Add generate_mask_labels_op to support Mask-RCNN and refine some code. (#15371)

* Add generate_mask_labels_op to support Mask-RCNN.
* Refine sigmoid_cross_entropy to support nomalize mode.
* Fix generator_proposals_label.
* Use DeviceTemporaryAllocator in roi_pool and roi_algin.
* Remove shape check in data_feeder.

Author: qingqing01

paddle/fluid/API.spec

@@ -197,7 +197,7 @@ paddle.fluid.layers.clip ArgSpec(args=['x', 'min', 'max', 'name'], varargs=None,
 paddle.fluid.layers.clip_by_norm ArgSpec(args=['x', 'max_norm', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.mean ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.mul ArgSpec(args=['x', 'y', 'x_num_col_dims', 'y_num_col_dims', 'name'], varargs=None, keywords=None, defaults=(1, 1, None))
-paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'ignore_index', 'name'], varargs=None, keywords=None, defaults=(-100, None))
+paddle.fluid.layers.sigmoid_cross_entropy_with_logits ArgSpec(args=['x', 'label', 'ignore_index', 'name', 'normalize'], varargs=None, keywords=None, defaults=(-100, None, False))
 paddle.fluid.layers.maxout ArgSpec(args=['x', 'groups', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.space_to_depth ArgSpec(args=['x', 'blocksize', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.affine_grid ArgSpec(args=['theta', 'out_shape', 'name'], varargs=None, keywords=None, defaults=(None,))
@@ -318,6 +318,7 @@ paddle.fluid.layers.anchor_generator ArgSpec(args=['input', 'anchor_sizes', 'asp
 paddle.fluid.layers.roi_perspective_transform ArgSpec(args=['input', 'rois', 'transformed_height', 'transformed_width', 'spatial_scale'], varargs=None, keywords=None, defaults=(1.0,))
 paddle.fluid.layers.generate_proposal_labels ArgSpec(args=['rpn_rois', 'gt_classes', 'is_crowd', 'gt_boxes', 'im_info', 'batch_size_per_im', 'fg_fraction', 'fg_thresh', 'bg_thresh_hi', 'bg_thresh_lo', 'bbox_reg_weights', 'class_nums', 'use_random'], varargs=None, keywords=None, defaults=(256, 0.25, 0.25, 0.5, 0.0, [0.1, 0.1, 0.2, 0.2], None, True))
 paddle.fluid.layers.generate_proposals ArgSpec(args=['scores', 'bbox_deltas', 'im_info', 'anchors', 'variances', 'pre_nms_top_n', 'post_nms_top_n', 'nms_thresh', 'min_size', 'eta', 'name'], varargs=None, keywords=None, defaults=(6000, 1000, 0.5, 0.1, 1.0, None))
+paddle.fluid.layers.generate_mask_labels ArgSpec(args=['im_info', 'gt_classes', 'is_crowd', 'gt_segms', 'rois', 'labels_int32', 'num_classes', 'resolution'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.layers.iou_similarity ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.layers.box_coder ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name'], varargs=None, keywords=None, defaults=('encode_center_size', True, None))
 paddle.fluid.layers.polygon_box_transform ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,))

paddle/fluid/operators/affine_channel_op.cu

@@ -83,7 +83,7 @@ __global__ void AffineChannelScaleBiasGradientCUDAKernel(
     T* dbias) {
   const int outer_size = C;
   const int inner_size = N * HxW;
-  typedef cub::BlockReduce<T, BlockDim> BlockReduce;
+  typedef cub::BlockReduce<double, BlockDim> BlockReduce;
   __shared__ typename BlockReduce::TempStorage ds_storage;
   __shared__ typename BlockReduce::TempStorage db_storage;
 
@@ -97,13 +97,16 @@ __global__ void AffineChannelScaleBiasGradientCUDAKernel(
       ds_sum += dy[index] * x[index];
       db_sum += dy[index];
     }
-    ds_sum = BlockReduce(ds_storage).Reduce(ds_sum, cub::Sum());
-    db_sum = BlockReduce(db_storage).Reduce(db_sum, cub::Sum());
+    __syncthreads();
+    auto ds_out =
+        BlockReduce(ds_storage).Reduce(static_cast<double>(ds_sum), cub::Sum());
+    auto db_out =
+        BlockReduce(db_storage).Reduce(static_cast<double>(db_sum), cub::Sum());
+    __syncthreads();
     if (threadIdx.x == 0) {
-      dscale[i] = ds_sum;
-      dbias[i] = db_sum;
+      dscale[i] = ds_out;
+      dbias[i] = db_out;
     }
-    __syncthreads();
   }
 }
 

paddle/fluid/operators/detection/CMakeLists.txt

@@ -45,3 +45,7 @@ detection_library(roi_perspective_transform_op SRCS roi_perspective_transform_op
 foreach(src ${LOCAL_DETECTION_LIBS})
     set(OP_LIBRARY ${src} ${OP_LIBRARY} CACHE INTERNAL "op libs")
 endforeach()
+
+cc_library(mask_util SRCS mask_util.cc DEPS memory)
+cc_test(mask_util_test SRCS mask_util_test.cc DEPS memory mask_util)
+detection_library(generate_mask_labels_op SRCS generate_mask_labels_op.cc DEPS mask_util)

paddle/fluid/operators/detection/bbox_util.h

@@ -1,13 +1,17 @@
 /* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+
 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
 You may obtain a copy of the License at
+
     http://www.apache.org/licenses/LICENSE-2.0
+
 Unless required by applicable law or agreed to in writing, software
 distributed under the License is distributed on an "AS IS" BASIS,
 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
+
 #pragma once
 #include <algorithm>
 #include "paddle/fluid/framework/eigen.h"
@@ -88,7 +92,9 @@ void BboxOverlaps(const framework::Tensor& r_boxes,
       inter_w = std::max(x_max - x_min + 1, zero);
       inter_h = std::max(y_max - y_min + 1, zero);
       inter_area = inter_w * inter_h;
-      overlaps_et(i, j) = inter_area / (r_box_area + c_box_area - inter_area);
+      overlaps_et(i, j) =
+          (inter_area == 0.) ? 0 : inter_area /
+                                       (r_box_area + c_box_area - inter_area);
     }
   }
 }