diff --git a/dygraph/configs/ppyolo/_base_/ppyolo_r50vd_dcn.yml b/dygraph/configs/ppyolo/_base_/ppyolo_r50vd_dcn.yml index 3bc6af116c4ed00e0ceb5175bab98bd9c18ef994..18111ad025d94ffe2c5517bf7f08981a9cfe0af2 100644 --- a/dygraph/configs/ppyolo/_base_/ppyolo_r50vd_dcn.yml +++ b/dygraph/configs/ppyolo/_base_/ppyolo_r50vd_dcn.yml @@ -22,7 +22,6 @@ ResNet: norm_decay: 0. PPYOLOFPN: - feat_channels: [2048, 1280, 640] coord_conv: true drop_block: true block_size: 3 diff --git a/dygraph/configs/ssd/_base_/ssd_mobilenet_v1_300.yml b/dygraph/configs/ssd/_base_/ssd_mobilenet_v1_300.yml index c277a5dfcfd8b6fdde3271aabc087b41d6480536..1eb9f68ef230962b733be1c24badcc6c28fb6561 100644 --- a/dygraph/configs/ssd/_base_/ssd_mobilenet_v1_300.yml +++ b/dygraph/configs/ssd/_base_/ssd_mobilenet_v1_300.yml @@ -16,22 +16,19 @@ MobileNet: feature_maps: [11, 13, 14, 15, 16, 17] SSDHead: - in_channels: [512, 1024, 512, 256, 256, 128] - anchor_generator: AnchorGeneratorSSD kernel_size: 1 padding: 0 - -AnchorGeneratorSSD: - steps: [0, 0, 0, 0, 0, 0] - aspect_ratios: [[2.], [2., 3.], [2., 3.], [2., 3.], [2., 3.], [2., 3.]] - min_ratio: 20 - max_ratio: 90 - base_size: 300 - min_sizes: [60.0, 105.0, 150.0, 195.0, 240.0, 285.0] - max_sizes: [[], 150.0, 195.0, 240.0, 285.0, 300.0] - offset: 0.5 - flip: true - min_max_aspect_ratios_order: false + anchor_generator: + steps: [0, 0, 0, 0, 0, 0] + aspect_ratios: [[2.], [2., 3.], [2., 3.], [2., 3.], [2., 3.], [2., 3.]] + min_ratio: 20 + max_ratio: 90 + base_size: 300 + min_sizes: [60.0, 105.0, 150.0, 195.0, 240.0, 285.0] + max_sizes: [[], 150.0, 195.0, 240.0, 285.0, 300.0] + offset: 0.5 + flip: true + min_max_aspect_ratios_order: false BBoxPostProcess: decode: diff --git a/dygraph/configs/ssd/_base_/ssd_vgg16_300.yml b/dygraph/configs/ssd/_base_/ssd_vgg16_300.yml index 63ebc4ae07e9b535fccc7253bbd79e055306dfdb..3b9f039a5e2a99396ba0845c9f90e7b06fe3bf21 100644 --- a/dygraph/configs/ssd/_base_/ssd_vgg16_300.yml +++ b/dygraph/configs/ssd/_base_/ssd_vgg16_300.yml @@ -1,6 +1,5 @@ architecture: SSD pretrain_weights: https://paddlemodels.bj.bcebos.com/object_detection/dygraph/VGG16_caffe_pretrained.pdparams -load_static_weights: True # Model Achitecture SSD: @@ -15,19 +14,16 @@ VGG: normalizations: [20., -1, -1, -1, -1, -1] SSDHead: - in_channels: [512, 1024, 512, 256, 256, 256] - anchor_generator: AnchorGeneratorSSD - -AnchorGeneratorSSD: - steps: [8, 16, 32, 64, 100, 300] - aspect_ratios: [[2.], [2., 3.], [2., 3.], [2., 3.], [2.], [2.]] - min_ratio: 20 - max_ratio: 90 - min_sizes: [30.0, 60.0, 111.0, 162.0, 213.0, 264.0] - max_sizes: [60.0, 111.0, 162.0, 213.0, 264.0, 315.0] - offset: 0.5 - flip: true - min_max_aspect_ratios_order: true + anchor_generator: + steps: [8, 16, 32, 64, 100, 300] + aspect_ratios: [[2.], [2., 3.], [2., 3.], [2., 3.], [2.], [2.]] + min_ratio: 20 + max_ratio: 90 + min_sizes: [30.0, 60.0, 111.0, 162.0, 213.0, 264.0] + max_sizes: [60.0, 111.0, 162.0, 213.0, 264.0, 315.0] + offset: 0.5 + flip: true + min_max_aspect_ratios_order: true BBoxPostProcess: decode: diff --git a/dygraph/configs/ssd/_base_/ssdlite_mobilenet_v1_300.yml b/dygraph/configs/ssd/_base_/ssdlite_mobilenet_v1_300.yml index 95847a0b92f8039acb325fcc1d72e74b8bdd832d..2e5a47485be13ad398840b15af9ddaed9316188e 100644 --- a/dygraph/configs/ssd/_base_/ssdlite_mobilenet_v1_300.yml +++ b/dygraph/configs/ssd/_base_/ssdlite_mobilenet_v1_300.yml @@ -15,23 +15,20 @@ MobileNet: feature_maps: [11, 13, 14, 15, 16, 17] SSDHead: - in_channels: [512, 1024, 512, 256, 256, 128] - anchor_generator: AnchorGeneratorSSD use_sepconv: True conv_decay: 0.00004 - -AnchorGeneratorSSD: - steps: [16, 32, 64, 100, 150, 300] - aspect_ratios: [[2.], [2., 3.], [2., 3.], [2., 3.], [2., 3.], [2., 3.]] - min_ratio: 20 - max_ratio: 95 - base_size: 300 - min_sizes: [] - max_sizes: [] - offset: 0.5 - flip: true - clip: true - min_max_aspect_ratios_order: False + anchor_generator: + steps: [16, 32, 64, 100, 150, 300] + aspect_ratios: [[2.], [2., 3.], [2., 3.], [2., 3.], [2., 3.], [2., 3.]] + min_ratio: 20 + max_ratio: 95 + base_size: 300 + min_sizes: [] + max_sizes: [] + offset: 0.5 + flip: true + clip: true + min_max_aspect_ratios_order: False BBoxPostProcess: decode: diff --git a/dygraph/configs/ssd/_base_/ssdlite_mobilenet_v3_large_320.yml b/dygraph/configs/ssd/_base_/ssdlite_mobilenet_v3_large_320.yml index e7c2b1c3d3a3ea6619eb6203beb8f5f6e15b6da2..a3586ee2f2dad0591a312df1f77cfb41021a6ee1 100644 --- a/dygraph/configs/ssd/_base_/ssdlite_mobilenet_v3_large_320.yml +++ b/dygraph/configs/ssd/_base_/ssdlite_mobilenet_v3_large_320.yml @@ -18,23 +18,20 @@ MobileNetV3: multiplier: 0.5 SSDHead: - in_channels: [672, 480, 512, 256, 256, 128] - anchor_generator: AnchorGeneratorSSD use_sepconv: True conv_decay: 0.00004 - -AnchorGeneratorSSD: - steps: [16, 32, 64, 107, 160, 320] - aspect_ratios: [[2.], [2., 3.], [2., 3.], [2., 3.], [2., 3.], [2., 3.]] - min_ratio: 20 - max_ratio: 95 - base_size: 320 - min_sizes: [] - max_sizes: [] - offset: 0.5 - flip: true - clip: true - min_max_aspect_ratios_order: false + anchor_generator: + steps: [16, 32, 64, 107, 160, 320] + aspect_ratios: [[2.], [2., 3.], [2., 3.], [2., 3.], [2., 3.], [2., 3.]] + min_ratio: 20 + max_ratio: 95 + base_size: 320 + min_sizes: [] + max_sizes: [] + offset: 0.5 + flip: true + clip: true + min_max_aspect_ratios_order: false BBoxPostProcess: decode: diff --git a/dygraph/configs/ssd/_base_/ssdlite_mobilenet_v3_small_320.yml b/dygraph/configs/ssd/_base_/ssdlite_mobilenet_v3_small_320.yml index 21dfe669a6a05456cdee52e416edbdaf420a3ab7..6169d9759c29f4a125285590cb4b8f50e54cd2fd 100644 --- a/dygraph/configs/ssd/_base_/ssdlite_mobilenet_v3_small_320.yml +++ b/dygraph/configs/ssd/_base_/ssdlite_mobilenet_v3_small_320.yml @@ -18,23 +18,20 @@ MobileNetV3: multiplier: 0.5 SSDHead: - in_channels: [288, 288, 512, 256, 256, 128] - anchor_generator: AnchorGeneratorSSD use_sepconv: True conv_decay: 0.00004 - -AnchorGeneratorSSD: - steps: [16, 32, 64, 107, 160, 320] - aspect_ratios: [[2.], [2., 3.], [2., 3.], [2., 3.], [2., 3.], [2., 3.]] - min_ratio: 20 - max_ratio: 95 - base_size: 320 - min_sizes: [] - max_sizes: [] - offset: 0.5 - flip: true - clip: true - min_max_aspect_ratios_order: false + anchor_generator: + steps: [16, 32, 64, 107, 160, 320] + aspect_ratios: [[2.], [2., 3.], [2., 3.], [2., 3.], [2., 3.], [2., 3.]] + min_ratio: 20 + max_ratio: 95 + base_size: 320 + min_sizes: [] + max_sizes: [] + offset: 0.5 + flip: true + clip: true + min_max_aspect_ratios_order: false BBoxPostProcess: decode: diff --git a/dygraph/configs/yolov3/_base_/yolov3_darknet53.yml b/dygraph/configs/yolov3/_base_/yolov3_darknet53.yml index c094ba71cb485b76233fc5297923db3929f4c093..796c245019bed347a79a192546a618df129fafdf 100644 --- a/dygraph/configs/yolov3/_base_/yolov3_darknet53.yml +++ b/dygraph/configs/yolov3/_base_/yolov3_darknet53.yml @@ -14,8 +14,8 @@ DarkNet: depth: 53 return_idx: [2, 3, 4] -YOLOv3FPN: - feat_channels: [1024, 768, 384] +# use default config +# YOLOv3FPN: YOLOv3Head: anchors: [[10, 13], [16, 30], [33, 23], diff --git a/dygraph/configs/yolov3/_base_/yolov3_mobilenet_v1.yml b/dygraph/configs/yolov3/_base_/yolov3_mobilenet_v1.yml index c2fa0e931b19d8753c7351d813aa57f72254d1e2..5e4a5b04f8e2510b91fc2930cd460588c8bff805 100644 --- a/dygraph/configs/yolov3/_base_/yolov3_mobilenet_v1.yml +++ b/dygraph/configs/yolov3/_base_/yolov3_mobilenet_v1.yml @@ -15,8 +15,8 @@ MobileNet: with_extra_blocks: false extra_block_filters: [] -YOLOv3FPN: - feat_channels: [1024, 768, 384] +# use default config +# YOLOv3FPN: YOLOv3Head: anchors: [[10, 13], [16, 30], [33, 23], diff --git a/dygraph/configs/yolov3/_base_/yolov3_mobilenet_v3_large.yml b/dygraph/configs/yolov3/_base_/yolov3_mobilenet_v3_large.yml index d9085fafb2c1de42b9df4146cefa8386ad1cc454..c95ec77a0a471dcbce4cc0aefc4fa18e83c5a0e8 100644 --- a/dygraph/configs/yolov3/_base_/yolov3_mobilenet_v3_large.yml +++ b/dygraph/configs/yolov3/_base_/yolov3_mobilenet_v3_large.yml @@ -16,8 +16,8 @@ MobileNetV3: extra_block_filters: [] feature_maps: [7, 13, 16] -YOLOv3FPN: - feat_channels: [160, 368, 168] +# use default config +# YOLOv3FPN: YOLOv3Head: anchors: [[10, 13], [16, 30], [33, 23], diff --git a/dygraph/configs/yolov3/_base_/yolov3_mobilenet_v3_small.yml b/dygraph/configs/yolov3/_base_/yolov3_mobilenet_v3_small.yml index a6f0ecb3dfe1b1822ff75c11346a1f10ea328a34..88291b22dd1ae2e88d46e5df27bd0f14c5061d43 100644 --- a/dygraph/configs/yolov3/_base_/yolov3_mobilenet_v3_small.yml +++ b/dygraph/configs/yolov3/_base_/yolov3_mobilenet_v3_small.yml @@ -16,8 +16,8 @@ MobileNetV3: extra_block_filters: [] feature_maps: [4, 9, 12] -YOLOv3FPN: - feat_channels: [96, 304, 152] +# use default config +# YOLOv3FPN: YOLOv3Head: anchors: [[10, 13], [16, 30], [33, 23], diff --git a/dygraph/ppdet/modeling/architectures/ssd.py b/dygraph/ppdet/modeling/architectures/ssd.py index 92386db5151df35c7a9a05b1a11011e363eebeb1..55ff07efeee491362d6ed94f476979935aae79d9 100644 --- a/dygraph/ppdet/modeling/architectures/ssd.py +++ b/dygraph/ppdet/modeling/architectures/ssd.py @@ -2,7 +2,7 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -from ppdet.core.workspace import register +from ppdet.core.workspace import register, create from .meta_arch import BaseArch __all__ = ['SSD'] @@ -11,38 +11,47 @@ __all__ = ['SSD'] @register class SSD(BaseArch): __category__ = 'architecture' - __inject__ = ['backbone', 'neck', 'ssd_head', 'post_process'] + __inject__ = ['post_process'] - def __init__(self, backbone, ssd_head, post_process, neck=None): + def __init__(self, backbone, ssd_head, post_process): super(SSD, self).__init__() self.backbone = backbone - self.neck = neck self.ssd_head = ssd_head self.post_process = post_process - def model_arch(self): + @classmethod + def from_config(cls, cfg, *args, **kwargs): + # backbone + backbone = create(cfg['backbone']) + + # head + kwargs = {'input_shape': backbone.out_shape} + ssd_head = create(cfg['ssd_head'], **kwargs) + + return { + 'backbone': backbone, + "ssd_head": ssd_head, + } + + def _forward(self): # Backbone body_feats = self.backbone(self.inputs) - # Neck - if self.neck is not None: - body_feats, spatial_scale = self.neck(body_feats) - # SSD Head - self.ssd_head_outs, self.anchors = self.ssd_head(body_feats, - self.inputs['image']) + if self.training: + return self.ssd_head(body_feats, self.inputs['image'], + self.inputs['gt_bbox'], + self.inputs['gt_class']) + else: + boxes, scores, anchors = self.ssd_head(body_feats, + self.inputs['image']) + bbox, bbox_num = self.post_process((boxes, scores), anchors, + self.inputs['im_shape'], + self.inputs['scale_factor']) + return bbox, bbox_num def get_loss(self, ): - loss = self.ssd_head.get_loss(self.ssd_head_outs, self.inputs, - self.anchors) - return {"loss": loss} + return {"loss": self._forward()} def get_pred(self): - bbox, bbox_num = self.post_process(self.ssd_head_outs, self.anchors, - self.inputs['im_shape'], - self.inputs['scale_factor']) - outs = { - "bbox": bbox, - "bbox_num": bbox_num, - } - return outs + return dict(zip(['bbox', 'bbox_num'], self._forward())) diff --git a/dygraph/ppdet/modeling/architectures/yolo.py b/dygraph/ppdet/modeling/architectures/yolo.py index 19ec048a23020712584c91e8bd395430920e2aee..42391bb989a4d4c2fc44f724a6ea1bb0ab267ccb 100644 --- a/dygraph/ppdet/modeling/architectures/yolo.py +++ b/dygraph/ppdet/modeling/architectures/yolo.py @@ -2,7 +2,7 @@ from __future__ import absolute_import from __future__ import division from __future__ import print_function -from ppdet.core.workspace import register +from ppdet.core.workspace import register, create from .meta_arch import BaseArch __all__ = ['YOLOv3'] @@ -11,12 +11,7 @@ __all__ = ['YOLOv3'] @register class YOLOv3(BaseArch): __category__ = 'architecture' - __inject__ = [ - 'backbone', - 'neck', - 'yolo_head', - 'post_process', - ] + __inject__ = ['post_process'] def __init__(self, backbone='DarkNet', @@ -29,27 +24,50 @@ class YOLOv3(BaseArch): self.yolo_head = yolo_head self.post_process = post_process - def model_arch(self, ): - # Backbone - body_feats = self.backbone(self.inputs) + @classmethod + def from_config(cls, cfg, *args, **kwargs): + # backbone + backbone = create(cfg['backbone']) + + # fpn + kwargs = {'input_shape': backbone.out_shape} + neck = create(cfg['neck'], **kwargs) + + # head + kwargs = {'input_shape': neck.out_shape} + yolo_head = create(cfg['yolo_head'], **kwargs) - # neck + return { + 'backbone': backbone, + 'neck': neck, + "yolo_head": yolo_head, + } + + def _forward(self): + body_feats = self.backbone(self.inputs) body_feats = self.neck(body_feats) - # YOLO Head - self.yolo_head_outs = self.yolo_head(body_feats) + if self.training: + return self.yolo_head(body_feats, self.inputs) + else: + yolo_head_outs = self.yolo_head(body_feats) + bbox, bbox_num = self.post_process( + yolo_head_outs, self.yolo_head.mask_anchors, + self.inputs['im_shape'], self.inputs['scale_factor']) + return bbox, bbox_num - def get_loss(self, ): - loss = self.yolo_head.get_loss(self.yolo_head_outs, self.inputs) - return loss + def get_loss(self): + return self._forward() def get_pred(self): - yolo_head_outs = self.yolo_head.get_outputs(self.yolo_head_outs) - bbox, bbox_num = self.post_process( - yolo_head_outs, self.yolo_head.mask_anchors, - self.inputs['im_shape'], self.inputs['scale_factor']) - outs = { - "bbox": bbox, - "bbox_num": bbox_num, + bbox_pred, bbox_num = self._forward() + label = bbox_pred[:, 0] + score = bbox_pred[:, 1] + bbox = bbox_pred[:, 2:] + output = { + 'bbox': bbox, + 'score': score, + 'label': label, + 'bbox_num': bbox_num } - return outs + return output diff --git a/dygraph/ppdet/modeling/backbones/darknet.py b/dygraph/ppdet/modeling/backbones/darknet.py index 4403dc48dc12969173b13f3164a36fb40eff564b..9dd5e07d1cead6702768e690ceab380cf17fd545 100755 --- a/dygraph/ppdet/modeling/backbones/darknet.py +++ b/dygraph/ppdet/modeling/backbones/darknet.py @@ -19,6 +19,7 @@ from paddle import ParamAttr from paddle.regularizer import L2Decay from ppdet.core.workspace import register, serializable from ppdet.modeling.ops import batch_norm +from ..shape_spec import ShapeSpec __all__ = ['DarkNet', 'ConvBNLayer'] @@ -193,6 +194,7 @@ class DarkNet(nn.Layer): norm_decay=norm_decay, name='yolo_input.downsample') + self._out_channels = [] self.darknet_conv_block_list = [] self.downsample_list = [] ch_in = [64, 128, 256, 512, 1024] @@ -208,6 +210,8 @@ class DarkNet(nn.Layer): norm_decay=norm_decay, name=name)) self.darknet_conv_block_list.append(conv_block) + if i in return_idx: + self._out_channels.append(64 * (2**i)) for i in range(num_stages - 1): down_name = 'stage.{}.downsample'.format(i) downsample = self.add_sublayer( @@ -235,3 +239,7 @@ class DarkNet(nn.Layer): if i < self.num_stages - 1: out = self.downsample_list[i](out) return blocks + + @property + def out_shape(self): + return [ShapeSpec(channels=c) for c in self._out_channels] diff --git a/dygraph/ppdet/modeling/backbones/mobilenet_v1.py b/dygraph/ppdet/modeling/backbones/mobilenet_v1.py index 198773b5edb22ccd7e26f3f3e2355d7489fe051c..5b4d1287847808f0e4c0d20b7809a1efee3cd82b 100644 --- a/dygraph/ppdet/modeling/backbones/mobilenet_v1.py +++ b/dygraph/ppdet/modeling/backbones/mobilenet_v1.py @@ -24,6 +24,7 @@ from paddle.regularizer import L2Decay from paddle.nn.initializer import KaimingNormal from ppdet.core.workspace import register, serializable from numbers import Integral +from ..shape_spec import ShapeSpec __all__ = ['MobileNet'] @@ -201,6 +202,8 @@ class MobileNet(nn.Layer): self.with_extra_blocks = with_extra_blocks self.extra_block_filters = extra_block_filters + self._out_channels = [] + self.conv1 = ConvBNLayer( in_channels=3, out_channels=int(32 * scale), @@ -229,6 +232,7 @@ class MobileNet(nn.Layer): norm_type=norm_type, name="conv2_1")) self.dwsl.append(dws21) + self._update_out_channels(64, len(self.dwsl), feature_maps) dws22 = self.add_sublayer( "conv2_2", sublayer=DepthwiseSeparable( @@ -244,6 +248,7 @@ class MobileNet(nn.Layer): norm_type=norm_type, name="conv2_2")) self.dwsl.append(dws22) + self._update_out_channels(128, len(self.dwsl), feature_maps) # 1/4 dws31 = self.add_sublayer( "conv3_1", @@ -260,6 +265,7 @@ class MobileNet(nn.Layer): norm_type=norm_type, name="conv3_1")) self.dwsl.append(dws31) + self._update_out_channels(128, len(self.dwsl), feature_maps) dws32 = self.add_sublayer( "conv3_2", sublayer=DepthwiseSeparable( @@ -275,6 +281,7 @@ class MobileNet(nn.Layer): norm_type=norm_type, name="conv3_2")) self.dwsl.append(dws32) + self._update_out_channels(256, len(self.dwsl), feature_maps) # 1/8 dws41 = self.add_sublayer( "conv4_1", @@ -291,6 +298,7 @@ class MobileNet(nn.Layer): norm_type=norm_type, name="conv4_1")) self.dwsl.append(dws41) + self._update_out_channels(256, len(self.dwsl), feature_maps) dws42 = self.add_sublayer( "conv4_2", sublayer=DepthwiseSeparable( @@ -306,6 +314,7 @@ class MobileNet(nn.Layer): norm_type=norm_type, name="conv4_2")) self.dwsl.append(dws42) + self._update_out_channels(512, len(self.dwsl), feature_maps) # 1/16 for i in range(5): tmp = self.add_sublayer( @@ -323,6 +332,7 @@ class MobileNet(nn.Layer): norm_type=norm_type, name="conv5_" + str(i + 1))) self.dwsl.append(tmp) + self._update_out_channels(512, len(self.dwsl), feature_maps) dws56 = self.add_sublayer( "conv5_6", sublayer=DepthwiseSeparable( @@ -338,6 +348,7 @@ class MobileNet(nn.Layer): norm_type=norm_type, name="conv5_6")) self.dwsl.append(dws56) + self._update_out_channels(1024, len(self.dwsl), feature_maps) # 1/32 dws6 = self.add_sublayer( "conv6", @@ -354,6 +365,7 @@ class MobileNet(nn.Layer): norm_type=norm_type, name="conv6")) self.dwsl.append(dws6) + self._update_out_channels(1024, len(self.dwsl), feature_maps) if self.with_extra_blocks: self.extra_blocks = [] @@ -371,6 +383,13 @@ class MobileNet(nn.Layer): norm_type=norm_type, name="conv7_" + str(i + 1))) self.extra_blocks.append(conv_extra) + self._update_out_channels( + block_filter[1], + len(self.dwsl) + len(self.extra_blocks), feature_maps) + + def _update_out_channels(self, channel, feature_idx, feature_maps): + if feature_idx in feature_maps: + self._out_channels.append(channel) def forward(self, inputs): outs = [] @@ -390,3 +409,7 @@ class MobileNet(nn.Layer): if idx + 1 in self.feature_maps: outs.append(y) return outs + + @property + def out_shape(self): + return [ShapeSpec(channels=c) for c in self._out_channels] diff --git a/dygraph/ppdet/modeling/backbones/mobilenet_v3.py b/dygraph/ppdet/modeling/backbones/mobilenet_v3.py index 40ee4e86ff0dba64c2c749df4c7f09e6422f80b3..1cebf5ef1e08c341f4ffb2670f1737974750ba6b 100644 --- a/dygraph/ppdet/modeling/backbones/mobilenet_v3.py +++ b/dygraph/ppdet/modeling/backbones/mobilenet_v3.py @@ -23,6 +23,7 @@ from paddle import ParamAttr from paddle.regularizer import L2Decay from ppdet.core.workspace import register, serializable from numbers import Integral +from ..shape_spec import ShapeSpec __all__ = ['MobileNetV3'] @@ -383,6 +384,7 @@ class MobileNetV3(nn.Layer): freeze_norm=freeze_norm, name="conv1") + self._out_channels = [] self.block_list = [] i = 0 inplanes = make_divisible(inplanes * scale) @@ -413,6 +415,9 @@ class MobileNetV3(nn.Layer): self.block_list.append(block) inplanes = make_divisible(scale * c) i += 1 + self._update_out_channels( + make_divisible(scale * exp) + if return_list else inplanes, i + 1, feature_maps) if self.with_extra_blocks: self.extra_block_list = [] @@ -438,6 +443,7 @@ class MobileNetV3(nn.Layer): name="conv" + str(i + 2))) self.extra_block_list.append(conv_extra) i += 1 + self._update_out_channels(extra_out_c, i + 1, feature_maps) for j, block_filter in enumerate(self.extra_block_filters): in_c = extra_out_c if j == 0 else self.extra_block_filters[j - @@ -457,6 +463,11 @@ class MobileNetV3(nn.Layer): name='conv' + str(i + 2))) self.extra_block_list.append(conv_extra) i += 1 + self._update_out_channels(block_filter[1], i + 1, feature_maps) + + def _update_out_channels(self, channel, feature_idx, feature_maps): + if feature_idx in feature_maps: + self._out_channels.append(channel) def forward(self, inputs): x = self.conv1(inputs['image']) @@ -479,3 +490,7 @@ class MobileNetV3(nn.Layer): if idx + 2 in self.feature_maps: outs.append(x) return outs + + @property + def out_shape(self): + return [ShapeSpec(channels=c) for c in self._out_channels] diff --git a/dygraph/ppdet/modeling/backbones/vgg.py b/dygraph/ppdet/modeling/backbones/vgg.py index 9d1b41ce6c1dade75c5e5b59610164bdc0cc4949..dd03872cbe5fd845a4d2ddb4b452c1ec3eb46196 100755 --- a/dygraph/ppdet/modeling/backbones/vgg.py +++ b/dygraph/ppdet/modeling/backbones/vgg.py @@ -7,6 +7,7 @@ from paddle import ParamAttr from paddle.regularizer import L2Decay from paddle.nn import Conv2D, MaxPool2D from ppdet.core.workspace import register, serializable +from ..shape_spec import ShapeSpec __all__ = ['VGG'] @@ -129,6 +130,8 @@ class VGG(nn.Layer): self.normalizations = normalizations self.extra_block_filters = extra_block_filters + self._out_channels = [] + self.conv_block_0 = ConvBlock( 3, 64, self.groups[0], 2, 2, 0, name="conv1_") self.conv_block_1 = ConvBlock( @@ -139,6 +142,7 @@ class VGG(nn.Layer): 256, 512, self.groups[3], 2, 2, 0, name="conv4_") self.conv_block_4 = ConvBlock( 512, 512, self.groups[4], 3, 1, 1, name="conv5_") + self._out_channels.append(512) self.fc6 = Conv2D( in_channels=512, @@ -153,6 +157,7 @@ class VGG(nn.Layer): kernel_size=1, stride=1, padding=0) + self._out_channels.append(1024) # extra block self.extra_convs = [] @@ -164,6 +169,7 @@ class VGG(nn.Layer): v[2], v[3], v[4])) last_channels = v[1] self.extra_convs.append(extra_conv) + self._out_channels.append(last_channels) self.norms = [] for i, n in enumerate(self.normalizations): @@ -192,7 +198,7 @@ class VGG(nn.Layer): outputs.append(out) if not self.extra_block_filters: - return out + return outputs # extra block for extra_conv in self.extra_convs: @@ -204,3 +210,7 @@ class VGG(nn.Layer): outputs[i] = self.norms[i](outputs[i]) return outputs + + @property + def out_shape(self): + return [ShapeSpec(channels=c) for c in self._out_channels] diff --git a/dygraph/ppdet/modeling/heads/ssd_head.py b/dygraph/ppdet/modeling/heads/ssd_head.py index 505a8e2980097b845d7bcac8b0cb4e1814ec727c..fb004c498b1ab30b908df9e50d1ca26bc80114d7 100644 --- a/dygraph/ppdet/modeling/heads/ssd_head.py +++ b/dygraph/ppdet/modeling/heads/ssd_head.py @@ -5,6 +5,8 @@ from ppdet.core.workspace import register from paddle.regularizer import L2Decay from paddle import ParamAttr +from ..layers import AnchorGeneratorSSD + class SepConvLayer(nn.Layer): def __init__(self, @@ -58,7 +60,7 @@ class SSDHead(nn.Layer): def __init__(self, num_classes=81, in_channels=(512, 1024, 512, 256, 256, 256), - anchor_generator='AnchorGeneratorSSD', + anchor_generator=AnchorGeneratorSSD().__dict__, kernel_size=3, padding=1, use_sepconv=False, @@ -69,8 +71,11 @@ class SSDHead(nn.Layer): self.in_channels = in_channels self.anchor_generator = anchor_generator self.loss = loss - self.num_priors = self.anchor_generator.num_priors + if isinstance(anchor_generator, dict): + self.anchor_generator = AnchorGeneratorSSD(**anchor_generator) + + self.num_priors = self.anchor_generator.num_priors self.box_convs = [] self.score_convs = [] for i, num_prior in enumerate(self.num_priors): @@ -116,7 +121,11 @@ class SSDHead(nn.Layer): name=score_conv_name)) self.score_convs.append(score_conv) - def forward(self, feats, image): + @classmethod + def from_config(cls, cfg, input_shape): + return {'in_channels': [i.channels for i in input_shape], } + + def forward(self, feats, image, gt_bbox=None, gt_class=None): box_preds = [] cls_scores = [] prior_boxes = [] @@ -134,10 +143,11 @@ class SSDHead(nn.Layer): prior_boxes = self.anchor_generator(feats, image) - outputs = {} - outputs['boxes'] = box_preds - outputs['scores'] = cls_scores - return outputs, prior_boxes + if self.training: + return self.get_loss(box_preds, cls_scores, gt_bbox, gt_class, + prior_boxes) + else: + return box_preds, cls_scores, prior_boxes - def get_loss(self, inputs, targets, prior_boxes): - return self.loss(inputs, targets, prior_boxes) + def get_loss(self, boxes, scores, gt_bbox, gt_class, prior_boxes): + return self.loss(boxes, scores, gt_bbox, gt_class, prior_boxes) diff --git a/dygraph/ppdet/modeling/heads/yolo_head.py b/dygraph/ppdet/modeling/heads/yolo_head.py index cd225d576c795895e78c7ab648881e2aaae7a8ff..ab32ce1e6e3efab1d75ee9cc4f78bc2d262baa4a 100644 --- a/dygraph/ppdet/modeling/heads/yolo_head.py +++ b/dygraph/ppdet/modeling/heads/yolo_head.py @@ -67,38 +67,36 @@ class YOLOv3Head(nn.Layer): assert mask < anchor_num, "anchor mask index overflow" self.mask_anchors[-1].extend(anchors[mask]) - def forward(self, feats): + def forward(self, feats, targets=None): assert len(feats) == len(self.anchors) yolo_outputs = [] for i, feat in enumerate(feats): yolo_output = self.yolo_outputs[i](feat) yolo_outputs.append(yolo_output) - return yolo_outputs - def get_loss(self, inputs, targets): - return self.loss(inputs, targets, self.anchors) - - def get_outputs(self, outputs): - if self.iou_aware: - y = [] - for i, out in enumerate(outputs): - na = len(self.anchors[i]) - ioup, x = out[:, 0:na, :, :], out[:, na:, :, :] - b, c, h, w = x.shape - no = c // na - x = x.reshape((b, na, no, h * w)) - ioup = ioup.reshape((b, na, 1, h * w)) - obj = x[:, :, 4:5, :] - ioup = F.sigmoid(ioup) - obj = F.sigmoid(obj) - obj_t = (obj**(1 - self.iou_aware_factor)) * ( - ioup**self.iou_aware_factor) - obj_t = _de_sigmoid(obj_t) - loc_t = x[:, :, :4, :] - cls_t = x[:, :, 5:, :] - y_t = paddle.concat([loc_t, obj_t, cls_t], axis=2) - y_t = y_t.reshape((b, c, h, w)) - y.append(y_t) - return y + if self.training: + return self.loss(yolo_outputs, targets, self.anchors) else: - return outputs + if self.iou_aware: + y = [] + for i, out in enumerate(yolo_outputs): + na = len(self.anchors[i]) + ioup, x = out[:, 0:na, :, :], out[:, na:, :, :] + b, c, h, w = x.shape + no = c // na + x = x.reshape((b, na, no, h * w)) + ioup = ioup.reshape((b, na, 1, h * w)) + obj = x[:, :, 4:5, :] + ioup = F.sigmoid(ioup) + obj = F.sigmoid(obj) + obj_t = (obj**(1 - self.iou_aware_factor)) * ( + ioup**self.iou_aware_factor) + obj_t = _de_sigmoid(obj_t) + loc_t = x[:, :, :4, :] + cls_t = x[:, :, 5:, :] + y_t = paddle.concat([loc_t, obj_t, cls_t], axis=2) + y_t = y_t.reshape((b, c, h, w)) + y.append(y_t) + return y + else: + return yolo_outputs diff --git a/dygraph/ppdet/modeling/layers.py b/dygraph/ppdet/modeling/layers.py index 4af24110fc6e2c3a633bb52c6b20826e1f370908..2b96e2c984ca3156f48ec8c935c481881f061cf5 100644 --- a/dygraph/ppdet/modeling/layers.py +++ b/dygraph/ppdet/modeling/layers.py @@ -403,7 +403,7 @@ class MatrixNMS(object): self.gaussian_sigma = gaussian_sigma self.background_label = background_label - def __call__(self, bbox, score): + def __call__(self, bbox, score, *args): return ops.matrix_nms( bboxes=bbox, scores=score, @@ -469,7 +469,7 @@ class SSDBox(object): im_shape, scale_factor, var_weight=None): - boxes, scores = preds['boxes'], preds['scores'] + boxes, scores = preds outputs = [] for box, score, prior_box in zip(boxes, scores, prior_boxes): pb_w = prior_box[:, 2] - prior_box[:, 0] + self.norm_delta diff --git a/dygraph/ppdet/modeling/losses/ssd_loss.py b/dygraph/ppdet/modeling/losses/ssd_loss.py index aa1d2c944a5defdefd6e94c45035af971c2a39d9..8561a83cb771d949ae56c7e361face894741a5b3 100644 --- a/dygraph/ppdet/modeling/losses/ssd_loss.py +++ b/dygraph/ppdet/modeling/losses/ssd_loss.py @@ -109,12 +109,11 @@ class SSDLoss(nn.Layer): neg_mask = (idx_rank < num_neg).astype(conf_loss.dtype) return neg_mask - def forward(self, inputs, targets, anchors): - boxes = paddle.concat(inputs['boxes'], axis=1) - scores = paddle.concat(inputs['scores'], axis=1) + def forward(self, boxes, scores, gt_box, gt_class, anchors): + boxes = paddle.concat(boxes, axis=1) + scores = paddle.concat(scores, axis=1) prior_boxes = paddle.concat(anchors, axis=0) - gt_box = targets['gt_bbox'] - gt_label = targets['gt_class'].unsqueeze(-1) + gt_label = gt_class.unsqueeze(-1) batch_size, num_priors, num_classes = scores.shape def _reshape_to_2d(x): diff --git a/dygraph/ppdet/modeling/necks/fpn.py b/dygraph/ppdet/modeling/necks/fpn.py index 0b817a2c11e9df49b5ee71d570edf7e6e69d9dee..99434c42224b9e1295202a2e7392edb61b6e6e2b 100644 --- a/dygraph/ppdet/modeling/necks/fpn.py +++ b/dygraph/ppdet/modeling/necks/fpn.py @@ -23,6 +23,8 @@ from paddle.regularizer import L2Decay from ppdet.core.workspace import register, serializable from ..shape_spec import ShapeSpec +__all__ = ['FPN'] + @register @serializable diff --git a/dygraph/ppdet/modeling/necks/ttf_fpn.py b/dygraph/ppdet/modeling/necks/ttf_fpn.py index 234451dfd0c2c46f3b3537f9ef2bccc3d1141abf..92ada3925db1f94723649d39ee9ae2b75b58578c 100644 --- a/dygraph/ppdet/modeling/necks/ttf_fpn.py +++ b/dygraph/ppdet/modeling/necks/ttf_fpn.py @@ -25,6 +25,8 @@ from ppdet.modeling.layers import DeformableConvV2 import math from ppdet.modeling.ops import batch_norm +__all__ = ['TTFFPN'] + class Upsample(nn.Layer): def __init__(self, ch_in, ch_out, name=None): diff --git a/dygraph/ppdet/modeling/necks/yolo_fpn.py b/dygraph/ppdet/modeling/necks/yolo_fpn.py index 0def906ba9c0b63610d0192a9d0a7261c952a193..4ef6935b3e718b86b70615ec93ddf681e7a19018 100644 --- a/dygraph/ppdet/modeling/necks/yolo_fpn.py +++ b/dygraph/ppdet/modeling/necks/yolo_fpn.py @@ -20,6 +20,10 @@ from ppdet.core.workspace import register, serializable from ..backbones.darknet import ConvBNLayer import numpy as np +from ..shape_spec import ShapeSpec + +__all__ = ['YOLOv3FPN', 'PPYOLOFPN'] + class YoloDetBlock(nn.Layer): def __init__(self, ch_in, channel, norm_type, name): @@ -163,23 +167,30 @@ class PPYOLODetBlock(nn.Layer): class YOLOv3FPN(nn.Layer): __shared__ = ['norm_type'] - def __init__(self, feat_channels=[1024, 768, 384], norm_type='bn'): + def __init__(self, in_channels=[256, 512, 1024], norm_type='bn'): super(YOLOv3FPN, self).__init__() - assert len(feat_channels) > 0, "feat_channels length should > 0" - self.feat_channels = feat_channels - self.num_blocks = len(feat_channels) + assert len(in_channels) > 0, "in_channels length should > 0" + self.in_channels = in_channels + self.num_blocks = len(in_channels) + + self._out_channels = [] self.yolo_blocks = [] self.routes = [] for i in range(self.num_blocks): name = 'yolo_block.{}'.format(i) + in_channel = in_channels[-i - 1] + if i > 0: + in_channel += 512 // (2**i) yolo_block = self.add_sublayer( name, YoloDetBlock( - feat_channels[i], + in_channel, channel=512 // (2**i), norm_type=norm_type, name=name)) self.yolo_blocks.append(yolo_block) + # tip layer output channel doubled + self._out_channels.append(1024 // (2**i)) if i < self.num_blocks - 1: name = 'yolo_transition.{}'.format(i) @@ -211,20 +222,25 @@ class YOLOv3FPN(nn.Layer): return yolo_feats + @classmethod + def from_config(cls, cfg, input_shape): + return {'in_channels': [i.channels for i in input_shape], } + + @property + def out_shape(self): + return [ShapeSpec(channels=c) for c in self._out_channels] + @register @serializable class PPYOLOFPN(nn.Layer): __shared__ = ['norm_type'] - def __init__(self, - feat_channels=[2048, 1280, 640], - norm_type='bn', - **kwargs): + def __init__(self, in_channels=[512, 1024, 2048], norm_type='bn', **kwargs): super(PPYOLOFPN, self).__init__() - assert len(feat_channels) > 0, "feat_channels length should > 0" - self.feat_channels = feat_channels - self.num_blocks = len(feat_channels) + assert len(in_channels) > 0, "in_channels length should > 0" + self.in_channels = in_channels + self.num_blocks = len(in_channels) # parse kwargs self.coord_conv = kwargs.get('coord_conv', False) self.drop_block = kwargs.get('drop_block', False) @@ -246,9 +262,12 @@ class PPYOLOFPN(nn.Layer): else: dropblock_cfg = [] + self._out_channels = [] self.yolo_blocks = [] self.routes = [] - for i, ch_in in enumerate(self.feat_channels): + for i, ch_in in enumerate(self.in_channels[::-1]): + if i > 0: + ch_in += 512 // (2**i) channel = 64 * (2**self.num_blocks) // (2**i) base_cfg = [ # name of layer, Layer, args @@ -279,6 +298,7 @@ class PPYOLOFPN(nn.Layer): name = 'yolo_block.{}'.format(i) yolo_block = self.add_sublayer(name, PPYOLODetBlock(cfg, name)) self.yolo_blocks.append(yolo_block) + self._out_channels.append(channel * 2) if i < self.num_blocks - 1: name = 'yolo_transition.{}'.format(i) route = self.add_sublayer( @@ -307,4 +327,12 @@ class PPYOLOFPN(nn.Layer): route = self.routes[i](route) route = F.interpolate(route, scale_factor=2.) - return yolo_feats \ No newline at end of file + return yolo_feats + + @classmethod + def from_config(cls, cfg, input_shape): + return {'in_channels': [i.channels for i in input_shape], } + + @property + def out_shape(self): + return [ShapeSpec(channels=c) for c in self._out_channels]