update ssd infer bug

Author: CycloneBoy

configs/ssd/_base_/ssd_mobilenet_reader.yml

@@ -6,7 +6,7 @@ TrainReader:
     - Decode: {}
     - RandomDistort: {brightness: [0.5, 1.125, 0.875], random_apply: False}
     - RandomExpand: {fill_value: [127.5, 127.5, 127.5]}
-    - RandomCrop: {allow_no_crop: Fasle}
+    - RandomCrop: {allow_no_crop: False}
     - RandomFlip: {}
     - Resize: {target_size: [300, 300], keep_ratio: False, interp: 1}
     - NormalizeBox: {}

configs/ssd/_base_/ssd_vgg16_300.yml

@@ -1,7 +1,7 @@
 architecture: SSD
 pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/VGG16_caffe_pretrained.pdparams
 
-# Model Achitecture
+# Model Architecture
 SSD:
   # model feat info flow
   backbone: VGG

ppdettorch/modeling/architectures/ssd.py

@@ -75,9 +75,12 @@ def _forward(self):
                                  self.inputs['gt_class'])
         else:
             preds, anchors = self.ssd_head(body_feats, self.inputs['image'])
-            bbox, bbox_num = self.post_process(preds, anchors,
-                                               self.inputs['im_shape'],
-                                               self.inputs['scale_factor'])
+
+            bbox, bbox_num, before_nms_indexes = self.post_process(preds, anchors,
+                                                                   self.inputs['im_shape'],
+                                                                   self.inputs['scale_factor'])
+
+            return bbox, bbox_num
             return bbox, bbox_num
 
     def get_loss(self, ):

ppdettorch/modeling/backbones/__init__.py

@@ -5,7 +5,7 @@
 # @Author  ：sl
 # @Date    ：2022/11/1 14:45
 
-# from . import vgg
+from . import vgg
 from . import resnet
 from . import darknet
 from . import mobilenet_v1
@@ -32,7 +32,7 @@
 # from . import vision_transformer
 # from . import mobileone
 
-# from .vgg import *
+from .vgg import *
 from .resnet import *
 from .darknet import *
 from .mobilenet_v1 import *

ppdettorch/modeling/backbones/mobilenet_v3.py

@@ -202,14 +202,15 @@ def __init__(self,
                  norm_type='bn',
                  norm_decay=0.,
                  freeze_norm=False,
-                 name=None):
+                 name=None,
+                 padding=3):
         super(ExtraBlockDW, self).__init__()
         self.pointwise_conv = ConvBNLayer(
             in_c=in_c,
             out_c=ch_1,
             filter_size=1,
             stride=1,
-            padding='SAME',
+            padding=0,
             act='relu6',
             lr_mult=lr_mult,
             conv_decay=conv_decay,
@@ -222,7 +223,7 @@ def __init__(self,
             out_c=ch_2,
             filter_size=3,
             stride=stride,
-            padding='SAME',
+            padding=padding,  # TODO: same padding
             num_groups=int(ch_1),
             act='relu6',
             lr_mult=lr_mult,
@@ -236,7 +237,7 @@ def __init__(self,
             out_c=ch_2,
             filter_size=1,
             stride=1,
-            padding='SAME',
+            padding=0,
             act='relu6',
             lr_mult=lr_mult,
             conv_decay=conv_decay,
@@ -412,7 +413,8 @@ def __init__(
                                           norm_type=norm_type,
                                           norm_decay=norm_decay,
                                           freeze_norm=freeze_norm,
-                                          name='conv' + str(i + 2))
+                                          name='conv' + str(i + 2),
+                                          padding=3)  # TODO: calc same padding
                 self.add_module("conv" + str(i + 2), module=conv_extra)
                 self.extra_block_list.append(conv_extra)
                 i += 1

ppdettorch/modeling/backbones/vgg.py

@@ -4,7 +4,6 @@
 import torch.nn as nn
 import torch.nn.functional as F
 
-
 from torch.nn import Conv2d, MaxPool2d
 from ppdettorch.core.workspace import register, serializable
 from ..shape_spec import ShapeSpec
@@ -34,21 +33,18 @@ def __init__(self,
             padding=1)
         self.conv_out_list = []
         for i in range(1, groups):
-            conv_out = self.add_sublayer(
-                'conv{}'.format(i),
-                Conv2d(
-                    in_channels=out_channels,
-                    out_channels=out_channels,
-                    kernel_size=3,
-                    stride=1,
-                    padding=1))
+            conv_out = Conv2d(in_channels=out_channels,
+                              out_channels=out_channels,
+                              kernel_size=3,
+                              stride=1,
+                              padding=1)
+            self.add_module('conv{}'.format(i), conv_out)
             self.conv_out_list.append(conv_out)
 
-        self.pool = MaxPool2d(
-            kernel_size=pool_size,
-            stride=pool_stride,
-            padding=pool_padding,
-            ceil_mode=True)
+        self.pool = MaxPool2d(kernel_size=pool_size,
+                              stride=pool_stride,
+                              padding=pool_padding,
+                              ceil_mode=True)
 
     def forward(self, inputs):
         out = self.conv0(inputs)
@@ -95,12 +91,10 @@ def forward(self, inputs):
 class L2NormScale(nn.Module):
     def __init__(self, num_channels, scale=1.0):
         super(L2NormScale, self).__init__()
-        self.scale = self.create_parameter(
-            attr=ParamAttr(initializer=paddle.nn.initializer.Constant(scale)),
-            shape=[num_channels])
+        self.scale = nn.Parameter(torch.ones([num_channels]), requires_grad=False)
 
     def forward(self, inputs):
-        out = F.normalize(inputs, axis=1, epsilon=1e-10)
+        out = F.normalize(inputs, dim=1, eps=1e-10)
         # out = self.scale.unsqueeze(0).unsqueeze(2).unsqueeze(3).expand_as(
         #     out) * out
         out = self.scale.unsqueeze(0).unsqueeze(2).unsqueeze(3) * out
@@ -119,7 +113,7 @@ def __init__(self,
         super(VGG, self).__init__()
 
         assert depth in [16, 19], \
-                "depth as 16/19 supported currently, but got {}".format(depth)
+            "depth as 16/19 supported currently, but got {}".format(depth)
         self.depth = depth
         self.groups = VGG_cfg[depth]
         self.normalizations = normalizations
@@ -159,19 +153,18 @@ def __init__(self,
         last_channels = 1024
         for i, v in enumerate(self.extra_block_filters):
             assert len(v) == 5, "extra_block_filters size not fix"
-            extra_conv = self.add_sublayer("conv{}".format(6 + i),
-                                           ExtraBlock(last_channels, v[0], v[1],
-                                                      v[2], v[3], v[4]))
+            extra_conv = ExtraBlock(last_channels, v[0], v[1],
+                                    v[2], v[3], v[4])
+            self.add_module("conv{}".format(6 + i), extra_conv)
             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):
             if n != -1:
-                norm = self.add_sublayer("norm{}".format(i),
-                                         L2NormScale(
-                                             self.extra_block_filters[i][1], n))
+                norm = L2NormScale(self.extra_block_filters[i][1], n)
+                self.add_module("norm{}".format(i), norm)
             else:
                 norm = None
             self.norms.append(norm)

ppdettorch/modeling/bbox_utils.py

@@ -885,7 +885,7 @@ def ssd_prior_box_np(
             Default: None, means [] and will not be used.
        aspect_ratios (list|tuple|float, optional): the aspect ratios of generated
             prior boxes. Default: [1.0].
-       variances (list|tuple, optional): the variances to be encoded in prior boxes.
+       variance (list|tuple, optional): the variances to be encoded in prior boxes.
             Default:[0.1, 0.1, 0.2, 0.2].
        flip (bool): Whether to flip aspect ratios. Default:False.
        clip (bool): Whether to clip out-of-boundary boxes. Default: False.
@@ -941,6 +941,10 @@ def ssd_prior_box_np(
             if flip:
                 real_aspect_ratios.append(1.0 / ar)
 
+    if step_w == 0 or step_h == 0:
+        step_w = image_w / layer_w
+        step_h = image_h / layer_h
+
     num_priors = len(real_aspect_ratios) * len(min_sizes)
     if max_sizes is None:
         max_sizes = []

ppdettorch/modeling/layers.py

@@ -258,29 +258,29 @@ def __init__(self,
             stride=stride,
             groups=in_channels,
             norm_type=norm_type,
-            )
+        )
         conv2 = ConvNormLayer(
             in_channels,
             out_channels,
             filter_size=1,
             stride=stride,
             norm_type=norm_type,
-            )
+        )
         conv3 = ConvNormLayer(
             out_channels,
             out_channels,
             filter_size=1,
             stride=stride,
             norm_type=norm_type,
-            )
+        )
         conv4 = ConvNormLayer(
             out_channels,
             out_channels,
             filter_size=5,
             stride=stride,
             groups=out_channels,
             norm_type=norm_type,
-            )
+        )
         conv_list = [conv1, conv2, conv3, conv4]
         self.lite_conv.add_sublayer('conv1', conv1)
         self.lite_conv.add_sublayer('relu6_1', nn.ReLU6())
@@ -627,6 +627,7 @@ def __call__(self,
 
         return yolo_boxes, yolo_scores
 
+
 class YOLOLayer(nn.Module):
     """Detection layer"""
 
@@ -720,7 +721,7 @@ def __call__(self,
             output_boxes *= im_shape
         else:
             output_boxes[..., -2:] -= 1.0
-        output_scores = F.softmax(torch.concat(scores, dim=1)).permute(0, 2, 1)
+        output_scores = F.softmax(torch.concat(scores, dim=1), dim=-1).permute(0, 2, 1)
 
         return output_boxes, output_scores
 

ppdettorch/modeling/post_process.py

@@ -62,7 +62,7 @@ def __call__(self, head_out, rois, im_shape, scale_factor):
         """
         if self.nms is not None:
             bboxes, score = self.decode(head_out, rois, im_shape, scale_factor)
-            bbox_pred, bbox_num, _ = self.nms(bboxes, score, self.num_classes)
+            bbox_pred, bbox_num, before_nms_indexes = self.nms(bboxes, score, self.num_classes)
 
         else:
             bbox_pred, bbox_num = self.decode(head_out, rois, im_shape,
@@ -77,7 +77,10 @@ def __call__(self, head_out, rois, im_shape, scale_factor):
             bbox_pred = torch.concat([bbox_pred, fake_bboxes])
             bbox_num = bbox_num + 1
 
-        return bbox_pred, bbox_num
+        if self.nms is not None:
+            return bbox_pred, bbox_num, before_nms_indexes
+        else:
+            return bbox_pred, bbox_num
 
     def get_pred(self, bboxes, bbox_num, im_shape, scale_factor):
         """

tests/process/infer/run_detection_infer.py

@@ -163,7 +163,10 @@ def run_picodet_coco(self, config_name=None):
         # config_name = f"yolov8_n_500e_coco.yml"
 
         # ssd
-        config_name = f"ssd_mobilenet_v1_300_120e_voc.yml"
+        # config_name = f"ssd_mobilenet_v1_300_120e_voc.yml"
+        config_name = f"ssd_vgg16_300_240e_voc.yml"
+        # config_name = f"ssdlite_mobilenet_v3_large_320_coco.yml"
+        # config_name = f"ssdlite_mobilenet_v3_small_320_coco.yml"
 
 
         # run_arg = DetectionInferUtils.init_args()
@@ -176,6 +179,8 @@ def run_picodet_coco(self, config_name=None):
             model_class = "picodet"
         elif "ppyoloe" in config_name:
             model_class = "ppyoloe"
+        elif "ssdlite_" in config_name:
+            model_class = "ssd"
         else:
             config_name_end_index = FileUtils.get_file_name(config_name).find("_")
             model_class = config_name[:config_name_end_index]
@@ -229,14 +234,27 @@ def run_picodet_coco_batch():
     # model_class = "yolov6"
     # model_class = "yolov7"
     # model_class = "rtmdet"
-    model_class = "yolov8"
+    # model_class = "yolov8"
+    model_class = "ssd"
 
     with_application = False
     # with_application = True
 
     # do_transform = False
     do_transform = True
 
+    # 需要跳过执行验证的列表
+    skip_config_name_dict = {
+        "yolov3": [
+            "yolov3_darknet53_original_270e_coco.yml",
+            "yolov3_mobilenet_v1_roadsign.yml"
+        ],
+        "ssd": [
+            "ssd_r34_70e_coco.yml",
+            "ssdlite_ghostnet_320_coco.yml"
+        ]
+    }
+
     # base_dir = f"/home/mqq/shenglei/ocr/PaddleDetection/configs/{model_class}"
     base_dir = f"{Constants.WORK_DIR}/configs/{model_class}"
     if with_application:
@@ -249,14 +267,20 @@ def run_picodet_coco_batch():
                                                 end_with=".yml", )
 
     logger.info(f"total: {len(file_name_list)}")
-    skip = 0
+
+    skip_config_name_list = skip_config_name_dict.get(model_class, [])
+    skip = 1
     detection_runner = DetectionRunInfer()
 
     for index, file_name in enumerate(file_name_list):
         if index < skip:
             logger.info(f"跳过已经执行的：{index} - {file_name}")
             continue
 
+        if f"{FileUtils.get_file_name(file_name)}.yml" in skip_config_name_list:
+            logger.info(f"跳过无需测试的：{index} - {file_name}")
+            continue
+
         if "_xpu" in file_name:
             continue
         if "ppyoloe_crn_m_80e_pcb" in file_name:
@@ -269,5 +293,5 @@ def run_picodet_coco_batch():
 
 
 if __name__ == '__main__':
-    demo_run_detection_infer()
-    # run_picodet_coco_batch()
+    # demo_run_detection_infer()
+    run_picodet_coco_batch()