Add IoU-loss for yolov3 (#192)

* add iou loss to yolov3
* modify the comment and delete redundant yml file
* fix the low speed bug in dropblock module

configs/dcn/yolov3_r50vd_dcn_iouloss_obj365_pretrained_coco.yml

+architecture: YOLOv3
+use_gpu: true
+max_iters: 55000
+log_smooth_window: 20
+save_dir: output
+snapshot_iter: 10000
+metric: COCO
+pretrain_weights: https://paddlemodels.bj.bcebos.com/object_detection/ResNet50_vd_obj365_pretrained.tar
+weights: output/yolov3_r50vd_dcn_iouloss_obj365_pretrained_coco/model_final
+num_classes: 80
+use_fine_grained_loss: true
+
+YOLOv3:
+  backbone: ResNet
+  yolo_head: YOLOv3Head
+  use_fine_grained_loss: true
+
+ResNet:
+  norm_type: sync_bn
+  freeze_at: 0
+  freeze_norm: false
+  norm_decay: 0.
+  depth: 50
+  feature_maps: [3, 4, 5]
+  variant: d
+  dcn_v2_stages: [5]
+
+YOLOv3Head:
+  anchor_masks: [[6, 7, 8], [3, 4, 5], [0, 1, 2]]
+  anchors: [[10, 13], [16, 30], [33, 23],
+            [30, 61], [62, 45], [59, 119],
+            [116, 90], [156, 198], [373, 326]]
+  norm_decay: 0.
+  yolo_loss: YOLOv3Loss
+  nms:
+    background_label: -1
+    keep_top_k: 100
+    nms_threshold: 0.45
+    nms_top_k: 1000
+    normalized: false
+    score_threshold: 0.01
+
+YOLOv3Loss:
+  batch_size: 8
+  ignore_thresh: 0.7
+  label_smooth: false
+  use_fine_grained_loss: true
+  iou_loss: IouLoss
+
+IouLoss:
+  loss_weight: 2.5
+  max_height: 608
+  max_width: 608
+
+LearningRate:
+  base_lr: 0.001
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 0.1
+    milestones:
+    - 40000
+    - 50000
+  - !LinearWarmup
+    start_factor: 0.
+    steps: 4000
+
+OptimizerBuilder:
+  optimizer:
+    momentum: 0.9
+    type: Momentum
+  regularizer:
+    factor: 0.0005
+    type: L2
+
+_READER_: '../yolov3_reader.yml'

ppdet/modeling/losses/__init__.py

@@ -18,8 +18,10 @@ from . import yolo_loss
 from . import smooth_l1_loss
 from . import giou_loss
 from . import diou_loss
+from . import iou_loss
 
 from .yolo_loss import *
 from .smooth_l1_loss import *
 from .giou_loss import *
 from .diou_loss import *
+from .iou_loss import *

ppdet/modeling/losses/iou_loss.py

+# Copyright (c) 2020 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.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import numpy as np
+from paddle.fluid.param_attr import ParamAttr
+from paddle.fluid.initializer import NumpyArrayInitializer
+
+from paddle import fluid
+from ppdet.core.workspace import register, serializable
+
+__all__ = ['IouLoss']
+
+
+@register
+@serializable
+class IouLoss(object):
+    """
+    iou loss, see https://arxiv.org/abs/1908.03851
+    loss = 1.0 - iou * iou
+    Args:
+        loss_weight (float): iou loss weight, default is 2.5
+        max_height (int): max height of input to support random shape input
+        max_width (int): max width of input to support random shape input
+    """
+    def __init__(self,
+                 loss_weight=2.5,
+                 max_height=608,
+                 max_width=608):
+        self._loss_weight = loss_weight
+        self._MAX_HI = max_height
+        self._MAX_WI = max_width
+
+    def __call__(self, x, y, w, h, tx, ty, tw, th,
+                 anchors, downsample_ratio, batch_size, eps=1.e-10):
+        '''
+        Args:
+            x  | y | w | h  ([Variables]): the output of yolov3 for encoded x|y|w|h
+            tx |ty |tw |th  ([Variables]): the target of yolov3 for encoded x|y|w|h
+            anchors ([float]): list of anchors for current output layer
+            downsample_ratio (float): the downsample ratio for current output layer
+            batch_size (int): training batch size
+            eps (float): the decimal to prevent the denominator eqaul zero
+        '''
+        x1, y1, x2, y2 = self._bbox_transform(x, y, w, h, anchors,
+            downsample_ratio, batch_size, False)
+        x1g, y1g, x2g, y2g = self._bbox_transform(tx, ty, tw, th,
+            anchors, downsample_ratio, batch_size, True)
+
+        x2 = fluid.layers.elementwise_max(x1, x2)
+        y2 = fluid.layers.elementwise_max(y1, y2)
+
+        xkis1 = fluid.layers.elementwise_max(x1, x1g)
+        ykis1 = fluid.layers.elementwise_max(y1, y1g)
+        xkis2 = fluid.layers.elementwise_min(x2, x2g)
+        ykis2 = fluid.layers.elementwise_min(y2, y2g)
+
+        xc1 = fluid.layers.elementwise_min(x1, x1g)
+        yc1 = fluid.layers.elementwise_min(y1, y1g)
+        xc2 = fluid.layers.elementwise_max(x2, x2g)
+        yc2 = fluid.layers.elementwise_max(y2, y2g)
+
+        intsctk = (xkis2 - xkis1) * (ykis2 - ykis1)
+        intsctk = intsctk * fluid.layers.greater_than(
+            xkis2, xkis1) * fluid.layers.greater_than(ykis2, ykis1)
+        unionk = (x2 - x1) * (y2 - y1) + (x2g - x1g) * (y2g - y1g) - intsctk + eps
+        iouk = intsctk / unionk
+        loss_iou = 1. - iouk * iouk
+        loss_iou = loss_iou * self._loss_weight
+
+        return loss_iou
+
+    def _bbox_transform(self, dcx, dcy, dw, dh, anchors, downsample_ratio, batch_size, is_gt):
+        grid_x = int(self._MAX_WI / downsample_ratio)
+        grid_y = int(self._MAX_HI / downsample_ratio)
+        an_num = len(anchors) // 2
+
+        shape_fmp = fluid.layers.shape(dcx)
+        shape_fmp.stop_gradient = True
+        # generate the grid_w x grid_h center of feature map
+        idx_i = np.array([[i for i in range(grid_x)]])
+        idx_j = np.array([[j for j in range(grid_y)]]).transpose()
+        gi_np = np.repeat(idx_i, grid_y, axis=0)
+        gi_np = np.reshape(gi_np, newshape=[1, 1, grid_y, grid_x])
+        gi_np = np.tile(gi_np, reps=[batch_size, an_num, 1, 1])
+        gj_np = np.repeat(idx_j, grid_x, axis=1)
+        gj_np = np.reshape(gj_np, newshape=[1, 1, grid_y, grid_x])
+        gj_np = np.tile(gj_np, reps=[batch_size, an_num, 1, 1])
+        gi_max = self._create_tensor_from_numpy(gi_np.astype(np.float32))
+        gi = fluid.layers.crop(x=gi_max, shape=dcx)
+        gi.stop_gradient = True
+        gj_max = self._create_tensor_from_numpy(gj_np.astype(np.float32))
+        gj = fluid.layers.crop(x=gj_max, shape=dcx)
+        gj.stop_gradient = True
+
+        grid_x_act = fluid.layers.cast(shape_fmp[3], dtype="float32")
+        grid_x_act.stop_gradient = True
+        grid_y_act = fluid.layers.cast(shape_fmp[2], dtype="float32")
+        grid_y_act.stop_gradient = True
+        if is_gt:
+            cx = fluid.layers.elementwise_add(dcx, gi) / grid_x_act
+            cx.gradient = True
+            cy = fluid.layers.elementwise_add(dcy, gj) / grid_y_act
+            cy.gradient = True
+        else:
+            dcx_sig = fluid.layers.sigmoid(dcx)
+            cx = fluid.layers.elementwise_add(dcx_sig, gi) / grid_x_act
+            dcy_sig = fluid.layers.sigmoid(dcy)
+            cy = fluid.layers.elementwise_add(dcy_sig, gj) / grid_y_act
+
+        anchor_w_ = [anchors[i] for i in range(0, len(anchors)) if i % 2 == 0]
+        anchor_w_np = np.array(anchor_w_)
+        anchor_w_np = np.reshape(anchor_w_np, newshape=[1, an_num, 1, 1])
+        anchor_w_np = np.tile(anchor_w_np, reps=[batch_size, 1, grid_y, grid_x])
+        anchor_w_max = self._create_tensor_from_numpy(anchor_w_np.astype(np.float32))
+        anchor_w = fluid.layers.crop(x=anchor_w_max, shape=dcx)
+        anchor_w.stop_gradient = True
+        anchor_h_ = [anchors[i] for i in range(0, len(anchors)) if i % 2 == 1]
+        anchor_h_np = np.array(anchor_h_)
+        anchor_h_np = np.reshape(anchor_h_np, newshape=[1, an_num, 1, 1])
+        anchor_h_np = np.tile(anchor_h_np, reps=[batch_size, 1, grid_y, grid_x])
+        anchor_h_max = self._create_tensor_from_numpy(anchor_h_np.astype(np.float32))
+        anchor_h = fluid.layers.crop(x=anchor_h_max, shape=dcx)
+        anchor_h.stop_gradient = True
+        # e^tw e^th
+        exp_dw = fluid.layers.exp(dw)
+        exp_dh = fluid.layers.exp(dh)
+        pw = fluid.layers.elementwise_mul(exp_dw, anchor_w) / \
+            (grid_x_act * downsample_ratio)
+        ph = fluid.layers.elementwise_mul(exp_dh, anchor_h) / \
+            (grid_y_act * downsample_ratio)
+        if is_gt:
+            exp_dw.stop_gradient = True
+            exp_dh.stop_gradient = True
+            pw.stop_gradient = True
+            ph.stop_gradient = True
+        
+
+        x1 = cx - 0.5 * pw
+        y1 = cy - 0.5 * ph
+        x2 = cx + 0.5 * pw
+        y2 = cy + 0.5 * ph
+        if is_gt:
+            x1.stop_gradient = True
+            y1.stop_gradient = True
+            x2.stop_gradient = True
+            y2.stop_gradient = True
+
+        return x1, y1, x2, y2
+
+    def _create_tensor_from_numpy(self, numpy_array):
+        paddle_array = fluid.layers.create_parameter(
+            attr=ParamAttr(),
+            shape=numpy_array.shape,
+            dtype=numpy_array.dtype,
+            default_initializer=NumpyArrayInitializer(numpy_array))
+        paddle_array.stop_gradient = True
+        return paddle_array
+

ppdet/modeling/losses/yolo_loss.py

@@ -34,17 +34,20 @@ class YOLOv3Loss(object):
         use_fine_grained_loss (bool): whether use fine grained YOLOv3 loss
                                       instead of fluid.layers.yolov3_loss
     """
+    __inject__ = ['iou_loss']
     __shared__ = ['use_fine_grained_loss']
 
     def __init__(self,
                  batch_size=8,
                  ignore_thresh=0.7,
                  label_smooth=True,
-                 use_fine_grained_loss=False):
+                 use_fine_grained_loss=False,
+                 iou_loss=None):
         self._batch_size = batch_size
         self._ignore_thresh = ignore_thresh
         self._label_smooth = label_smooth
         self._use_fine_grained_loss = use_fine_grained_loss
+        self._iou_loss = iou_loss
 
     def __call__(self, outputs, gt_box, gt_label, gt_score, targets, anchors,
                  anchor_masks, mask_anchors, num_classes, prefix_name):
@@ -104,7 +107,10 @@ class YOLOv3Loss(object):
             "YOLOv3 output layer number not equal target number"
 
         downsample = 32
+        if self._iou_loss is None:
             loss_xys, loss_whs, loss_objs, loss_clss = [], [], [], []
+        else:
+            loss_xys, loss_whs, loss_ious, loss_objs, loss_clss = [], [], [], [], []
         for i, (output, target,
                 anchors) in enumerate(zip(outputs, targets, mask_anchors)):
             an_num = len(anchors) // 2
@@ -124,6 +130,12 @@ class YOLOv3Loss(object):
             loss_w = fluid.layers.reduce_sum(loss_w, dim=[1, 2, 3])
             loss_h = fluid.layers.abs(h - th) * tscale_tobj
             loss_h = fluid.layers.reduce_sum(loss_h, dim=[1, 2, 3])
+            if self._iou_loss is not None:
+                loss_iou = self._iou_loss(x, y, w, h, tx, ty, tw, th,
+                    anchors, downsample, self._batch_size)
+                loss_iou = loss_iou * tscale_tobj
+                loss_iou = fluid.layers.reduce_sum(loss_iou, dim=[1, 2, 3])
+                loss_ious.append(fluid.layers.reduce_mean(loss_iou))
 
             loss_obj_pos, loss_obj_neg = self._calc_obj_loss(
                 output, obj, tobj, gt_box, self._batch_size, anchors,
@@ -140,13 +152,15 @@ class YOLOv3Loss(object):
             loss_clss.append(fluid.layers.reduce_mean(loss_cls))
 
             downsample //= 2
-
-        return {
+        losses_all = {
             "loss_xy": fluid.layers.sum(loss_xys),
             "loss_wh": fluid.layers.sum(loss_whs),
             "loss_obj": fluid.layers.sum(loss_objs),
             "loss_cls": fluid.layers.sum(loss_clss),
         }
+        if self._iou_loss is not None:
+            losses_all["loss_iou"] = fluid.layers.sum(loss_ious)
+        return losses_all
 
     def _split_output(self, output, an_num, num_classes):
         """

ppdet/modeling/ops.py

@@ -155,14 +155,12 @@ def DropBlock(input, block_size, keep_prob, is_test):
     mask = 1.0 - mask_flag
 
     elem_numel = fluid.layers.reduce_prod(input_shape)
-    elem_numel = fluid.layers.cast(elem_numel, dtype="float32")
-    elem_numel_tmp = fluid.layers.reshape(elem_numel, [1, 1, 1, 1])
-    elem_numel_m = fluid.layers.expand_as(elem_numel_tmp, input)
+    elem_numel_m = fluid.layers.cast(elem_numel, dtype="float32")
+    elem_numel_m.stop_gradient = True
 
     elem_sum = fluid.layers.reduce_sum(mask)
-    elem_sum_tmp = fluid.layers.cast(elem_sum, dtype="float32")
-    elem_sum_tmp = fluid.layers.reshape(elem_sum_tmp, [1, 1, 1, 1])
-    elem_sum_m = fluid.layers.expand_as(elem_sum_tmp, input)
+    elem_sum_m = fluid.layers.cast(elem_sum, dtype="float32")
+    elem_sum_m.stop_gradient = True
 
     output = input * mask * elem_numel_m / elem_sum_m
     return output