) *Warning: THIS FUNCTION IS DEPRECATED. Please use YoloV2DetectionOutput instead. */ REG_OP(YoloV2DetectionOutputD) - .INPUT(coord_data, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(obj_prob, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(classes_prob, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(img_info, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(windex, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(hindex, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(coord_data, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(obj_prob, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(classes_prob, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(img_info, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(windex, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(hindex, TensorType({DT_FLOAT16, DT_FLOAT})) .REQUIRED_ATTR(biases, ListFloat) .ATTR(boxes, Int, 5) .ATTR(coords, Int, 4) @@ -663,7 +663,7 @@ REG_OP(YoloV2DetectionOutputD) .ATTR(score_threshold, Float, 0.5) .ATTR(iou_threshold, Float, 0.45) .ATTR(pre_nms_topn, Int, 512) - .OUTPUT(box_out, TensorType({DT_FLOAT16,DT_FLOAT})) + .OUTPUT(box_out, TensorType({DT_FLOAT16, DT_FLOAT})) .OUTPUT(box_out_num, TensorType({DT_INT32})) .OP_END_FACTORY_REG(YoloV2DetectionOutputD) @@ -707,16 +707,16 @@ REG_OP(YoloV2DetectionOutputD) * It is a custom operator. It has no corresponding operator in Caffe. */ REG_OP(YoloV3DetectionOutput) - .INPUT(coord_data_low, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(coord_data_mid, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(coord_data_high, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(obj_prob_low, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(obj_prob_mid, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(obj_prob_high, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(classes_prob_low, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(classes_prob_mid, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(classes_prob_high, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(img_info, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(coord_data_low, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(coord_data_mid, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(coord_data_high, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(obj_prob_low, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(obj_prob_mid, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(obj_prob_high, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(classes_prob_low, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(classes_prob_mid, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(classes_prob_high, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(img_info, TensorType({DT_FLOAT16, DT_FLOAT})) .REQUIRED_ATTR(biases_low, ListFloat) .REQUIRED_ATTR(biases_mid, ListFloat) .REQUIRED_ATTR(biases_high, ListFloat) @@ -729,7 +729,7 @@ REG_OP(YoloV3DetectionOutput) .ATTR(score_threshold, Float, 0.5) .ATTR(iou_threshold, Float, 0.45) .ATTR(pre_nms_topn, Int, 512) - .OUTPUT(box_out, TensorType({DT_FLOAT16,DT_FLOAT})) + .OUTPUT(box_out, TensorType({DT_FLOAT16, DT_FLOAT})) .OUTPUT(box_out_num, TensorType({DT_INT32})) .OP_END_FACTORY_REG(YoloV3DetectionOutput) @@ -776,22 +776,22 @@ s *Warning: THIS FUNCTION IS DEPRECATED. Please use YoloV3DetectionOutput instead. */ REG_OP(YoloV3DetectionOutputD) - .INPUT(coord_data_low, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(coord_data_mid, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(coord_data_high, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(obj_prob_low, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(obj_prob_mid, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(obj_prob_high, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(classes_prob_low, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(classes_prob_mid, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(classes_prob_high, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(img_info, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(windex1, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(windex2, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(windex3, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(hindex1, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(hindex2, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(hindex3, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(coord_data_low, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(coord_data_mid, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(coord_data_high, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(obj_prob_low, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(obj_prob_mid, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(obj_prob_high, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(classes_prob_low, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(classes_prob_mid, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(classes_prob_high, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(img_info, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(windex1, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(windex2, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(windex3, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(hindex1, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(hindex2, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(hindex3, TensorType({DT_FLOAT16, DT_FLOAT})) .REQUIRED_ATTR(biases_low, ListFloat) .REQUIRED_ATTR(biases_mid, ListFloat) .REQUIRED_ATTR(biases_high, ListFloat) @@ -804,7 +804,7 @@ REG_OP(YoloV3DetectionOutputD) .ATTR(score_threshold, Float, 0.5) .ATTR(iou_threshold, Float, 0.45) .ATTR(pre_nms_topn, Int, 512) - .OUTPUT(box_out, TensorType({DT_FLOAT16,DT_FLOAT})) + .OUTPUT(box_out, TensorType({DT_FLOAT16, DT_FLOAT})) .OUTPUT(box_out_num, TensorType({DT_INT32})) .OP_END_FACTORY_REG(YoloV3DetectionOutputD) @@ -848,7 +848,7 @@ There are three Yolo operators at Yolov3DetectionOutput's preceding layer on Yol * It is a custom operator. It has no corresponding operator in Caffe. */ REG_OP(YoloV3DetectionOutputV2) - .DYNAMIC_INPUT(x, TensorType({DT_FLOAT16,DT_FLOAT})) + .DYNAMIC_INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) .REQUIRED_ATTR(biases, ListFloat) .ATTR(boxes, Int, 3) .ATTR(coords, Int, 4) @@ -862,7 +862,7 @@ REG_OP(YoloV3DetectionOutputV2) .ATTR(N, Int, 10) .ATTR(resize_origin_img_to_net, Bool, false) .ATTR(out_box_dim, Int, 3) - .OUTPUT(box_out, TensorType({DT_FLOAT16,DT_FLOAT})) + .OUTPUT(box_out, TensorType({DT_FLOAT16, DT_FLOAT})) .OUTPUT(box_out_num, TensorType({DT_INT32})) .OP_END_FACTORY_REG(YoloV3DetectionOutputV2) @@ -910,9 +910,9 @@ REG_OP(YoloV3DetectionOutputV2) * Warning: THIS FUNCTION IS DEPRECATED. Please use YoloV3DetectionOutputV2 instead. */ REG_OP(YoloV3DetectionOutputV2D) - .DYNAMIC_INPUT(x, TensorType({DT_FLOAT16,DT_FLOAT})) - .DYNAMIC_INPUT(windex, TensorType({DT_FLOAT16,DT_FLOAT})) - .DYNAMIC_INPUT(hindex, TensorType({DT_FLOAT16,DT_FLOAT})) + .DYNAMIC_INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) + .DYNAMIC_INPUT(windex, TensorType({DT_FLOAT16, DT_FLOAT})) + .DYNAMIC_INPUT(hindex, TensorType({DT_FLOAT16, DT_FLOAT})) .REQUIRED_ATTR(biases, ListFloat) .ATTR(boxes, Int, 3) .ATTR(coords, Int, 4) @@ -926,7 +926,7 @@ REG_OP(YoloV3DetectionOutputV2D) .ATTR(N, Int, 10) .ATTR(resize_origin_img_to_net, Bool, false) .ATTR(out_box_dim, Int, 3) - .OUTPUT(box_out, TensorType({DT_FLOAT16,DT_FLOAT})) + .OUTPUT(box_out, TensorType({DT_FLOAT16, DT_FLOAT})) .OUTPUT(box_out_num, TensorType({DT_INT32})) .OP_END_FACTORY_REG(YoloV3DetectionOutputV2D) @@ -1466,9 +1466,9 @@ REG_OP(NormalizeBBox) * y: A Tensor. Must have the same type as box_predictions. */ REG_OP(DecodeBboxV2) - .INPUT(boxes, TensorType({DT_FLOAT16,DT_FLOAT})) - .INPUT(anchors, TensorType({DT_FLOAT16,DT_FLOAT})) - .OUTPUT(y, TensorType({DT_FLOAT16,DT_FLOAT})) + .INPUT(boxes, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(anchors, TensorType({DT_FLOAT16, DT_FLOAT})) + .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) .ATTR(scales, ListFloat, {1.0, 1.0, 1.0, 1.0}) .ATTR(decode_clip, Float, 0.0) .ATTR(reversed_box, Bool, false) @@ -1592,7 +1592,6 @@ selected indices from the boxes tensor, where M <= max_output_size. \n *Compatible with onnx NonMaxSuppression operator. */ - REG_OP(NonMaxSuppressionV7) .INPUT(boxes, TensorType({DT_FLOAT16, DT_FLOAT})) .INPUT(scores, TensorType({DT_FLOAT16, DT_FLOAT})) @@ -1641,7 +1640,7 @@ REG_OP(RoiExtractor) .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) .ATTR(finest_scale, Int, 56) .ATTR(roi_scale_factor, Float, 0) - .ATTR(spatial_scale, ListFloat, { 1.f/4, 1.f/8, 1.f/16, 1.f/32 }) + .ATTR(spatial_scale, ListFloat, {1.f / 4, 1.f / 8, 1.f / 16, 1.f / 32}) .ATTR(pooled_height, Int, 7) .ATTR(pooled_width, Int, 7) .ATTR(sample_num, Int, 0) @@ -1649,6 +1648,84 @@ REG_OP(RoiExtractor) .ATTR(aligned, Bool, true) .OP_END_FACTORY_REG(RoiExtractor) +/** +*@brief Performs Position Sensitive PS ROI Pooling . \n + +*@par Inputs: +* Two inputs, including: +*@li x: An NC1HWC0 tensor of type float16 or float32, describing the feature +* map, dimension C1 must be equal to +* (int(output_dim+15)/C0))*group_size*group_size. +*@li rois: A tensor of type float16 or float32, with shape +* [batch, 5, rois_num], describing the ROIs, each ROI consists of five +* elements: "batch_id", "x1", "y1", "x2", and "y2", which "batch_id" indicates +* the index of the input feature map, "x1", "y1", "x2", or "y2" must be +* greater than or equal to "0.0" . \n + +*@par Attributes: +*@li output_dim: A required int32, specifying the number of output channels, +* must be greater than 0. +*@li group_size: A required int32, specifying the number of groups to encode +* position-sensitive score maps, must be within the range (0, 128). +*@li spatial_scale: A required float32, scaling factor for mapping the input +* coordinates to the ROI coordinates . \n + +*@par Outputs: +*y: An NC1HWC0 tensor of type float16 or float32, describing the result +* feature map . \n + +*@attention Constraints: +* HC1HWC0: channel must be Group_size squared, rois_num is a multiple of 16 +*/ +REG_OP(PSROIPoolingV2) + .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(rois, TensorType({DT_FLOAT16, DT_FLOAT})) + .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) + .REQUIRED_ATTR(spatial_scale, Float) + .REQUIRED_ATTR(output_dim, Int) + .REQUIRED_ATTR(group_size, Int) + .OP_END_FACTORY_REG(PSROIPoolingV2) + +/** +*@brief Performs Position Sensitive PS ROI Pooling Grad . \n + +*@par Inputs: +* Two inputs, including: +*@li x: An NC1HWC0 tensor of type float16 or float32, describing the result +* feature map . \n +*@li rois: A tensor of type float16 or float32, with shape +* [batch, 5, rois_num], describing the ROIs, each ROI consists of five +* elements: "batch_id", "x1", "y1", "x2", and "y2", which "batch_id" indicates +* the index of the input feature map, "x1", "y1", "x2", or "y2" must be +* greater than or equal to "0.0" . \n + +*@par Attributes: +*@li output_dim: A required int32, specifying the number of output channels, +* must be greater than 0. +*@li group_size: A required int32, specifying the number of groups to encode +* position-sensitive score maps, must be within the range (0, 128). +*@li spatial_scale: A required float32, scaling factor for mapping the input +* coordinates to the ROI coordinates . \n +*@li input_size: A required listInt, mapping the gradinput size: (H, W) + +*@par Outputs: +*y: An NC1HWC0 tensor of type float16 or float32, describing the feature +* map, dimension C1 must be equal to +* (int(output_dim+15)/C0))*group_size*group_size. + +*@attention Constraints: +* HC1HWC0: channel must be Group_size squared, rois_num is a multiple of 16 +*/ +REG_OP(PSROIPoolingGradV2D) + .INPUT(x, TensorType({DT_FLOAT16, DT_FLOAT})) + .INPUT(rois, TensorType({DT_FLOAT16, DT_FLOAT})) + .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT})) + .REQUIRED_ATTR(spatial_scale, Float) + .REQUIRED_ATTR(output_dim, Int) + .REQUIRED_ATTR(group_size, Int) + .REQUIRED_ATTR(input_size, ListInt) + .OP_END_FACTORY_REG(PSROIPoolingGradV2D) + } // namespace ge #endif // OPS_BUILT_IN_OP_PROTO_INC_NN_DETECT_OPS_H_ diff --git a/third_party/fwkacllib/inc/ops/nn_norm_ops.h b/third_party/fwkacllib/inc/ops/nn_norm_ops.h index 7d38beb5..2bc87ec1 100644 --- a/third_party/fwkacllib/inc/ops/nn_norm_ops.h +++ b/third_party/fwkacllib/inc/ops/nn_norm_ops.h @@ -525,6 +525,31 @@ REG_OP(LayerNorm) .ATTR(epsilon, Float, 0.0000001) .OP_END_FACTORY_REG(LayerNorm) +/** +*@brief Returns a tensor where each sub-tensor of input along dimension +* dim is normalized such that the p-norm of the sub-tensor is lower than the value maxnorm. \n + +*@par Inputs: +*One input, including: +* @li x: A Tensor. Must be one of the following types: float16, float32 . \n + +*@par Attributes: +* @li p: Specify L_p norm, the type is float. +* @li dim: The processed dim, the type is int. +* @li maxnorm: Threshold for comparison, the type is float. \n + +*@par Outputs: +*One outputs, including: +* @li y: shape and dtype of output, should be same shape and type as input. +*/ +REG_OP(Renorm) + .INPUT(x, TensorType::BasicType()) + .OUTPUT(y, TensorType::BasicType()) + .REQUIRED_ATTR(p, Float) + .REQUIRED_ATTR(dim, Int) + .REQUIRED_ATTR(maxnorm, Float) + .OP_END_FACTORY_REG(Renorm) + /** *@brief LayerNormGrad operator interface implementation * calculating: dy, x, variance, mean, gamma diff --git a/third_party/fwkacllib/inc/ops/nn_pooling_ops.h b/third_party/fwkacllib/inc/ops/nn_pooling_ops.h index 32eb148b..8251fac3 100644 --- a/third_party/fwkacllib/inc/ops/nn_pooling_ops.h +++ b/third_party/fwkacllib/inc/ops/nn_pooling_ops.h @@ -397,8 +397,8 @@ No default value. specifying the stride of the sliding window for each dimension of the input tensor. No default value. *@li padding: A required string type of float16. -*@li pads: A list type of int32. Default value {0, 0, 0}. -*@li dilation: A list type of int32. Default value {1, 1, 1}. +*@li pads: A list type of int32. Default value {0,0,0,0,0,0}. +*@li dilation: A list type of int32. Default value {1,1,1,1,1,1}. *@li ceil_mode: A ceil mode number of int32 . Default value 0. *@li data_format: An optional string. Defaults to "NDHWC" . \n @@ -421,8 +421,8 @@ REG_OP(MaxPool3D) .REQUIRED_ATTR(ksize, ListInt) .REQUIRED_ATTR(strides, ListInt) .REQUIRED_ATTR(padding, String) - .ATTR(pads, ListInt, {0,0,0}) - .ATTR(dilation, ListInt, {1,1,1}) + .ATTR(pads, ListInt, {0,0,0,0,0,0}) + .ATTR(dilation, ListInt, {1,1,1,1,1,1}) .ATTR(ceil_mode, Int, 0) .ATTR(data_format, String, "NDHWC") .OP_END_FACTORY_REG(MaxPool3D) diff --git a/third_party/fwkacllib/inc/ops/pad_ops.h b/third_party/fwkacllib/inc/ops/pad_ops.h index 50eb98c8..42da3828 100644 --- a/third_party/fwkacllib/inc/ops/pad_ops.h +++ b/third_party/fwkacllib/inc/ops/pad_ops.h @@ -418,7 +418,7 @@ REG_OP(EmbeddingRankId) */ REG_OP(FillV2) .INPUT(dims, TensorType({DT_INT16, DT_INT32, DT_INT64})) - .OUTPUT(y, TensorType({DT_FLOAT, DT_FLOAT, DT_DOUBLE, DT_INT8, DT_INT16, DT_INT32, DT_INT64})) + .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT, DT_DOUBLE, DT_INT8, DT_INT16, DT_INT32, DT_INT64})) .ATTR(value, Float, 0) .OP_END_FACTORY_REG(FillV2) @@ -437,7 +437,7 @@ REG_OP(FillV2) * Compatible with the ONNX operator ConstantOfShape. */ REG_OP(FillV2D) - .OUTPUT(y, TensorType({DT_FLOAT, DT_FLOAT, DT_DOUBLE, DT_INT8, DT_UINT8, DT_INT16, DT_INT32, DT_INT64})) + .OUTPUT(y, TensorType({DT_FLOAT16, DT_FLOAT, DT_DOUBLE, DT_INT8, DT_UINT8, DT_INT16, DT_INT32, DT_INT64})) .ATTR(value, Float, 0) .REQUIRED_ATTR(dims, ListInt) .OP_END_FACTORY_REG(FillV2D)