test=dygraph add cascade rcnn/mask (#1013)

add cascade rcnn and modify make net method
add new cfg method 
add auto debug tool

ppdet/modeling/__init__.py

 from . import ops
-from . import anchor
+from . import bbox
+from . import mask
 from . import backbone
 from . import head
 from . import architecture
 
 from .ops import *
-from .anchor import *
+from .bbox import *
+from .mask import *
 from .backbone import *
 from .head import *
 from .architecture import *

ppdet/modeling/architecture/__init__.py

@@ -9,8 +9,10 @@ from . import meta_arch
 from . import faster_rcnn
 from . import mask_rcnn
 from . import yolo
+from . import cascade_rcnn
 
 from .meta_arch import *
 from .faster_rcnn import *
 from .mask_rcnn import *
 from .yolo import *
+from .cascade_rcnn import *

ppdet/modeling/architecture/cascade_rcnn.py

+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from paddle import fluid
+from ppdet.core.workspace import register
+from .meta_arch import BaseArch
+
+__all__ = ['CascadeRCNN']
+
+
+@register
+class CascadeRCNN(BaseArch):
+    __category__ = 'architecture'
+    __shared__ = ['num_stages']
+    __inject__ = [
+        'anchor',
+        'proposal',
+        'mask',
+        'backbone',
+        'rpn_head',
+        'bbox_head',
+        'mask_head',
+    ]
+
+    def __init__(self,
+                 anchor,
+                 proposal,
+                 mask,
+                 backbone,
+                 rpn_head,
+                 bbox_head,
+                 mask_head,
+                 num_stages=3,
+                 *args,
+                 **kwargs):
+        super(CascadeRCNN, self).__init__(*args, **kwargs)
+        self.anchor = anchor
+        self.proposal = proposal
+        self.mask = mask
+        self.backbone = backbone
+        self.rpn_head = rpn_head
+        self.bbox_head = bbox_head
+        self.mask_head = mask_head
+        self.num_stages = num_stages
+
+    def model_arch(self, ):
+        # Backbone
+        bb_out = self.backbone(self.gbd)
+        self.gbd.update(bb_out)
+
+        # RPN
+        rpn_head_out = self.rpn_head(self.gbd)
+        self.gbd.update(rpn_head_out)
+
+        # Anchor
+        anchor_out = self.anchor(self.gbd)
+        self.gbd.update(anchor_out)
+
+        self.gbd['stage'] = 0
+        for i in range(self.num_stages):
+            self.gbd.update_v('stage', i)
+            # Proposal BBox
+            proposal_out = self.proposal(self.gbd)
+            self.gbd.update({"proposal_" + str(i): proposal_out})
+
+            # BBox Head
+            bbox_head_out = self.bbox_head(self.gbd)
+            self.gbd.update({'bbox_head_' + str(i): bbox_head_out})
+
+            refine_bbox_out = self.proposal.refine_bbox(self.gbd)
+            self.gbd['proposal_' + str(i)].update(refine_bbox_out)
+
+        if self.gbd['mode'] == 'infer':
+            bbox_out = self.proposal.post_process(self.gbd)
+            self.gbd.update(bbox_out)
+
+        # Mask 
+        mask_out = self.mask(self.gbd)
+        self.gbd.update(mask_out)
+
+        # Mask Head 
+        mask_head_out = self.mask_head(self.gbd)
+        self.gbd.update(mask_head_out)
+
+        if self.gbd['mode'] == 'infer':
+            mask_out = self.mask.post_process(self.gbd)
+            self.gbd.update(mask_out)
+
+    def loss(self, ):
+        outs = {}
+        losses = []
+
+        rpn_cls_loss, rpn_reg_loss = self.rpn_head.loss(self.gbd)
+        outs['loss_rpn_cls'] = rpn_cls_loss
+        outs['loss_rpn_reg'] = rpn_reg_loss
+        losses.extend([rpn_cls_loss, rpn_reg_loss])
+
+        bbox_cls_loss_list = []
+        bbox_reg_loss_list = []
+        for i in range(self.num_stages):
+            self.gbd.update_v('stage', i)
+            bbox_cls_loss, bbox_reg_loss = self.bbox_head.loss(self.gbd)
+            bbox_cls_loss_list.append(bbox_cls_loss)
+            bbox_reg_loss_list.append(bbox_reg_loss)
+            outs['loss_bbox_cls_' + str(i)] = bbox_cls_loss
+            outs['loss_bbox_reg_' + str(i)] = bbox_reg_loss
+        losses.extend(bbox_cls_loss_list)
+        losses.extend(bbox_reg_loss_list)
+
+        mask_loss = self.mask_head.loss(self.gbd)
+        outs['mask_loss'] = mask_loss
+        losses.append(mask_loss)
+
+        loss = fluid.layers.sum(losses)
+        outs['loss'] = loss
+        return outs
+
+    def infer(self, ):
+        outs = {
+            'bbox': self.gbd['predicted_bbox'].numpy(),
+            'bbox_nums': self.gbd['predicted_bbox_nums'].numpy(),
+            'mask': self.gbd['predicted_mask'].numpy(),
+            'im_id': self.gbd['im_id'].numpy(),
+            'im_shape': self.gbd['im_shape'].numpy()
+        }
+        return inputs

ppdet/modeling/architecture/faster_rcnn.py

@@ -3,7 +3,6 @@ from __future__ import division
 from __future__ import print_function
 
 from paddle import fluid
-
 from ppdet.core.workspace import register
 from .meta_arch import BaseArch
 
@@ -21,27 +20,16 @@ class FasterRCNN(BaseArch):
         'bbox_head',
     ]
 
-    def __init__(self,
-                 anchor,
-                 proposal,
-                 backbone,
-                 rpn_head,
-                 bbox_head,
-                 rpn_only=False,
-                 mode='train'):
-        super(FasterRCNN, self).__init__()
+    def __init__(self, anchor, proposal, backbone, rpn_head, bbox_head, *args,
+                 **kwargs):
+        super(FasterRCNN, self).__init__(*args, **kwargs)
         self.anchor = anchor
         self.proposal = proposal
         self.backbone = backbone
         self.rpn_head = rpn_head
         self.bbox_head = bbox_head
-        self.rpn_only = rpn_only
-        self.mode = mode
-
-    def forward(self, inputs, inputs_keys):
-        self.gbd = self.build_inputs(inputs, inputs_keys)
-        self.gbd['mode'] = self.mode
 
+    def model_arch(self, ):
         # Backbone
         bb_out = self.backbone(self.gbd)
         self.gbd.update(bb_out)
@@ -55,29 +43,21 @@ class FasterRCNN(BaseArch):
         self.gbd.update(anchor_out)
 
         # Proposal BBox
+        self.gbd['stage'] = 0
         proposal_out = self.proposal(self.gbd)
-        self.gbd.update(proposal_out)
+        self.gbd.update({'proposal_0': proposal_out})
 
         # BBox Head
-        bbox_head_out = self.bbox_head(self.gbd)
-        self.gbd.update(bbox_head_out)
+        bboxhead_out = self.bbox_head(self.gbd)
+        self.gbd.update({'bbox_head_0': bboxhead_out})
 
         if self.gbd['mode'] == 'infer':
             bbox_out = self.proposal.post_process(self.gbd)
             self.gbd.update(bbox_out)
 
-        # result  
-        if self.gbd['mode'] == 'train':
-            return self.loss(self.gbd)
-        elif self.gbd['mode'] == 'infer':
-            return self.infer(self.gbd)
-        else:
-            raise "Now, only support train or infer mode!"
-
-    def loss(self, inputs):
-        losses = []
-        rpn_cls_loss, rpn_reg_loss = self.rpn_head.loss(inputs)
-        bbox_cls_loss, bbox_reg_loss = self.bbox_head.loss(inputs)
+    def loss(self, ):
+        rpn_cls_loss, rpn_reg_loss = self.rpn_head.loss(self.gbd)
+        bbox_cls_loss, bbox_reg_loss = self.bbox_head.loss(self.gbd)
         losses = [rpn_cls_loss, rpn_reg_loss, bbox_cls_loss, bbox_reg_loss]
         loss = fluid.layers.sum(losses)
         out = {
@@ -89,11 +69,11 @@ class FasterRCNN(BaseArch):
         }
         return out
 
-    def infer(self, inputs):
+    def infer(self, ):
         outs = {
-            "bbox": inputs['predicted_bbox'].numpy(),
-            "bbox_nums": inputs['predicted_bbox_nums'].numpy(),
-            'im_id': inputs['im_id'].numpy(),
-            'im_shape': inputs['im_shape'].numpy()
+            "bbox": self.gbd['predicted_bbox'].numpy(),
+            "bbox_nums": self.gbd['predicted_bbox_nums'].numpy(),
+            'im_id': self.gbd['im_id'].numpy(),
+            'im_shape': self.gbd['im_shape'].numpy()
         }
         return outs

ppdet/modeling/architecture/mask_rcnn.py

@@ -3,11 +3,10 @@ from __future__ import division
 from __future__ import print_function
 
 from paddle import fluid
-
 from ppdet.core.workspace import register
 from ppdet.utils.data_structure import BufferDict
-
 from .meta_arch import BaseArch
+
 __all__ = ['MaskRCNN']
 
 
@@ -24,18 +23,9 @@ class MaskRCNN(BaseArch):
         'mask_head',
     ]
 
-    def __init__(self,
-                 anchor,
-                 proposal,
-                 mask,
-                 backbone,
-                 rpn_head,
-                 bbox_head,
-                 mask_head,
-                 rpn_only=False,
-                 mode='train'):
-        super(MaskRCNN, self).__init__()
-
+    def __init__(self, anchor, proposal, mask, backbone, rpn_head, bbox_head,
+                 mask_head, *args, **kwargs):
+        super(MaskRCNN, self).__init__(*args, **kwargs)
         self.anchor = anchor
         self.proposal = proposal
         self.mask = mask
@@ -43,12 +33,8 @@ class MaskRCNN(BaseArch):
         self.rpn_head = rpn_head
         self.bbox_head = bbox_head
         self.mask_head = mask_head
-        self.mode = mode
-
-    def forward(self, inputs, inputs_keys):
-        self.gbd = self.build_inputs(inputs, inputs_keys)
-        self.gbd['mode'] = mode
 
+    def model_arch(self, ):
         # Backbone
         bb_out = self.backbone(self.gbd)
         self.gbd.update(bb_out)
@@ -62,12 +48,13 @@ class MaskRCNN(BaseArch):
         self.gbd.update(anchor_out)
 
         # Proposal BBox
+        self.gbd['stage'] = 0
         proposal_out = self.proposal(self.gbd)
-        self.gbd.update(proposal_out)
+        self.gbd.update({'proposal_0': proposal_out})
 
         # BBox Head
-        bbox_head_out = self.bbox_head(self.gbd)
-        self.gbd.update(bbox_head_out)
+        bboxhead_out = self.bbox_head(self.gbd)
+        self.gbd.update({'bbox_head_0': bboxhead_out})
 
         if self.gbd['mode'] == 'infer':
             bbox_out = self.proposal.post_process(self.gbd)
@@ -85,19 +72,11 @@ class MaskRCNN(BaseArch):
             mask_out = self.mask.post_process(self.gbd)
             self.gbd.update(mask_out)
 
-        # result  
-        if self.gbd['mode'] == 'train':
-            return self.loss(self.gbd)
-        elif self.gbd['mode'] == 'infer':
-            self.infer(self.gbd)
-        else:
-            raise "Now, only support train or infer mode!"
-
-    def loss(self, inputs):
+    def loss(self, ):
         losses = []
-        rpn_cls_loss, rpn_reg_loss = self.rpn_head.loss(inputs)
-        bbox_cls_loss, bbox_reg_loss = self.bbox_head.loss(inputs)
-        mask_loss = self.mask_head.loss(inputs)
+        rpn_cls_loss, rpn_reg_loss = self.rpn_head.loss(self.gbd)
+        bbox_cls_loss, bbox_reg_loss = self.bbox_head.loss(self.gbd)
+        mask_loss = self.mask_head.loss(self.gbd)
         losses = [
             rpn_cls_loss, rpn_reg_loss, bbox_cls_loss, bbox_reg_loss, mask_loss
         ]
@@ -112,12 +91,12 @@ class MaskRCNN(BaseArch):
         }
         return out
 
-    def infer(self, inputs):
+    def infer(self, ):
         outs = {
-            'bbox': inputs['predicted_bbox'].numpy(),
-            'bbox_nums': inputs['predicted_bbox_nums'].numpy(),
-            'mask': inputs['predicted_mask'].numpy(),
-            'im_id': inputs['im_id'].numpy(),
-            'im_shape': inputs['im_shape'].numpy()
+            'bbox': self.gbd['predicted_bbox'].numpy(),
+            'bbox_nums': self.gbd['predicted_bbox_nums'].numpy(),
+            'mask': self.gbd['predicted_mask'].numpy(),
+            'im_id': self.gbd['im_id'].numpy(),
+            'im_shape': self.gbd['im_shape'].numpy()
         }
         return inputs

ppdet/modeling/architecture/meta_arch.py

@@ -3,11 +3,8 @@ from __future__ import division
 from __future__ import print_function
 
 import numpy as np
-
-from paddle import fluid
 from paddle.fluid.dygraph import Layer
 from paddle.fluid.dygraph.base import to_variable
-
 from ppdet.core.workspace import register
 from ppdet.utils.data_structure import BufferDict
 
@@ -16,22 +13,43 @@ __all__ = ['BaseArch']
 
 @register
 class BaseArch(Layer):
-    def __init__(self, mode='train', *args, **kwargs):
+    def __init__(self, *args, **kwargs):
         super(BaseArch, self).__init__()
-        self.mode = mode
+        self.args = args
+        self.kwargs = kwargs
 
-    def forward(self, inputs, inputs_keys, mode='train'):
-        raise NotImplementedError("Should implement forward method!")
+    def forward(self, inputs, inputs_keys):
+        self.gbd = BufferDict()
+        self.gbd.update(self.kwargs)
+        assert self.gbd[
+            'mode'] is not None, "Please specify mode train or infer in config file!"
+        if self.kwargs['open_debug'] is None:
+            self.gbd['open_debug'] = False
 
-    def loss(self, inputs):
-        raise NotImplementedError("Should implement loss method!")
+        self.build_inputs(inputs, inputs_keys)
 
-    def infer(self, inputs):
-        raise NotImplementedError("Should implement infer method!")
+        self.model_arch()
+
+        self.gbd.debug()
+
+        if self.gbd['mode'] == 'train':
+            out = self.loss()
+        elif self.gbd['mode'] == 'infer':
+            out = self.infer()
+        else:
+            raise "Now, only support train or infer mode!"
+        return out
 
     def build_inputs(self, inputs, inputs_keys):
-        gbd = BufferDict()
         for i, k in enumerate(inputs_keys):
             v = to_variable(np.array([x[i] for x in inputs]))
-            gbd.set(k, v)
-        return gbd
+            self.gbd.set(k, v)
+
+    def model_arch(self, ):
+        raise NotImplementedError("Should implement model_arch method!")
+
+    def loss(self, ):
+        raise NotImplementedError("Should implement loss method!")
+
+    def infer(self, ):
+        raise NotImplementedError("Should implement infer method!")

ppdet/modeling/architecture/yolo.py

@@ -2,8 +2,6 @@ from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
 
-from paddle import fluid
-
 from ppdet.core.workspace import register
 from .meta_arch import BaseArch
 
@@ -19,17 +17,13 @@ class YOLOv3(BaseArch):
         'yolo_head',
     ]
 
-    def __init__(self, anchor, backbone, yolo_head, mode='train'):
-        super(YOLOv3, self).__init__()
+    def __init__(self, anchor, backbone, yolo_head, *args, **kwargs):
+        super(YOLOv3, self).__init__(*args, **kwargs)
         self.anchor = anchor
         self.backbone = backbone
         self.yolo_head = yolo_head
-        self.mode = mode
-
-    def forward(self, inputs, inputs_keys):
-        self.gbd = self.build_inputs(inputs, inputs_keys)
-        self.gbd['mode'] = self.mode
 
+    def model_arch(self, ):
         # Backbone
         bb_out = self.backbone(self.gbd)
         self.gbd.update(bb_out)
@@ -46,23 +40,14 @@ class YOLOv3(BaseArch):
             bbox_out = self.anchor.post_process(self.gbd)
             self.gbd.update(bbox_out)
 
-        # result  
-        if self.gbd['mode'] == 'train':
-            return self.loss(self.gbd)
-        elif self.gbd['mode'] == 'infer':
-            return self.infer(self.gbd)
-        else:
-            raise "Now, only support train or infer mode!"
-
-    def loss(self, inputs):
-        yolo_loss = self.yolo_head.loss(inputs)
-        out = {'loss': yolo_loss, }
+    def loss(self, ):
+        yolo_loss = self.yolo_head.loss(self.gbd)
+        out = {'loss': yolo_loss}
         return out
 
-    def infer(self, inputs):
+    def infer(self, ):
         outs = {
-            "bbox": inputs['predicted_bbox'].numpy(),
-            "bbox_nums": inputs['predicted_bbox_nums']
+            "bbox": self.gbd['predicted_bbox'].numpy(),
+            "bbox_nums": self.gbd['predicted_bbox_nums']
         }
-        print(outs['bbox_nums'])
         return outs

ppdet/modeling/backbone/resnet.py

@@ -4,7 +4,6 @@ from paddle.fluid.dygraph import Layer
 from paddle.fluid.dygraph import Conv2D, Pool2D, BatchNorm
 from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.initializer import Constant
-
 from ppdet.core.workspace import register, serializable
 
 
@@ -17,10 +16,10 @@ class ConvBNLayer(Layer):
                  stride,
                  padding,
                  act='relu',
-                 learning_rate=1.0):
+                 lr=1.0):
         super(ConvBNLayer, self).__init__()
 
-        self._conv = Conv2D(
+        self.conv = Conv2D(
             num_channels=ch_in,
             num_filters=ch_out,
             filter_size=filter_size,
@@ -29,26 +28,23 @@ class ConvBNLayer(Layer):
             groups=1,
             act=act,
             param_attr=ParamAttr(
-                name=name_scope + "_weights", learning_rate=learning_rate),
+                name=name_scope + "_weights", learning_rate=lr),
             bias_attr=ParamAttr(name=name_scope + "_bias"))
-
         if name_scope == "conv1":
             bn_name = "bn_" + name_scope
         else:
             bn_name = "bn" + name_scope[3:]
-
-        self._bn = BatchNorm(
+        self.bn = BatchNorm(
             num_channels=ch_out,
             act=act,
             param_attr=ParamAttr(name=bn_name + '_scale'),
-            bias_attr=ParamAttr(bn_name + '_offset'),
+            bias_attr=ParamAttr(name=bn_name + '_offset'),
             moving_mean_name=bn_name + '_mean',
-            moving_variance_name=bn_name + '_variance',
-            is_test=True)
+            moving_variance_name=bn_name + '_variance')
 
     def forward(self, inputs):
-        x = self._conv(inputs)
-        out = self._bn(x)
+        out = self.conv(inputs)
+        out = self.bn(out)
         return out
 
 
@@ -60,11 +56,11 @@ class ConvAffineLayer(Layer):
                  filter_size,
                  stride,
                  padding,
-                 learning_rate=1.0,
+                 lr=1.0,
                  act='relu'):
         super(ConvAffineLayer, self).__init__()
 
-        self._conv = Conv2D(
+        self.conv = Conv2D(
             num_channels=ch_in,
             num_filters=ch_out,
             filter_size=filter_size,
@@ -72,36 +68,34 @@ class ConvAffineLayer(Layer):
             padding=padding,
             act=None,
             param_attr=ParamAttr(
-                name=name_scope + "_weights", learning_rate=learning_rate),
+                name=name_scope + "_weights", learning_rate=lr),
             bias_attr=False)
-
         if name_scope == "conv1":
             bn_name = "bn_" + name_scope
         else:
             bn_name = "bn" + name_scope[3:]
-        self.name_scope = name_scope
-
-        self.scale = fluid.Layer.create_parameter(
+        self.scale = fluid.layers.create_parameter(
             shape=[ch_out],
             dtype='float32',
             attr=ParamAttr(
                 name=bn_name + '_scale', learning_rate=0.),
             default_initializer=Constant(1.))
-        self.bias = fluid.layers.create_parameter(
+
+        self.offset = fluid.layers.create_parameter(
             shape=[ch_out],
             dtype='float32',
             attr=ParamAttr(
-                bn_name + '_offset', learning_rate=0.),
+                name=bn_name + '_offset', learning_rate=0.),
             default_initializer=Constant(0.))
 
         self.act = act
 
     def forward(self, inputs):
-        conv = self._conv(inputs)
+        out = self.conv(inputs)
         out = fluid.layers.affine_channel(
-            x=conv, scale=self.scale, bias=self.bias)
+            out, scale=self.scale, bias=self.offset)
         if self.act == 'relu':
-            out = fluid.layers.relu(x=out)
+            out = fluid.layers.relu(out)
         return out
 
 
@@ -112,12 +106,21 @@ class BottleNeck(Layer):
                  ch_out,
                  stride,
                  shortcut=True,
-                 learning_rate=1.0):
+                 lr=1.0,
+                 norm_type='bn'):
         super(BottleNeck, self).__init__()
+        self.name_scope = name_scope
+        if norm_type == 'bn':
+            atom_block = ConvBNLayer
+        elif norm_type == 'affine':
+            atom_block = ConvAffineLayer
+        else:
+            atom_block = None
+        assert atom_block != None, 'NormType only support BatchNorm and Affine!'
 
         self.shortcut = shortcut
         if not shortcut:
-            self.short = ConvBNLayer(
+            self.branch1 = atom_block(
                 name_scope + "_branch1",
                 ch_in=ch_in,
                 ch_out=ch_out * 4,
@@ -125,52 +128,48 @@ class BottleNeck(Layer):
                 stride=stride,
                 padding=0,
                 act=None,
-                learning_rate=learning_rate)
+                lr=lr)
 
-        self.conv1 = ConvBNLayer(
+        self.branch2a = atom_block(
             name_scope + "_branch2a",
             ch_in=ch_in,
             ch_out=ch_out,
             filter_size=1,
             stride=stride,
             padding=0,
-            learning_rate=learning_rate, )
+            lr=lr)
 
-        self.conv2 = ConvBNLayer(
+        self.branch2b = atom_block(
             name_scope + "_branch2b",
             ch_in=ch_out,
             ch_out=ch_out,
             filter_size=3,
             stride=1,
             padding=1,
-            learning_rate=learning_rate)
+            lr=lr)
 
-        self.conv3 = ConvBNLayer(
+        self.branch2c = atom_block(
             name_scope + "_branch2c",
             ch_in=ch_out,
             ch_out=ch_out * 4,
             filter_size=1,
             stride=1,
             padding=0,
-            learning_rate=learning_rate,
+            lr=lr,
             act=None)
-        self.name_scope = name_scope
 
     def forward(self, inputs):
         if self.shortcut:
             short = inputs
         else:
-            short = self.short(inputs)
+            short = self.branch1(inputs)
 
-        conv1 = self.conv1(inputs)
-        conv2 = self.conv2(conv1)
-        conv3 = self.conv3(conv2)
+        out = self.branch2a(inputs)
+        out = self.branch2b(out)
+        out = self.branch2c(out)
 
         out = fluid.layers.elementwise_add(
-            x=short,
-            y=conv3,
-            act='relu',
-            name=self.name_scope + ".add.output.5")
+            x=short, y=out, act='relu', name=self.name_scope + ".add.output.5")
 
         return out
 
@@ -182,7 +181,8 @@ class Blocks(Layer):
                  ch_out,
                  count,
                  stride,
-                 learning_rate=1.0):
+                 lr=1.0,
+                 norm_type='bn'):
         super(Blocks, self).__init__()
 
         self.blocks = []
@@ -204,7 +204,8 @@ class Blocks(Layer):
                     ch_out=ch_out,
                     stride=self.stride,
                     shortcut=self.shortcut,
-                    learning_rate=learning_rate))
+                    lr=lr,
+                    norm_type=norm_type))
             self.blocks.append(block)
             shortcut = True
 
@@ -215,62 +216,75 @@ class Blocks(Layer):
         return res_out
 
 
+ResNet_cfg = {'50': [3, 4, 6, 3], '101': [3, 4, 23, 3], '152': [3, 8, 36, 3]}
+
+
 @register
 @serializable
 class ResNet(Layer):
-    def __init__(
-            self,
-            norm_type='bn',
-            depth=50,
-            feature_maps=4,
-            freeze_at=2, ):
+    def __init__(self, depth=50, norm_type='bn', freeze_at='res2'):
         super(ResNet, self).__init__()
+        self.depth = depth
+        self.norm_type = norm_type
+        self.freeze_at = freeze_at
+
+        block_nums = ResNet_cfg[str(self.depth)]
+        if self.norm_type == 'bn':
+            atom_block = ConvBNLayer
+        elif self.norm_type == 'affine':
+            atom_block = ConvAffineLayer
+        else:
+            atom_block = None
+        assert atom_block != None, 'NormType only support BatchNorm and Affine!'
 
-        if depth == 50:
-            blocks = [3, 4, 6, 3]
-        elif depth == 101:
-            blocks = [3, 4, 23, 3]
-        elif depth == 152:
-            blocks = [3, 8, 36, 3]
-
-        self.conv = ConvBNLayer(
-            "conv1",
-            ch_in=3,
-            ch_out=64,
-            filter_size=7,
-            stride=2,
-            padding=3,
-            learning_rate=0.)
+        self.conv1 = atom_block(
+            'conv1', ch_in=3, ch_out=64, filter_size=7, stride=2, padding=3)
 
-        self.pool2d_max = Pool2D(
+        self.pool = Pool2D(
             pool_type='max', pool_size=3, pool_stride=2, pool_padding=1)
 
         self.stage2 = Blocks(
             "res2",
             ch_in=64,
             ch_out=64,
-            count=blocks[0],
+            count=block_nums[0],
             stride=1,
-            learning_rate=0.)
+            norm_type=norm_type)
 
         self.stage3 = Blocks(
-            "res3", ch_in=256, ch_out=128, count=blocks[1], stride=2)
+            "res3",
+            ch_in=256,
+            ch_out=128,
+            count=block_nums[1],
+            stride=2,
+            norm_type=norm_type)
 
         self.stage4 = Blocks(
-            "res4", ch_in=512, ch_out=256, count=blocks[2], stride=2)
+            "res4",
+            ch_in=512,
+            ch_out=256,
+            count=block_nums[2],
+            stride=2,
+            norm_type=norm_type)
 
     def forward(self, inputs):
         x = inputs['image']
 
-        conv1 = self.conv(x)
-        poo1 = self.pool2d_max(conv1)
+        conv1 = self.conv1(x)
+
+        pool1 = self.pool(conv1)
 
-        res2 = self.stage2(poo1)
-        res2.stop_gradient = True
+        res2 = self.stage2(pool1)
 
         res3 = self.stage3(res2)
 
         res4 = self.stage4(res3)
 
-        outs = {'res2': res2, 'res3': res3, 'res4': res4}
+        outs = {
+            'res2': res2,
+            'res3': res3,
+            'res4': res4,
+            'res_norm_type': self.norm_type
+        }
+        outs[self.freeze_at].stop_gradient = True
         return outs

ppdet/modeling/anchor.py → ppdet/modeling/bbox.py

 import numpy as np
 import paddle.fluid as fluid
-
 from ppdet.core.workspace import register
-from ppdet.modeling.ops import (
-    AnchorGeneratorYOLO, AnchorTargetGeneratorYOLO, AnchorGeneratorRPN,
-    AnchorTargetGeneratorRPN, ProposalGenerator, ProposalTargetGenerator,
-    MaskTargetGenerator, DecodeClipNms, YOLOBox, MultiClassNMS)
-# TODO: modify here into ppdet.modeling.ops like DecodeClipNms 
-from ppdet.py_op.post_process import mask_post_process
 
 
 @register
 class BBoxPostProcess(object):
+    __shared__ = ['num_classes', 'num_stages']
+    __inject__ = ['decode_clip_nms']
+
     def __init__(self,
+                 decode_clip_nms,
+                 num_classes=81,
+                 num_stages=1,
                  decode=None,
                  clip=None,
-                 nms=None,
-                 decode_clip_nms=DecodeClipNms().__dict__):
+                 nms=None):
         super(BBoxPostProcess, self).__init__()
+        self.num_classes = num_classes
+        self.num_stages = num_stages
         self.decode = decode
         self.clip = clip
         self.nms = nms
         self.decode_clip_nms = decode_clip_nms
-        if isinstance(decode_clip_nms, dict):
-            self.decode_clip_nms = DecodeClipNms(**decode_clip_nms)
 
     def __call__(self, inputs):
         # TODO: split into 3 steps
@@ -31,8 +29,25 @@ class BBoxPostProcess(object):
         # decode
         # clip
         # nms
-        outs = self.decode_clip_nms(inputs['rpn_rois'], inputs['bbox_prob'],
-                                    inputs['bbox_delta'], inputs['im_info'])
+        if self.num_stages > 0:
+            bbox_prob_list = []
+            for i in range(self.num_stages):
+                bbox_prob_list.append(inputs['bbox_head_' + str(i)][
+                    'bbox_prob'])
+            bbox_prob = fluid.layers.sum(bbox_prob_list) / float(
+                len(bbox_prob_list))
+            bbox_delta = inputs['bbox_head_' + str(i)]['bbox_delta']
+            if inputs['bbox_head_0']['cls_agnostic_bbox_reg'] == 2:
+                bbox_delta = fluid.layers.slice(
+                    bbox_delta, axes=1, starts=[1], ends=[2])
+                bbox_delta = fluid.layers.expand(bbox_delta,
+                                                 [1, self.num_classes, 1])
+        else:
+            bbox_prob = inputs['bbox_prob']
+            bbox_delta = inputs['bbox_delta']
+
+        outs = self.decode_clip_nms(inputs['rpn_rois'], bbox_prob, bbox_delta,
+                                    inputs['im_info'])
         outs = {"predicted_bbox_nums": outs[0], "predicted_bbox": outs[1]}
         return outs
 
@@ -40,22 +55,15 @@ class BBoxPostProcess(object):
 @register
 class BBoxPostProcessYOLO(object):
     __shared__ = ['num_classes']
+    __inject__ = ['yolo_box', 'nms']
 
-    def __init__(self,
-                 num_classes=80,
-                 decode=None,
-                 clip=None,
-                 yolo_box=YOLOBox().__dict__,
-                 nms=MultiClassNMS().__dict__):
+    def __init__(self, yolo_box, nms, num_classes=80, decode=None, clip=None):
         super(BBoxPostProcessYOLO, self).__init__()
+        self.yolo_box = yolo_box
+        self.nms = nms
         self.num_classes = num_classes
         self.decode = decode
         self.clip = clip
-        self.nms = nms
-        if isinstance(yolo_box, dict):
-            self.yolo_box = YOLOBox(**yolo_box)
-        if isinstance(nms, dict):
-            self.nms = MultiClassNMS(**nms)
 
     def __call__(self, inputs):
         # TODO: split yolo_box into 2 steps
@@ -80,40 +88,14 @@ class BBoxPostProcessYOLO(object):
         return outs
 
 
-@register
-class MaskPostProcess(object):
-    __shared__ = ['num_classes']
-
-    def __init__(self, num_classes=81):
-        super(MaskPostProcess, self).__init__()
-        self.num_classes = num_classes
-
-    def __call__(self, inputs):
-        # TODO: modify related ops for deploying
-        outs = mask_post_process(inputs['predicted_bbox_nums'].numpy(),
-                                 inputs['predicted_bbox'].numpy(),
-                                 inputs['mask_logits'].numpy(),
-                                 inputs['im_info'].numpy())
-        outs = {'predicted_mask': outs}
-        return outs
-
-
 @register
 class AnchorRPN(object):
     __inject__ = ['anchor_generator', 'anchor_target_generator']
 
-    def __init__(self,
-                 anchor_type='rpn',
-                 anchor_generator=AnchorGeneratorRPN().__dict__,
-                 anchor_target_generator=AnchorTargetGeneratorRPN().__dict__):
+    def __init__(self, anchor_generator, anchor_target_generator):
         super(AnchorRPN, self).__init__()
         self.anchor_generator = anchor_generator
         self.anchor_target_generator = anchor_target_generator
-        if isinstance(anchor_generator, dict):
-            self.anchor_generator = AnchorGeneratorRPN(**anchor_generator)
-        if isinstance(anchor_target_generator, dict):
-            self.anchor_target_generator = AnchorTargetGeneratorRPN(
-                **anchor_target_generator)
 
     def __call__(self, inputs):
         outs = self.generate_anchors(inputs)
@@ -122,7 +104,7 @@ class AnchorRPN(object):
     def generate_anchors(self, inputs):
         # TODO: update here to use int to specify featmap size
         outs = self.anchor_generator(inputs['rpn_feat'])
-        outs = {'anchor': outs[0], 'var': outs[1], 'anchor_module': self}
+        outs = {'anchor': outs[0], 'anchor_var': outs[1], 'anchor_module': self}
         return outs
 
     def generate_anchors_target(self, inputs):
@@ -143,7 +125,8 @@ class AnchorRPN(object):
             anchor_box=anchor,
             gt_boxes=inputs['gt_bbox'],
             is_crowd=inputs['is_crowd'],
-            im_info=inputs['im_info'])
+            im_info=inputs['im_info'],
+            open_debug=inputs['open_debug'])
         outs = {
             'rpn_score_pred': score_pred,
             'rpn_score_target': score_tgt,
@@ -160,22 +143,12 @@ class AnchorYOLO(object):
         'anchor_generator', 'anchor_target_generator', 'anchor_post_process'
     ]
 
-    def __init__(self,
-                 anchor_generator=AnchorGeneratorYOLO().__dict__,
-                 anchor_target_generator=AnchorTargetGeneratorYOLO().__dict__,
-                 anchor_post_process=BBoxPostProcessYOLO().__dict__):
+    def __init__(self, anchor_generator, anchor_target_generator,
+                 anchor_post_process):
         super(AnchorYOLO, self).__init__()
         self.anchor_generator = anchor_generator
         self.anchor_target_generator = anchor_target_generator
         self.anchor_post_process = anchor_post_process
-        if isinstance(anchor_generator, dict):
-            self.anchor_generator = AnchorGeneratorYOLO(**anchor_generator)
-        if isinstance(anchor_target_generator, dict):
-            self.anchor_target_generator = AnchorTargetGeneratorYOLO(
-                **anchor_target_generator)
-        if isinstance(anchor_post_process, dict):
-            self.anchor_post_process = BBoxPostProcessYOLO(
-                **anchor_post_process)
 
     def __call__(self, inputs):
         outs = self.generate_anchors(inputs)
@@ -200,30 +173,20 @@ class Proposal(object):
         'proposal_generator', 'proposal_target_generator', 'bbox_post_process'
     ]
 
-    def __init__(
-            self,
-            proposal_generator=ProposalGenerator().__dict__,
-            proposal_target_generator=ProposalTargetGenerator().__dict__,
-            bbox_post_process=BBoxPostProcess().__dict__, ):
+    def __init__(self, proposal_generator, proposal_target_generator,
+                 bbox_post_process):
         super(Proposal, self).__init__()
         self.proposal_generator = proposal_generator
         self.proposal_target_generator = proposal_target_generator
         self.bbox_post_process = bbox_post_process
-        if isinstance(proposal_generator, dict):
-            self.proposal_generator = ProposalGenerator(**proposal_generator)
-        if isinstance(proposal_target_generator, dict):
-            self.proposal_target_generator = ProposalTargetGenerator(
-                **proposal_target_generator)
-        if isinstance(bbox_post_process, dict):
-            self.bbox_post_process = BBoxPostProcess(**bbox_post_process)
-
-    def __call__(self, inputs, stage=0):
+
+    def __call__(self, inputs):
         outs = {}
-        if stage == 0:
+        if inputs['stage'] == 0:
             proposal_out = self.generate_proposal(inputs)
             inputs.update(proposal_out)
         if inputs['mode'] == 'train':
-            proposal_target_out = self.generate_proposal_target(inputs, stage)
+            proposal_target_out = self.generate_proposal_target(inputs)
             outs.update(proposal_target_out)
         return outs
 
@@ -234,7 +197,7 @@ class Proposal(object):
             scores=rpn_rois_prob,
             bbox_deltas=inputs['rpn_rois_delta'],
             anchors=inputs['anchor'],
-            variances=inputs['var'],
+            variances=inputs['anchor_var'],
             im_info=inputs['im_info'],
             mode=inputs['mode'])
         outs = {
@@ -244,15 +207,24 @@ class Proposal(object):
         }
         return outs
 
-    def generate_proposal_target(self, inputs, stage=0):
+    def generate_proposal_target(self, inputs):
+        if inputs['stage'] == 0:
+            rois = inputs['rpn_rois']
+            rois_num = inputs['rpn_rois_nums']
+        elif inputs['stage'] > 0:
+            last_proposal_out = inputs['proposal_' + str(inputs['stage'] - 1)]
+            rois = last_proposal_out['refined_bbox']
+            rois_num = last_proposal_out['rois_nums']
+
         outs = self.proposal_target_generator(
-            rpn_rois=inputs['rpn_rois'],
-            rpn_rois_nums=inputs['rpn_rois_nums'],
+            rpn_rois=rois,
+            rpn_rois_nums=rois_num,
             gt_classes=inputs['gt_class'],
             is_crowd=inputs['is_crowd'],
             gt_boxes=inputs['gt_bbox'],
             im_info=inputs['im_info'],
-            stage=stage)
+            stage=inputs['stage'],
+            open_debug=inputs['open_debug'])
         outs = {
             'rois': outs[0],
             'labels_int32': outs[1],
@@ -263,49 +235,31 @@ class Proposal(object):
         }
         return outs
 
-    def post_process(self, inputs):
-        outs = self.bbox_post_process(inputs)
-        return outs
-
-
-@register
-class Mask(object):
-    __inject__ = ['mask_target_generator', 'mask_post_process']
-
-    def __init__(self,
-                 mask_target_generator=MaskTargetGenerator().__dict__,
-                 mask_post_process=MaskPostProcess().__dict__):
-        super(Mask, self).__init__()
-        self.mask_target_generator = mask_target_generator
-        self.mask_post_process = mask_post_process
-        if isinstance(mask_target_generator, dict):
-            self.mask_target_generator = MaskTargetGenerator(
-                **mask_target_generator)
-        if isinstance(mask_post_process, dict):
-            self.mask_post_process = MaskPostProcess(**mask_post_process)
-
-    def __call__(self, inputs):
-        outs = {}
+    def refine_bbox(self, inputs):
         if inputs['mode'] == 'train':
-            outs = self.generate_mask_target(inputs)
-        return outs
-
-    def generate_mask_target(self, inputs):
-        outs = self.mask_target_generator(
-            im_info=inputs['im_info'],
-            gt_classes=inputs['gt_class'],
-            is_crowd=inputs['is_crowd'],
-            gt_segms=inputs['gt_mask'],
-            rois=inputs['rois'],
-            rois_nums=inputs['rois_nums'],
-            labels_int32=inputs['labels_int32'], )
-        outs = {
-            'mask_rois': outs[0],
-            'rois_has_mask_int32': outs[1],
-            'mask_int32': outs[2]
-        }
+            rois = inputs['proposal_' + str(inputs['stage'])]['rois']
+        else:
+            rois = inputs['rpn_rois']
+        bbox_head_out = inputs['bbox_head_' + str(inputs['stage'])]
+
+        bbox_delta_r = fluid.layers.reshape(
+            bbox_head_out['bbox_delta'],
+            (-1, inputs['bbox_head_0']['cls_agnostic_bbox_reg'], 4))
+        bbox_delta_s = fluid.layers.slice(
+            bbox_delta_r, axes=[1], starts=[1], ends=[2])
+
+        refined_bbox = fluid.layers.box_coder(
+            prior_box=rois,
+            prior_box_var=self.proposal_target_generator.bbox_reg_weights[
+                inputs['stage']],
+            target_box=bbox_delta_s,
+            code_type='decode_center_size',
+            box_normalized=False,
+            axis=1)
+        refined_bbox = fluid.layers.reshape(refined_bbox, shape=[-1, 4])
+        outs = {'refined_bbox': refined_bbox}
         return outs
 
     def post_process(self, inputs):
-        outs = self.mask_post_process(inputs)
+        outs = self.bbox_post_process(inputs)
         return outs

ppdet/modeling/head/bbox_head.py

 import paddle.fluid as fluid
 from paddle.fluid.dygraph import Layer
-
 from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.initializer import Normal, MSRA
 from paddle.fluid.regularizer import L2Decay
 from paddle.fluid.dygraph.nn import Conv2D, Pool2D
 from ppdet.core.workspace import register
+# TODO: del import and use inject
 from ..backbone.resnet import Blocks
-from ..ops import RoIExtractor
 
 
 @register
 class BBoxFeat(Layer):
     __inject__ = ['roi_extractor']
+    __shared__ = ['num_stages']
 
-    def __init__(self,
-                 feat_in=1024,
-                 feat_out=512,
-                 roi_extractor=RoIExtractor().__dict__,
-                 stage=0):
+    def __init__(self, roi_extractor, feat_in=1024, feat_out=512, num_stages=1):
         super(BBoxFeat, self).__init__()
         self.roi_extractor = roi_extractor
-        if isinstance(roi_extractor, dict):
-            self.roi_extractor = RoIExtractor(**roi_extractor)
-        if stage == 0:
+        self.num_stages = num_stages
+        self.res5s = []
+        for i in range(self.num_stages):
+            if i == 0:
                 postfix = ''
             else:
-            postfix = '_' + str(stage)
-        self.res5 = Blocks(
-            "res5", ch_in=feat_in, ch_out=feat_out, count=3, stride=2)
+                postfix = '_' + str(i)
+            # TODO: set norm type  
+            res5 = Blocks(
+                "res5" + postfix,
+                ch_in=feat_in,
+                ch_out=feat_out,
+                count=3,
+                stride=2)
+            self.res5s.append(res5)
         self.res5_pool = fluid.dygraph.Pool2D(
             pool_type='avg', global_pooling=True)
 
     def forward(self, inputs):
+
         if inputs['mode'] == 'train':
-            rois = inputs['rois']
-            rois_num = inputs['rois_nums']
+            in_rois = inputs['proposal_' + str(inputs['stage'])]
+            rois = in_rois['rois']
+            rois_num = in_rois['rois_nums']
         elif inputs['mode'] == 'infer':
             rois = inputs['rpn_rois']
             rois_num = inputs['rpn_rois_nums']
@@ -44,14 +49,14 @@ class BBoxFeat(Layer):
 
         rois_feat = self.roi_extractor(inputs['res4'], rois, rois_num)
         # TODO: add others 
-        y_res5 = self.res5(rois_feat)
+        y_res5 = self.res5s[inputs['stage']](rois_feat)
         y = self.res5_pool(y_res5)
         y = fluid.layers.squeeze(y, axes=[2, 3])
         outs = {
             'rois_feat': rois_feat,
             'res5': y_res5,
             "bbox_feat": y,
-            'shared_res5_block': self.res5,
+            'shared_res5_block': self.res5s[inputs['stage']],
             'shared_roi_extractor': self.roi_extractor
         }
         return outs
@@ -59,25 +64,30 @@ class BBoxFeat(Layer):
 
 @register
 class BBoxHead(Layer):
-    __shared__ = ['num_classes']
     __inject__ = ['bbox_feat']
+    __shared__ = ['num_classes', 'num_stages']
 
     def __init__(self,
-                 in_feat=2048,
+                 bbox_feat,
+                 feat_in=2048,
                  num_classes=81,
-                 bbox_feat=BBoxFeat().__dict__,
-                 stage=0):
+                 cls_agnostic_bbox_reg=81,
+                 num_stages=1):
         super(BBoxHead, self).__init__()
-        self.num_classes = num_classes
         self.bbox_feat = bbox_feat
-        if isinstance(bbox_feat, dict):
-            self.bbox_feat = BBoxFeat(**bbox_feat)
-        if stage == 0:
+        self.num_classes = num_classes
+        self.cls_agnostic_bbox_reg = cls_agnostic_bbox_reg
+        self.num_stages = num_stages
+
+        self.bbox_scores = []
+        self.bbox_deltas = []
+        for i in range(self.num_stages):
+            if i == 0:
                 postfix = ''
             else:
-            postfix = '_' + str(stage)
-        self.bbox_score = fluid.dygraph.Linear(
-            input_dim=in_feat,
+                postfix = '_' + str(i)
+            bbox_score = fluid.dygraph.Linear(
+                input_dim=feat_in,
                 output_dim=1 * self.num_classes,
                 act=None,
                 param_attr=ParamAttr(
@@ -89,9 +99,9 @@ class BBoxHead(Layer):
                     learning_rate=2.,
                     regularizer=L2Decay(0.)))
 
-        self.bbox_delta = fluid.dygraph.Linear(
-            input_dim=in_feat,
-            output_dim=4 * self.num_classes,
+            bbox_delta = fluid.dygraph.Linear(
+                input_dim=feat_in,
+                output_dim=4 * self.cls_agnostic_bbox_reg,
                 act=None,
                 param_attr=ParamAttr(
                     name='bbox_pred_w' + postfix,
@@ -101,37 +111,45 @@ class BBoxHead(Layer):
                     name='bbox_pred_b' + postfix,
                     learning_rate=2.,
                     regularizer=L2Decay(0.)))
+            self.bbox_scores.append(bbox_score)
+            self.bbox_deltas.append(bbox_delta)
 
     def forward(self, inputs):
         outs = self.bbox_feat(inputs)
         x = outs['bbox_feat']
-        bs = self.bbox_score(x)
-        bd = self.bbox_delta(x)
+        bs = self.bbox_scores[inputs['stage']](x)
+        bd = self.bbox_deltas[inputs['stage']](x)
         outs.update({'bbox_score': bs, 'bbox_delta': bd})
+        if inputs['stage'] == 0:
+            outs.update({"cls_agnostic_bbox_reg": self.cls_agnostic_bbox_reg})
         if inputs['mode'] == 'infer':
             bbox_prob = fluid.layers.softmax(bs, use_cudnn=False)
             outs['bbox_prob'] = bbox_prob
         return outs
 
     def loss(self, inputs):
+        bbox_out = inputs['bbox_head_' + str(inputs['stage'])]
+        bbox_target = inputs['proposal_' + str(inputs['stage'])]
+
         # bbox cls  
         labels_int64 = fluid.layers.cast(
-            x=inputs['labels_int32'], dtype='int64')
+            x=bbox_target['labels_int32'], dtype='int64')
         labels_int64.stop_gradient = True
-        bbox_score = fluid.layers.reshape(inputs['bbox_score'],
+        bbox_score = fluid.layers.reshape(bbox_out['bbox_score'],
                                           (-1, self.num_classes))
         loss_bbox_cls = fluid.layers.softmax_with_cross_entropy(
             logits=bbox_score, label=labels_int64)
         loss_bbox_cls = fluid.layers.reduce_mean(
-            loss_bbox_cls, name='loss_bbox_cls')
+            loss_bbox_cls, name='loss_bbox_cls_' + str(inputs['stage']))
+
         # bbox reg
         loss_bbox_reg = fluid.layers.smooth_l1(
-            x=inputs['bbox_delta'],
-            y=inputs['bbox_targets'],
-            inside_weight=inputs['bbox_inside_weights'],
-            outside_weight=inputs['bbox_outside_weights'],
+            x=bbox_out['bbox_delta'],
+            y=bbox_target['bbox_targets'],
+            inside_weight=bbox_target['bbox_inside_weights'],
+            outside_weight=bbox_target['bbox_outside_weights'],
             sigma=1.0)
         loss_bbox_reg = fluid.layers.reduce_mean(
-            loss_bbox_reg, name='loss_bbox_loc')
+            loss_bbox_reg, name='loss_bbox_loc_' + str(inputs['stage']))
 
         return loss_bbox_cls, loss_bbox_reg

ppdet/modeling/head/mask_head.py

@@ -6,29 +6,23 @@ from paddle.fluid.initializer import Normal, MSRA
 from paddle.fluid.regularizer import L2Decay
 from paddle.fluid.dygraph.nn import Conv2D, Pool2D
 from ppdet.core.workspace import register
-from ..ops import RoIExtractor
+# TODO: del it and use inject 
 from ..backbone.resnet import Blocks
 
 
 @register
 class MaskFeat(Layer):
-    __inject__ = ['mask_roi_extractor']
-
-    def __init__(self,
-                 feat_in=2048,
-                 feat_out=256,
-                 mask_roi_extractor=RoIExtractor().__dict__,
-                 stage=0):
+    def __init__(self, feat_in=2048, feat_out=256, mask_stages=1):
         super(MaskFeat, self).__init__()
         self.feat_in = feat_in
         self.feat_out = feat_out
-        self.mask_roi_extractor = mask_roi_extractor
-        if isinstance(mask_roi_extractor, dict):
-            self.mask_roi_extractor = RoIExtractor(**mask_roi_extractor)
-        if stage == 0:
+        self.mask_stages = mask_stages
+
+        for i in range(self.mask_stages):
+            if i == 0:
                 postfix = ''
             else:
-            postfix = '_' + str(stage)
+                postfix = '_' + str(i)
             self.upsample = fluid.dygraph.Conv2DTranspose(
                 num_channels=self.feat_in,
                 num_filters=self.feat_out,
@@ -36,25 +30,30 @@ class MaskFeat(Layer):
                 stride=2,
                 act='relu',
                 param_attr=ParamAttr(
-                name='conv5_mask_w' + postfix, initializer=MSRA(uniform=False)),
+                    name='conv5_mask_w' + postfix,
+                    initializer=MSRA(uniform=False)),
                 bias_attr=ParamAttr(
                     name='conv5_mask_b' + postfix,
                     learning_rate=2.,
                     regularizer=L2Decay(0.)))
 
     def forward(self, inputs):
+        bbox_head_out = inputs['bbox_head_' + str(inputs['stage'])]
         if inputs['mode'] == 'train':
-            x = inputs['res5']
+            x = bbox_head_out['res5']
             rois_feat = fluid.layers.gather(x, inputs['rois_has_mask_int32'])
         elif inputs['mode'] == 'infer':
             rois = inputs['predicted_bbox'][:, 2:] * inputs['im_info'][:, 2]
             rois_num = inputs['predicted_bbox_nums']
             # TODO: optim here 
-            if callable(inputs['shared_roi_extractor']):
-                rois_feat = inputs['shared_roi_extractor'](inputs['res4'], rois,
-                                                           rois_num)
-            if callable(inputs['shared_res5_block']):
-                rois_feat = inputs['shared_res5_block'](rois_feat)
+            shared_roi_ext = bbox_head_out['shared_roi_extractor']
+            if callable(shared_roi_ext):
+                rois_feat = shared_roi_ext(inputs['res4'], rois, rois_num)
+
+            shared_res5 = bbox_head_out['shared_res5_block']
+            if callable(shared_res5):
+                rois_feat = shared_res5(rois_feat)
+
         # upsample 
         y = self.upsample(rois_feat)
         outs = {'mask_feat': y}
@@ -67,22 +66,23 @@ class MaskHead(Layer):
     __inject__ = ['mask_feat']
 
     def __init__(self,
+                 mask_feat,
+                 num_classes=81,
                  feat_in=256,
                  resolution=14,
-                 num_classes=81,
-                 mask_feat=MaskFeat().__dict__,
-                 stage=0):
+                 mask_stages=1):
         super(MaskHead, self).__init__()
+        self.mask_feat = mask_feat
         self.feat_in = feat_in
         self.resolution = resolution
         self.num_classes = num_classes
-        self.mask_feat = mask_feat
-        if isinstance(mask_feat, dict):
-            self.mask_feat = MaskFeat(**mask_feat)
-        if stage == 0:
+        self.mask_stages = mask_stages
+
+        for i in range(self.mask_stages):
+            if i == 0:
                 postfix = ''
             else:
-            postfix = '_' + str(stage)
+                postfix = '_' + str(i)
             self.mask_fcn_logits = fluid.dygraph.Conv2D(
                 num_channels=self.feat_in,
                 num_filters=self.num_classes,
@@ -115,7 +115,6 @@ class MaskHead(Layer):
         return outs
 
     def loss(self, inputs):
-        # input needs (model_out, target)
         reshape_dim = self.num_classes * self.resolution * self.resolution
         mask_logits = fluid.layers.reshape(inputs['mask_logits'],
                                            (-1, reshape_dim))

ppdet/modeling/head/rpn_head.py

@@ -4,7 +4,6 @@ from paddle.fluid.param_attr import ParamAttr
 from paddle.fluid.initializer import Normal
 from paddle.fluid.regularizer import L2Decay
 from paddle.fluid.dygraph.nn import Conv2D
-
 from ppdet.core.workspace import register
 
 
@@ -20,10 +19,10 @@ class RPNFeat(Layer):
             padding=1,
             act='relu',
             param_attr=ParamAttr(
-                "conv_rpn_w", initializer=Normal(
+                name="conv_rpn_w", initializer=Normal(
                     loc=0., scale=0.01)),
             bias_attr=ParamAttr(
-                "conv_rpn_b", learning_rate=2., regularizer=L2Decay(0.)))
+                name="conv_rpn_b", learning_rate=2., regularizer=L2Decay(0.)))
 
     def forward(self, inputs):
         x = inputs.get('res4')
@@ -36,12 +35,10 @@ class RPNFeat(Layer):
 class RPNHead(Layer):
     __inject__ = ['rpn_feat']
 
-    def __init__(self, anchor_per_position=15, rpn_feat=RPNFeat().__dict__):
+    def __init__(self, rpn_feat, anchor_per_position=15):
         super(RPNHead, self).__init__()
-        self.anchor_per_position = anchor_per_position
         self.rpn_feat = rpn_feat
-        if isinstance(rpn_feat, dict):
-            self.rpn_feat = RPNFeat(**rpn_feat)
+        self.anchor_per_position = anchor_per_position
 
         # rpn roi classification scores
         self.rpn_rois_score = Conv2D(

ppdet/modeling/head/yolo_head.py

@@ -132,19 +132,11 @@ class YOLOv3Head(Layer):
     __shared__ = ['num_classes']
     __inject__ = ['yolo_feat']
 
-    def __init__(
-            self,
-            num_classes=80,
-            anchor_per_position=3,
-            mode='train',
-            yolo_feat=YOLOFeat().__dict__, ):
+    def __init__(self, yolo_feat, num_classes=80, anchor_per_position=3):
         super(YOLOv3Head, self).__init__()
         self.num_classes = num_classes
         self.anchor_per_position = anchor_per_position
-        self.mode = mode
         self.yolo_feat = yolo_feat
-        if isinstance(yolo_feat, dict):
-            self.yolo_feat = YOLOFeat(**yolo_feat)
 
         self.yolo_outs = []
         for i in range(3):

ppdet/modeling/mask.py

+import numpy as np
+import paddle.fluid as fluid
+from ppdet.core.workspace import register
+
+# TODO: regitster mask_post_process op 
+from ppdet.py_op.post_process import mask_post_process
+
+
+@register
+class MaskPostProcess(object):
+    __shared__ = ['num_classes']
+
+    def __init__(self, num_classes=81):
+        super(MaskPostProcess, self).__init__()
+        self.num_classes = num_classes
+
+    def __call__(self, inputs):
+        # TODO: modify related ops for deploying
+        outs = mask_post_process(inputs['predicted_bbox_nums'].numpy(),
+                                 inputs['predicted_bbox'].numpy(),
+                                 inputs['mask_logits'].numpy(),
+                                 inputs['im_info'].numpy())
+        outs = {'predicted_mask': outs}
+        return outs
+
+
+@register
+class Mask(object):
+    __inject__ = ['mask_target_generator', 'mask_post_process']
+
+    def __init__(self, mask_target_generator, mask_post_process):
+        super(Mask, self).__init__()
+        self.mask_target_generator = mask_target_generator
+        self.mask_post_process = mask_post_process
+
+    def __call__(self, inputs):
+        outs = {}
+        if inputs['mode'] == 'train':
+            outs = self.generate_mask_target(inputs)
+        return outs
+
+    def generate_mask_target(self, inputs):
+        proposal_out = inputs['proposal_' + str(inputs['stage'])]
+        outs = self.mask_target_generator(
+            im_info=inputs['im_info'],
+            gt_classes=inputs['gt_class'],
+            is_crowd=inputs['is_crowd'],
+            gt_segms=inputs['gt_mask'],
+            rois=proposal_out['rois'],
+            rois_nums=proposal_out['rois_nums'],
+            labels_int32=proposal_out['labels_int32'])
+        outs = {
+            'mask_rois': outs[0],
+            'rois_has_mask_int32': outs[1],
+            'mask_int32': outs[2]
+        }
+        return outs
+
+    def post_process(self, inputs):
+        outs = self.mask_post_process(inputs)
+        return outs

ppdet/modeling/ops.py

@@ -49,13 +49,20 @@ class AnchorTargetGeneratorRPN(object):
         self.negative_overlap = negative_overlap
         self.use_random = use_random
 
-    def __call__(self, cls_logits, bbox_pred, anchor_box, gt_boxes, is_crowd,
-                 im_info):
+    def __call__(self,
+                 cls_logits,
+                 bbox_pred,
+                 anchor_box,
+                 gt_boxes,
+                 is_crowd,
+                 im_info,
+                 open_debug=False):
         anchor_box = anchor_box.numpy()
         gt_boxes = gt_boxes.numpy()
         is_crowd = is_crowd.numpy()
         im_info = im_info.numpy()
-
+        if open_debug:
+            self.use_random = False
         loc_indexes, score_indexes, tgt_labels, tgt_bboxes, bbox_inside_weights = generate_rpn_anchor_target(
             anchor_box, gt_boxes, is_crowd, im_info, self.straddle_thresh,
             self.batch_size_per_im, self.positive_overlap,
@@ -191,7 +198,7 @@ class ProposalTargetGenerator(object):
                  bg_thresh_lo=[0., ],
                  bbox_reg_weights=[[0.1, 0.1, 0.2, 0.2]],
                  num_classes=81,
-                 shuffle_before_sample=True,
+                 use_random=True,
                  is_cls_agnostic=False,
                  is_cascade_rcnn=False):
         super(ProposalTargetGenerator, self).__init__()
@@ -202,7 +209,7 @@ class ProposalTargetGenerator(object):
         self.bg_thresh_lo = bg_thresh_lo
         self.bbox_reg_weights = bbox_reg_weights
         self.num_classes = num_classes
-        self.use_random = shuffle_before_sample
+        self.use_random = use_random
         self.is_cls_agnostic = is_cls_agnostic,
         self.is_cascade_rcnn = is_cascade_rcnn
 
@@ -213,13 +220,17 @@ class ProposalTargetGenerator(object):
                  is_crowd,
                  gt_boxes,
                  im_info,
-                 stage=0):
+                 stage=0,
+                 open_debug=False):
         rpn_rois = rpn_rois.numpy()
         rpn_rois_nums = rpn_rois_nums.numpy()
         gt_classes = gt_classes.numpy()
         gt_boxes = gt_boxes.numpy()
         is_crowd = is_crowd.numpy()
         im_info = im_info.numpy()
+        if open_debug:
+            self.use_random = False
+
         outs = generate_proposal_target(
             rpn_rois, rpn_rois_nums, gt_classes, is_crowd, gt_boxes, im_info,
             self.batch_size_per_im, self.fg_fraction, self.fg_thresh[stage],
@@ -265,10 +276,10 @@ class MaskTargetGenerator(object):
 @register
 class RoIExtractor(object):
     def __init__(self,
-                 resolution=7,
+                 resolution=14,
                  spatial_scale=1. / 16,
                  sampling_ratio=0,
-                 extractor_type='RoIPool'):
+                 extractor_type='RoIAlign'):
         super(RoIExtractor, self).__init__()
         if isinstance(resolution, Integral):
             resolution = [resolution, resolution]

ppdet/utils/data_structure.py

@@ -21,25 +21,44 @@ class BufferDict(dict):
         for k, v in dict(*args, **kwargs).items():
             self[k] = v
 
+    def update_v(self, key, value):
+        if key in self.keys():
+            super(BufferDict, self).__setitem__(key, value)
+        else:
+            raise Exception("The %s is not in global inputs dict" % key)
+
     def get(self, key):
         return self.__getitem__(key)
 
     def set(self, key, value):
-        self.__setitem__(key, value)
+        return self.__setitem__(key, value)
+
+    def debug(self, dshape=True, dvalue=True, dtype=False):
+        if self['open_debug']:
+            if self['debug_names'] is None:
+                ditems = self.keys()
+            else:
+                ditems = self['debug_names']
 
-    def debug(self, dshape=True, dtype=False, dvalue=False, name='all'):
-        if name == 'all':
-            ditems = self.items()
+            infos = {}
+            for k in ditems:
+                if type(k) is dict:
+                    i_d = {}
+                    for i, j in k.items():
+                        if type(j) is list:
+                            for jj in j:
+                                i_d[jj] = self.get_debug_info(self[i][jj])
+                        infos[i] = i_d
                 else:
-            ditems = self.get(name)
+                    infos[k] = self.get_debug_info(self[k])
+            print(infos)
 
-        for k, v in ditems:
-            info = [k]
+    def get_debug_info(self, v, dshape=True, dvalue=True, dtype=False):
+        info = []
         if dshape == True and hasattr(v, 'shape'):
             info.append(v.shape)
-            if dtype == True:
-                info.append(type(v))
         if dvalue == True and hasattr(v, 'numpy'):
             info.append(np.mean(np.abs(v.numpy())))
-
-            print(info)
+        if dtype == True:
+            info.append(type(v))
+        return info