【Hackathon 8th No.38】Add RT-DETRv3 for PaddleDetection

configs/rtdetrv3/_base_/optimizer_6x.yml

@@ -0,0 +1,19 @@
+epoch: 72
+
+LearningRate:
+  base_lr: 0.0004
+  schedulers:
+  - !PiecewiseDecay
+    gamma: 1.0
+    milestones: [100]
+    use_warmup: true
+  - !LinearWarmup
+    start_factor: 0.001
+    steps: 2000
+
+OptimizerBuilder:
+  clip_grad_by_norm: 0.1
+  regularizer: false
+  optimizer:
+    type: AdamW
+    weight_decay: 0.0001

configs/rtdetrv3/_base_/rtdetr_reader.yml

@@ -0,0 +1,44 @@
+worker_num: 4
+TrainReader:
+  sample_transforms:
+    - Decode: {}
+    - RandomDistort: {prob: 0.8}
+    - RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
+    - RandomCrop: {prob: 0.8}
+    - RandomFlip: {}
+  batch_transforms:
+    - BatchRandomResize: {target_size: [480, 512, 544, 576, 608, 640, 640, 640, 672, 704, 736, 768, 800], random_size: True, random_interp: True, keep_ratio: False}
+    - NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
+    - NormalizeBox: {retain_origin_box: true}
+    - BboxXYXY2XYWH: {}
+    - Permute: {}
+    - PadGT: {only_origin_box: true}
+  batch_size: 16
+  shuffle: true
+  drop_last: true
+  collate_batch: false
+  use_shared_memory: true
+
+
+EvalReader:
+  sample_transforms:
+    - Decode: {}
+    - Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
+    - NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
+    - Permute: {}
+  batch_size: 16
+  shuffle: false
+  drop_last: false
+
+
+TestReader:
+  inputs_def:
+    image_shape: [3, 640, 640]
+  sample_transforms:
+    - Decode: {}
+    - Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
+    - NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
+    - Permute: {}
+  batch_size: 1
+  shuffle: false
+  drop_last: false

configs/rtdetrv3/_base_/rtdetrv3_r50vd.yml

@@ -0,0 +1,100 @@
+architecture: RTDETRV3
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_vd_ssld_v2_pretrained.pdparams
+norm_type: sync_bn
+use_ema: True
+ema_decay: 0.9999
+ema_decay_type: "exponential"
+ema_filter_no_grad: True
+hidden_dim: 256
+use_focal_loss: True
+eval_size: [640, 640]
+
+
+RTDETRV3:
+  backbone: ResNet
+  neck: HybridEncoder
+  transformer: RTDETRTransformerv3
+  detr_head: DINOv3Head
+  aux_o2m_head: PPYOLOEHead
+  post_process: DETRPostProcess
+
+ResNet:
+  # index 0 stands for res2
+  depth: 50
+  variant: d
+  norm_type: bn
+  freeze_at: 0
+  return_idx: [1, 2, 3]
+  lr_mult_list: [0.1, 0.1, 0.1, 0.1]
+  num_stages: 4
+  freeze_stem_only: True
+
+HybridEncoder:
+  hidden_dim: 256
+  use_encoder_idx: [2]
+  num_encoder_layers: 1
+  encoder_layer:
+    name: TransformerLayer
+    d_model: 256
+    nhead: 8
+    dim_feedforward: 1024
+    dropout: 0.
+    activation: 'gelu'
+  expansion: 1.0
+
+
+RTDETRTransformerv3:
+  num_queries: 300
+  position_embed_type: sine
+  feat_strides: [8, 16, 32]
+  num_levels: 3
+  nhead: 8
+  num_decoder_layers: 6
+  dim_feedforward: 1024
+  dropout: 0.0
+  activation: relu
+  num_denoising: 100
+  label_noise_ratio: 0.5
+  box_noise_scale: 1.0
+  learnt_init_query: False
+  num_noises: 0
+  num_noise_queries: []
+  num_noise_denoising: 100
+  learnt_init_query: False
+
+
+DINOv3Head:
+  o2m: 4
+  loss:
+    name: DINOv3Loss
+    loss_coeff: {class: 1, bbox: 5, giou: 2}
+    aux_loss: True
+    use_vfl: True
+    matcher:
+      name: HungarianMatcher
+      matcher_coeff: {class: 2, bbox: 5, giou: 2}
+
+PPYOLOEHead:
+  fpn_strides: [8, 16, 32]
+  grid_cell_scale: 5.0
+  grid_cell_offset: 0.5
+  static_assigner_epoch: 30
+  use_varifocal_loss: True
+  loss_weight: {class: 1.0, iou: 2.5, dfl: 0.5}
+  static_assigner:
+    name: ATSSAssigner
+    topk: 9
+  assigner:
+    name: TaskAlignedAssigner
+    topk: 13
+    alpha: 1.0
+    beta: 6.0
+  nms:
+    name: MultiClassNMS
+    nms_top_k: 1000
+    keep_top_k: 300
+    score_threshold: 0.01
+    nms_threshold: 0.7
+
+DETRPostProcess:
+  num_top_queries: 300

configs/rtdetrv3/rtdetrv3_r18vd_6x_coco.yml

@@ -0,0 +1,54 @@
+_BASE_: [
+  '../datasets/coco_detection.yml',
+  '../runtime.yml',
+  '_base_/optimizer_6x.yml',
+  '_base_/rtdetrv3_r50vd.yml',
+  '_base_/rtdetr_reader.yml',
+]
+
+weights: output/rtdetrv3_r18vd_6x_coco/model_final
+find_unused_parameters: True
+log_iter: 200
+
+o2m_branch: True
+num_queries_o2m: 450
+
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet18_vd_pretrained.pdparams
+
+RTDETRV3:
+  backbone: ResNet
+  neck: HybridEncoder
+  transformer: RTDETRTransformerv3
+  detr_head: DINOv3Head
+  aux_o2m_head: PPYOLOEHead
+  post_process: DETRPostProcess
+
+ResNet:
+  depth: 18
+  variant: d
+  return_idx: [1, 2, 3]
+  freeze_at: -1
+  freeze_norm: false
+  norm_decay: 0.
+
+HybridEncoder:
+  hidden_dim: 256
+  use_encoder_idx: [2]
+  num_encoder_layers: 1
+  encoder_layer:
+    name: TransformerLayer
+    d_model: 256
+    nhead: 8
+    dim_feedforward: 1024
+    dropout: 0.
+    activation: 'gelu'
+  expansion: 0.5
+  depth_mult: 1.0
+
+RTDETRTransformerv3:
+  eval_idx: -1
+  num_decoder_layers: 3
+  num_noises: 3
+  num_noise_queries: [300, 300, 300]
+  num_noise_denoising: 100
+  learnt_init_query: False

configs/rtdetrv3/rtdetrv3_r34vd_6x_coco.yml

@@ -0,0 +1,54 @@
+_BASE_: [
+  '../datasets/coco_detection.yml',
+  '../runtime.yml',
+  '_base_/optimizer_6x.yml',
+  '_base_/rtdetrv3_r50vd.yml',
+  '_base_/rtdetr_reader.yml',
+]
+
+weights: output/rtdetrv3_r34vd_6x_coco/model_final
+find_unused_parameters: True
+log_iter: 200
+
+o2m_branch: True
+num_queries_o2m: 450
+
+pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/ResNet34_vd_pretrained.pdparams
+
+RTDETRV3:
+  backbone: ResNet
+  neck: HybridEncoder
+  transformer: RTDETRTransformerv3
+  detr_head: DINOv3Head
+  aux_o2m_head: PPYOLOEHead
+  post_process: DETRPostProcess
+
+ResNet:
+  depth: 34
+  variant: d
+  return_idx: [1, 2, 3]
+  freeze_at: -1
+  freeze_norm: false
+  norm_decay: 0.
+
+HybridEncoder:
+  hidden_dim: 256
+  use_encoder_idx: [2]
+  num_encoder_layers: 1
+  encoder_layer:
+    name: TransformerLayer
+    d_model: 256
+    nhead: 8
+    dim_feedforward: 1024
+    dropout: 0.
+    activation: 'gelu'
+  expansion: 0.5
+  depth_mult: 1.0
+
+RTDETRTransformerv3:
+  eval_idx: -1
+  num_decoder_layers: 4
+  num_noises: 3
+  num_noise_queries: [300, 300, 300]
+  num_noise_denoising: 100
+  learnt_init_query: False

configs/rtdetrv3/rtdetrv3_r50vd_6x_coco.yml

@@ -0,0 +1,25 @@
+_BASE_: [
+  '../datasets/coco_detection.yml',
+  '../runtime.yml',
+  '_base_/optimizer_6x.yml',
+  '_base_/rtdetrv3_r50vd.yml',
+  '_base_/rtdetr_reader.yml',
+]
+
+weights: output/rtdetrv3_r50vd_6x_coco/model_final
+find_unused_parameters: True
+log_iter: 200
+
+o2m_branch: True
+num_queries_o2m: 450
+
+pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_vd_ssld_v2_pretrained.pdparams
+
+
+RTDETRTransformerv3:
+  eval_idx: -1
+  num_decoder_layers: 6
+  num_noises: 2
+  num_noise_queries: [300, 300]
+  num_noise_denoising: 100
+  learnt_init_query: False

ppdet/engine/naive_sync_bn.py

@@ -104,4 +104,4 @@ def convert_bn(model):
             bn.set_dict(m.state_dict())
             setattr(model, n, bn)
         else:
-            convert_bn(m)
+            convert_bn(m)

ppdet/modeling/architectures/__init__.py

@@ -83,3 +83,6 @@
 from .detr_ssod import *
 from .multi_stream_detector import *
 from .clrnet import *
+
+from . import rtdetrv3
+from .rtdetrv3 import *