[WIP] Add mask rcnn and yolo in vitdet (#7187)

* fix fpn

* add vitdet mask and yolo

* add vityolo

configs/vitdet/_base_/mask_rcnn_reader.yml

+worker_num: 2
+TrainReader:
+  sample_transforms:
+  - Decode: {}
+  # - RandomResizeCrop: {resizes: [400, 500, 600], cropsizes: [[384, 600], ], prob: 0.5}
+  - RandomResize: {target_size: [[640, 1333], [672, 1333], [704, 1333], [736, 1333], [768, 1333], [800, 1333]], interp: 2, keep_ratio: True}
+  - RandomFlip: {prob: 0.5}
+  - NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
+  - Permute: {}
+  batch_transforms:
+  - PadBatch: {pad_to_stride: 32}
+  batch_size: 1
+  shuffle: true
+  drop_last: true
+  collate_batch: false
+  use_shared_memory: true
+
+EvalReader:
+  sample_transforms:
+  - Decode: {}
+  - Resize: {interp: 2, target_size: [800, 1333], keep_ratio: True}
+  - NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
+  - Permute: {}
+  batch_transforms:
+  - PadBatch: {pad_to_stride: 32}
+  batch_size: 1
+  shuffle: false
+  drop_last: false
+
+
+TestReader:
+  sample_transforms:
+  - Decode: {}
+  - Resize: {interp: 2, target_size: [800, 1333], keep_ratio: True}
+  - NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
+  - Permute: {}
+  batch_transforms:
+  - PadBatch: {pad_to_stride: 32}
+  batch_size: 1
+  shuffle: false
+  drop_last: false

configs/vitdet/_base_/ppyoloe_reader.yml

+worker_num: 4
+eval_height: &eval_height 640
+eval_width: &eval_width 640
+eval_size: &eval_size [*eval_height, *eval_width]
+
+TrainReader:
+  sample_transforms:
+    - Decode: {}
+    - RandomDistort: {}
+    - RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
+    - RandomCrop: {}
+    - RandomFlip: {}
+  batch_transforms:
+    - BatchRandomResize: {target_size: [320, 352, 384, 416, 448, 480, 512, 544, 576, 608, 640, 672, 704, 736, 768], random_size: True, random_interp: True, keep_ratio: False}
+    - NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
+    - Permute: {}
+    - PadGT: {}
+  batch_size: 2
+  shuffle: true
+  drop_last: true
+  use_shared_memory: true
+  collate_batch: true
+
+EvalReader:
+  sample_transforms:
+    - Decode: {}
+    - Resize: {target_size: *eval_size, keep_ratio: False, interp: 2}
+    - NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
+    - Permute: {}
+  batch_size: 2
+
+TestReader:
+  inputs_def:
+    image_shape: [3, *eval_height, *eval_width]
+  sample_transforms:
+    - Decode: {}
+    - Resize: {target_size: *eval_size, keep_ratio: False, interp: 2}
+    - NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
+    - Permute: {}
+  batch_size: 1

configs/vitdet/cascade_rcnn_vit_base_hrfpn_cae_1x_coco.yml

@@ -2,7 +2,7 @@
 _BASE_: [
   '../datasets/coco_detection.yml',
   '../runtime.yml',
-  './_base_/reader.yml',
+  './_base_/faster_rcnn_reader.yml',
   './_base_/optimizer_base_1x.yml'
 ]
 

configs/vitdet/mask_rcnn_vit_base_hrfpn_cae_1x_coco.yml

+_BASE_: [
+  '../datasets/coco_instance.yml',
+  '../runtime.yml',
+  './_base_/mask_rcnn_reader.yml',
+  './_base_/optimizer_base_1x.yml'
+]
+
+weights: output/mask_rcnn_vit_base_hrfpn_cae_1x_coco/model_final
+
+
+# runtime
+log_iter: 100
+snapshot_epoch: 1
+norm_type: sync_bn
+use_fused_allreduce_gradients: &use_checkpoint False
+
+
+architecture: MaskRCNN
+MaskRCNN:
+  backbone: VisionTransformer
+  neck: HRFPN
+  rpn_head: RPNHead
+  bbox_head: BBoxHead
+  mask_head: MaskHead
+  # post process
+  bbox_post_process: BBoxPostProcess
+  mask_post_process: MaskPostProcess
+
+VisionTransformer:
+  patch_size: 16
+  embed_dim: 768
+  depth: 12
+  num_heads: 12
+  mlp_ratio: 4
+  qkv_bias: True
+  drop_rate: 0.0
+  drop_path_rate: 0.2
+  init_values: 0.1
+  final_norm: False
+  use_rel_pos_bias: False
+  use_sincos_pos_emb: True
+  epsilon: 0.000001 # 1e-6
+  out_indices: [3, 5, 7, 11]
+  with_fpn: True
+  use_checkpoint: *use_checkpoint
+  pretrained: https://bj.bcebos.com/v1/paddledet/models/pretrained/vit_base_cae_pretrained.pdparams
+
+HRFPN:
+  out_channel: 256
+  use_bias: True
+
+RPNHead:
+  anchor_generator:
+    aspect_ratios: [0.5, 1.0, 2.0]
+    anchor_sizes: [[32], [64], [128], [256], [512]]
+    strides: [4, 8, 16, 32, 64]
+  rpn_target_assign:
+    batch_size_per_im: 256
+    fg_fraction: 0.5
+    negative_overlap: 0.3
+    positive_overlap: 0.7
+    use_random: True
+  train_proposal:
+    min_size: 0.0
+    nms_thresh: 0.7
+    pre_nms_top_n: 2000
+    post_nms_top_n: 1000
+    topk_after_collect: True
+  test_proposal:
+    min_size: 0.0
+    nms_thresh: 0.7
+    pre_nms_top_n: 1000
+    post_nms_top_n: 1000
+  loss_rpn_bbox: SmoothL1Loss
+
+SmoothL1Loss:
+  beta: 0.1111111111111111
+
+
+BBoxHead:
+  head: XConvNormHead
+  roi_extractor:
+    resolution: 7
+    sampling_ratio: 0
+    aligned: True
+  bbox_assigner: BBoxAssigner
+  loss_normalize_pos: True
+  bbox_loss: GIoULoss
+
+BBoxAssigner:
+  batch_size_per_im: 512
+  bg_thresh: 0.5
+  fg_thresh: 0.5
+  fg_fraction: 0.25
+  use_random: True
+
+
+XConvNormHead:
+    num_convs: 4
+    norm_type: bn
+
+GIoULoss:
+  loss_weight: 10.
+  reduction: 'none'
+  eps: 0.000001
+
+
+
+BBoxPostProcess:
+  decode: RCNNBox
+  nms:
+    name: MultiClassNMS
+    keep_top_k: 100
+    score_threshold: 0.05
+    nms_threshold: 0.5
+
+MaskHead:
+  head: MaskFeat
+  roi_extractor:
+    resolution: 14
+    sampling_ratio: 0
+    aligned: True
+  mask_assigner: MaskAssigner
+  share_bbox_feat: False
+
+MaskFeat:
+  num_convs: 4
+  out_channel: 256
+  norm_type: ~
+
+MaskAssigner:
+  mask_resolution: 28
+
+MaskPostProcess:
+  binary_thresh: 0.5

configs/vitdet/ppyoloe_vit_base_csppan_cae_30e_coco.yml

+
+_BASE_: [
+  '../datasets/coco_detection.yml',
+  '../runtime.yml',
+  './_base_/ppyoloe_reader.yml',
+  './_base_/optimizer_base_30e.yml'
+]
+
+weights: output/ppyoloe_vit_base_csppan_cae_30e_coco/model_final
+
+
+snapshot_epoch: 2
+log_iter: 100
+
+
+use_ema: true
+ema_decay: 0.9999
+ema_skip_names: ['yolo_head.proj_conv.weight', 'backbone.pos_embed']
+custom_black_list: ['reduce_mean']
+use_fused_allreduce_gradients: &use_checkpoint False
+
+
+architecture: YOLOv3
+norm_type: sync_bn
+
+YOLOv3:
+  backbone: VisionTransformer
+  neck: YOLOCSPPAN
+  yolo_head: PPYOLOEHead
+  post_process: ~
+
+VisionTransformer:
+  patch_size: 16
+  embed_dim: 768
+  depth: 12
+  num_heads: 12
+  mlp_ratio: 4
+  qkv_bias: True
+  drop_rate: 0.0
+  drop_path_rate: 0.2
+  init_values: 0.1
+  final_norm: False
+  use_rel_pos_bias: False
+  use_sincos_pos_emb: True
+  epsilon: 0.000001 # 1e-6
+  out_indices: [11, ]
+  with_fpn: True
+  num_fpn_levels: 3
+  out_with_norm: False
+  use_checkpoint: *use_checkpoint
+  pretrained: https://bj.bcebos.com/v1/paddledet/models/pretrained/vit_base_cae_pretrained.pdparams
+
+YOLOCSPPAN:
+  in_channels: [768, 768, 768]
+  act: 'silu'
+
+PPYOLOEHead:
+  fpn_strides: [8, 16, 32]
+  in_channels: [768, 768, 768]
+  static_assigner_epoch: -1
+  grid_cell_scale: 5.0
+  grid_cell_offset: 0.5
+  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

ppdet/modeling/backbones/vision_transformer.py

@@ -340,6 +340,7 @@ class VisionTransformer(nn.Layer):
                  use_abs_pos_emb=False,
                  use_sincos_pos_emb=True,
                  with_fpn=True,
+                 num_fpn_levels=4,
                  use_checkpoint=False,
                  **args):
         super().__init__()
@@ -350,6 +351,8 @@ class VisionTransformer(nn.Layer):
         self.use_sincos_pos_emb = use_sincos_pos_emb
         self.use_rel_pos_bias = use_rel_pos_bias
         self.final_norm = final_norm
+        self.out_indices = out_indices
+        self.num_fpn_levels = num_fpn_levels
 
         if use_checkpoint:
             paddle.seed(0)
@@ -415,14 +418,15 @@ class VisionTransformer(nn.Layer):
 
         assert len(out_indices) <= 4, ''
         self.out_indices = out_indices
-        self.out_channels = [embed_dim for _ in range(len(out_indices))]
-        self.out_strides = [4, 8, 16, 32][-len(out_indices):] if with_fpn else [
-            8 for _ in range(len(out_indices))
+        self.out_channels = [embed_dim for _ in range(num_fpn_levels)]
+        self.out_strides = [4, 8, 16, 32][-num_fpn_levels:] if with_fpn else [
+            patch_size for _ in range(len(out_indices))
         ]
 
         self.norm = Identity()
 
         if self.with_fpn:
+            assert num_fpn_levels <= 4, ''
             self.init_fpn(
                 embed_dim=embed_dim,
                 patch_size=patch_size, )
@@ -611,9 +615,15 @@ class VisionTransformer(nn.Layer):
                 feats.append(xp)
 
         if self.with_fpn:
-            fpns = [self.fpn1, self.fpn2, self.fpn3, self.fpn4]
-            for i in range(len(feats)):
-                feats[i] = fpns[i](feats[i])
+            fpns = [self.fpn1, self.fpn2, self.fpn3, self.fpn4][
+                -self.num_fpn_levels:]
+            assert len(fpns) == len(feats) or len(feats) == 1, ''
+            outputs = []
+            for i, m in enumerate(fpns):
+                outputs.append(
+                    m(feats[i] if len(feats) == len(fpns) else feats[-1]))
+
+            return outputs
 
         return feats