Add lite hr net (PaddlePaddle#3793)

* add LiteHRNet backbone and config .YML * test lite18-network param acc is same with ori-model 1. fix default darkpose=ON, 2. += is not inplace add new keypoint model Lite-HRNet * add new keypoint model Lite-HRNet * 1. Add description of network type; 2. use channel_shuffle in ops.py * use normal to init conv2d * add network type description
cjxxu · Aug 4, 2021 · d578628 · d578628
1 parent 55fcc1f
commit d578628
Show file tree

Hide file tree

Showing 8 changed files with 1,191 additions and 24 deletions.
diff --git a/configs/keypoint/lite_hrnet/lite_hrnet_18_256x192_coco.yml b/configs/keypoint/lite_hrnet/lite_hrnet_18_256x192_coco.yml
@@ -0,0 +1,140 @@
+use_gpu: true
+log_iter: 5
+save_dir: output
+snapshot_epoch: 10
+weights: output/lite_hrnet_18_256x192_coco/model_final
+epoch: 210
+num_joints: &num_joints 17
+pixel_std: &pixel_std 200
+metric: KeyPointTopDownCOCOEval
+num_classes: 1
+train_height: &train_height 256
+train_width: &train_width 192
+trainsize: &trainsize [*train_width, *train_height]
+hmsize: &hmsize [48, 64]
+flip_perm: &flip_perm [[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12], [13, 14], [15, 16]]
+
+
+#####model
+architecture: TopDownHRNet
+
+TopDownHRNet:
+  backbone: LiteHRNet
+  post_process: HRNetPostProcess
+  flip_perm: *flip_perm
+  num_joints: *num_joints
+  width: &width 40
+  loss: KeyPointMSELoss
+  use_dark: false
+
+LiteHRNet:
+  network_type: lite_18
+  freeze_at: -1
+  freeze_norm: false
+  return_idx: [0]
+
+KeyPointMSELoss:
+  use_target_weight: true
+  loss_scale: 1.0
+
+#####optimizer
+LearningRate:
+  base_lr: 0.002
+  schedulers:
+  - !PiecewiseDecay
+    milestones: [170, 200]
+    gamma: 0.1
+  - !LinearWarmup
+    start_factor: 0.001
+    steps: 500
+
+OptimizerBuilder:
+  optimizer:
+    type: Adam
+  regularizer:
+    factor: 0.0
+    type: L2
+
+
+#####data
+TrainDataset:
+  !KeypointTopDownCocoDataset
+    image_dir: train2017
+    anno_path: annotations/person_keypoints_train2017.json
+    dataset_dir: dataset/coco
+    num_joints: *num_joints
+    trainsize: *trainsize
+    pixel_std: *pixel_std
+    use_gt_bbox: True
+
+
+EvalDataset:
+  !KeypointTopDownCocoDataset
+    image_dir: val2017
+    anno_path: annotations/person_keypoints_val2017.json
+    dataset_dir: dataset/coco
+    num_joints: *num_joints
+    trainsize: *trainsize
+    pixel_std: *pixel_std
+    use_gt_bbox: True
+    image_thre: 0.0
+
+
+TestDataset:
+  !ImageFolder
+    anno_path: dataset/coco/keypoint_imagelist.txt
+
+worker_num: 2
+global_mean: &global_mean [0.485, 0.456, 0.406]
+global_std: &global_std [0.229, 0.224, 0.225]
+TrainReader:
+  sample_transforms:
+    - RandomFlipHalfBodyTransform:
+        scale: 0.25
+        rot: 30
+        num_joints_half_body: 8
+        prob_half_body: 0.3
+        pixel_std: *pixel_std
+        trainsize: *trainsize
+        upper_body_ids: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
+        flip_pairs: *flip_perm
+    - TopDownAffine:
+        trainsize: *trainsize
+    - ToHeatmapsTopDown:
+        hmsize: *hmsize
+        sigma: 2
+  batch_transforms:
+    - NormalizeImage:
+        mean: *global_mean
+        std: *global_std
+        is_scale: true
+    - Permute: {}
+  batch_size: 64
+  shuffle: true
+  drop_last: false
+
+EvalReader:
+  sample_transforms:
+    - TopDownAffine:
+        trainsize: *trainsize
+  batch_transforms:
+    - NormalizeImage:
+        mean: *global_mean
+        std: *global_std
+        is_scale: true
+    - Permute: {}
+  batch_size: 16
+
+TestReader:
+  inputs_def:
+    image_shape: [3, *train_height, *train_width]
+  sample_transforms:
+    - Decode: {}
+    - TopDownEvalAffine:
+        trainsize: *trainsize
+    - NormalizeImage:
+        mean: *global_mean
+        std: *global_std
+        is_scale: true
+    - Permute: {}
+  batch_size: 1
diff --git a/configs/keypoint/lite_hrnet/lite_hrnet_30_256x192_coco.yml b/configs/keypoint/lite_hrnet/lite_hrnet_30_256x192_coco.yml
@@ -0,0 +1,140 @@
+use_gpu: true
+log_iter: 5
+save_dir: output
+snapshot_epoch: 10
+weights: output/lite_hrnet_30_256x192_coco/model_final
+epoch: 210
+num_joints: &num_joints 17
+pixel_std: &pixel_std 200
+metric: KeyPointTopDownCOCOEval
+num_classes: 1
+train_height: &train_height 256
+train_width: &train_width 192
+trainsize: &trainsize [*train_width, *train_height]
+hmsize: &hmsize [48, 64]
+flip_perm: &flip_perm [[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12], [13, 14], [15, 16]]
+
+
+#####model
+architecture: TopDownHRNet
+
+TopDownHRNet:
+  backbone: LiteHRNet
+  post_process: HRNetPostProcess
+  flip_perm: *flip_perm
+  num_joints: *num_joints
+  width: &width 40
+  loss: KeyPointMSELoss
+  use_dark: false
+
+LiteHRNet:
+  network_type: lite_30
+  freeze_at: -1
+  freeze_norm: false
+  return_idx: [0]
+
+KeyPointMSELoss:
+  use_target_weight: true
+  loss_scale: 1.0
+
+#####optimizer
+LearningRate:
+  base_lr: 0.002
+  schedulers:
+  - !PiecewiseDecay
+    milestones: [170, 200]
+    gamma: 0.1
+  - !LinearWarmup
+    start_factor: 0.001
+    steps: 500
+
+OptimizerBuilder:
+  optimizer:
+    type: Adam
+  regularizer:
+    factor: 0.0
+    type: L2
+
+
+#####data
+TrainDataset:
+  !KeypointTopDownCocoDataset
+    image_dir: train2017
+    anno_path: annotations/person_keypoints_train2017.json
+    dataset_dir: dataset/coco
+    num_joints: *num_joints
+    trainsize: *trainsize
+    pixel_std: *pixel_std
+    use_gt_bbox: True
+
+
+EvalDataset:
+  !KeypointTopDownCocoDataset
+    image_dir: val2017
+    anno_path: annotations/person_keypoints_val2017.json
+    dataset_dir: dataset/coco
+    num_joints: *num_joints
+    trainsize: *trainsize
+    pixel_std: *pixel_std
+    use_gt_bbox: True
+    image_thre: 0.0
+
+
+TestDataset:
+  !ImageFolder
+    anno_path: dataset/coco/keypoint_imagelist.txt
+
+worker_num: 4
+global_mean: &global_mean [0.485, 0.456, 0.406]
+global_std: &global_std [0.229, 0.224, 0.225]
+TrainReader:
+  sample_transforms:
+    - RandomFlipHalfBodyTransform:
+        scale: 0.25
+        rot: 30
+        num_joints_half_body: 8
+        prob_half_body: 0.3
+        pixel_std: *pixel_std
+        trainsize: *trainsize
+        upper_body_ids: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
+        flip_pairs: *flip_perm
+    - TopDownAffine:
+        trainsize: *trainsize
+    - ToHeatmapsTopDown:
+        hmsize: *hmsize
+        sigma: 2
+  batch_transforms:
+    - NormalizeImage:
+        mean: *global_mean
+        std: *global_std
+        is_scale: true
+    - Permute: {}
+  batch_size: 64
+  shuffle: true
+  drop_last: false
+
+EvalReader:
+  sample_transforms:
+    - TopDownAffine:
+        trainsize: *trainsize
+  batch_transforms:
+    - NormalizeImage:
+        mean: *global_mean
+        std: *global_std
+        is_scale: true
+    - Permute: {}
+  batch_size: 16
+
+TestReader:
+  inputs_def:
+    image_shape: [3, *train_height, *train_width]
+  sample_transforms:
+    - Decode: {}
+    - TopDownEvalAffine:
+        trainsize: *trainsize
+    - NormalizeImage:
+        mean: *global_mean
+        std: *global_std
+        is_scale: true
+    - Permute: {}
+  batch_size: 1
diff --git a/ppdet/modeling/architectures/keypoint_hrnet.py b/ppdet/modeling/architectures/keypoint_hrnet.py
@@ -41,18 +41,20 @@ def __init__(self,
                  post_process='HRNetPostProcess',
                  flip_perm=None,
                  flip=True,
-                 shift_heatmap=True):
+                 shift_heatmap=True,
+                 use_dark=True):
         """
-        HRNnet network, see https://arxiv.org/abs/1902.09212
+        HRNet network, see https://arxiv.org/abs/1902.09212
 
         Args:
             backbone (nn.Layer): backbone instance
             post_process (object): `HRNetPostProcess` instance
             flip_perm (list): The left-right joints exchange order list
+            use_dark(bool): Whether to use DARK in post processing
         """
         super(TopDownHRNet, self).__init__()
         self.backbone = backbone
-        self.post_process = HRNetPostProcess()
+        self.post_process = HRNetPostProcess(use_dark)
         self.loss = loss
         self.flip_perm = flip_perm
         self.flip = flip
@@ -218,7 +220,6 @@ def get_final_preds(self, heatmaps, center, scale, kernelsize=3):
             preds: numpy.ndarray([batch_size, num_joints, 2]), keypoints coords
             maxvals: numpy.ndarray([batch_size, num_joints, 1]), the maximum confidence of the keypoints
         """
-
         coords, maxvals = self.get_max_preds(heatmaps)
 
         heatmap_height = heatmaps.shape[2]

diff --git a/ppdet/modeling/backbones/__init__.py b/ppdet/modeling/backbones/__init__.py
@@ -18,6 +18,7 @@
 from . import mobilenet_v1
 from . import mobilenet_v3
 from . import hrnet
+from . import lite_hrnet
 from . import blazenet
 from . import ghostnet
 from . import senet
@@ -31,6 +32,7 @@
 from .mobilenet_v1 import *
 from .mobilenet_v3 import *
 from .hrnet import *
+from .lite_hrnet import *
 from .blazenet import *
 from .ghostnet import *
 from .senet import *