关于静态图中 MultiScaleTest时遇到的问题

我在使用Cascade RCNN的结构，在测试时发现 multi scale的效果会变得很差很差，而single scale每个scale的效果都还算可以，在debug过程中我有如下几个问题:

1. 为什么mutli scale过程中，每个尺度调用 self.bbox_head.get_prediction 的结果个数都为1000，而single scale 中调用 self.bbox_head.get_prediction的过程中结果为NMS之后的个数。请问这是什么原因造成mutli scale中bbox_head中的nms没有生效的呢？
2. 为何 performance会下降的很厉害，我试过将 mutli scale中每个scale 单独测试再用 soft_nms对所有结果 进行nms，发现performance 并没有变化太多，所以应该是 mutli scale 这里的细节问题。

如下是config文件:

architecture: CascadeRCNN
max_iters: 40968
snapshot_iter: 1707
use_gpu: true
log_iter: 20
save_dir: output
# pretrain_weights: https://paddlemodels.bj.bcebos.com/object_detection/cascade_rcnn_dcn_r50_fpn_1x.tar
pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_ssld_v2_pretrained.tar
weights: output/cascade_rcnn_dcn_r50_fpn_2x_gn/model_final
metric: COCO
num_classes: 4

CascadeRCNN:
  backbone: ResNet
  fpn: FPN
  rpn_head: FPNRPNHead
  roi_extractor: FPNRoIAlign
  bbox_head: CascadeBBoxHead
  bbox_assigner: CascadeBBoxAssigner

ResNet:
  norm_type: bn
  depth: 50
  feature_maps: [2, 3, 4, 5]
  freeze_at: 2
  variant: d
  dcn_v2_stages: [3, 4, 5]

FPN:
  min_level: 2
  max_level: 6
  num_chan: 256
  spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
  norm_type: gn

FPNRPNHead:
  anchor_generator:
    anchor_sizes: [32, 64, 128, 256, 512]
    aspect_ratios: [0.5, 1.0, 2.0]
    stride: [16.0, 16.0]
    variance: [1.0, 1.0, 1.0, 1.0]
  anchor_start_size: 32
  min_level: 2
  max_level: 6
  num_chan: 256
  rpn_target_assign:
    rpn_batch_size_per_im: 256
    rpn_fg_fraction: 0.5
    rpn_positive_overlap: 0.7
    rpn_negative_overlap: 0.3
    rpn_straddle_thresh: 0.0
  train_proposal:
    min_size: 0.0
    nms_thresh: 0.7
    pre_nms_top_n: 2000
    post_nms_top_n: 2000
  test_proposal:
    min_size: 0.0
    nms_thresh: 0.7
    pre_nms_top_n: 1000
    post_nms_top_n: 1000

FPNRoIAlign:
  canconical_level: 4
  canonical_size: 224
  min_level: 2
  max_level: 5
  box_resolution: 7
  sampling_ratio: 2

CascadeBBoxAssigner:
  batch_size_per_im: 512
  bbox_reg_weights: [10, 20, 30]
  bg_thresh_lo: [0.0, 0.0, 0.0]
  bg_thresh_hi: [0.5, 0.6, 0.7]
  fg_thresh: [0.5, 0.6, 0.7]
  fg_fraction: 0.25

CascadeBBoxHead:
  head: CascadeXConvNormHead
  nms: MultiClassSoftNMS

CascadeXConvNormHead:
  norm_type: gn

MultiClassSoftNMS:
  score_threshold: 0.01
  keep_top_k: 300
  softnms_sigma: 0.5

MultiScaleTEST:
  score_thresh: 0.05
  nms_thresh: 0.5
  detections_per_im: 100
  enable_voting: true
  vote_thresh: 0.9

LearningRate:
  base_lr: 0.02
  schedulers:
  - !PiecewiseDecay
    gamma: 0.1
    milestones: [27312, 37554]
  - !LinearWarmup
    start_factor: 0.001
    steps: 500

OptimizerBuilder:
  optimizer:
    momentum: 0.9
    type: Momentum
  regularizer:
    factor: 0.0001
    type: L2

_READER_: 'faster_fpn_reader.yml'
TrainReader:
  batch_size: 2

EvalReader:
  batch_size: 1
  inputs_def:
    fields: ['image', 'im_info', 'im_id', 'im_shape']
    multi_scale: true
    num_scales: 2
    use_flip: true
  sample_transforms:
  - !DecodeImage
    to_rgb: true
  - !NormalizeImage
    is_channel_first: false
    is_scale: true
    mean:
    - 0.485
    - 0.456
    - 0.406
    std:
    - 0.229
    - 0.224
    - 0.225
  - !MultiscaleTestResize
    origin_target_size: 1280
    origin_max_size: 2000
    use_flip: true
  - !Permute
    channel_first: true
    to_bgr: false
  - !PadMultiScaleTest
    pad_to_stride: 32
  worker_num: 2