读gaitpart代码

[BasicConv2d(
  (conv): Conv2d(1, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), bias=False)
), LeakyReLU(negative_slope=0.01, inplace=True), BasicConv2d(
  (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
), LeakyReLU(negative_slope=0.01, inplace=True), MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False), FocalConv2d(
  (conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
), LeakyReLU(negative_slope=0.01, inplace=True), FocalConv2d(
  (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
), LeakyReLU(negative_slope=0.01, inplace=True), MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False), FocalConv2d(
  (conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
), LeakyReLU(negative_slope=0.01, inplace=True), FocalConv2d(
  (conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
), LeakyReLU(negative_slope=0.01, inplace=True)]

GaitPart(
  输入sils[128, 30, 64, 44]-》[128, 1,30, 64, 44]
  (Backbone): SetBlockWrapper(
    转置成[3840, 1, 64, 44]
    (forward_block): Plain(
      (feature): Sequential(
        (0): BasicConv2d(
          (conv): Conv2d(1, 32, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2), bias=False)
        )
        [3840, 32, 64, 44]
        (1): LeakyReLU(negative_slope=0.01, inplace=True)
        (2): BasicConv2d(
          (conv): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (3): LeakyReLU(negative_slope=0.01, inplace=True)
        (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
        [3840, 32, 32, 22]
        (5): FocalConv2d(
          (conv): Conv2d(32, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        [3840, 64, 32, 22]
        (6): LeakyReLU(negative_slope=0.01, inplace=True)
        (7): FocalConv2d(
          (conv): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (8): LeakyReLU(negative_slope=0.01, inplace=True)
        (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
        [3840, 64, 16, 11]
        (10): FocalConv2d(
          (conv): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        [3840, 128, 16, 11]
        (11): LeakyReLU(negative_slope=0.01, inplace=True)
        (12): FocalConv2d(
          (conv): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        )
        (13): LeakyReLU(negative_slope=0.01, inplace=True)
      )
      [3840, 128, 16, 11]再转置得到out[128,128, 30, 16, 11]
    )
  )
  
  (HPP): SetBlockWrapper()
    输入out转置成[3840, 128, 16, 11]
  	其实是HorizontalPoolingPyramid
  	不过这里bin只有16
  	所以只有一个[3840, 128, 16]
  	再转置成out[128,128, 30, 16]
  
  
  (TFA): PackSequenceWrapper(
    (pooling_func): TemporalFeatureAggregator(
      out转置为x[16, 128, 128, 30]
      再split成16个[1, 128, 128, 30]组成的tuple feature
      让每一个feature经过下面的conv1d3x1每一层
      (conv1d3x1): ModuleList(
        (0): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (1): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (2): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (3): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (4): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (5): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (6): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (7): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (8): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (9): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (10): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (11): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (12): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (13): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (14): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
        (15): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(1,), stride=(1,), bias=False)
          )
        )
      )
      每层的结果是[1, 128, 128, 30]
      拼接得到logits3x1[16, 128, 128, 30]
      sigmoid后得到scores3x1[16, 128, 128, 30]
      
      (avg_pool3x1): AvgPool1d(kernel_size=(3,), stride=(1,), padding=(1,))
      (max_pool3x1): MaxPool1d(kernel_size=3, stride=1, padding=1, dilation=1, ceil_mode=False)
      x[2048, 128, 30]送入上面两种池化后求和得到feature3x1[2048, 128, 30]
      view成[16, 128, 128, 30]
      再与scores3x1相乘得到feature3x1

      (conv1d3x3): ModuleList(
        (0): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (1): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (2): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (3): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (4): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (5): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (6): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (7): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (8): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (9): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (10): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (11): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (12): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (13): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (14): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
        (15): Sequential(
          (0): BasicConv1d(
            (conv): Conv1d(128, 32, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
          (1): LeakyReLU(negative_slope=0.01, inplace=True)
          (2): BasicConv1d(
            (conv): Conv1d(32, 128, kernel_size=(3,), stride=(1,), padding=(1,), bias=False)
          )
        )
      )
      每层同样得到[1, 128, 128, 30]
      拼接得到logits3x3也是一样的[16, 128, 128, 30]
      sigmoid后是scores3x3
      
      (avg_pool3x3): AvgPool1d(kernel_size=(5,), stride=(1,), padding=(2,))
      (max_pool3x3): MaxPool1d(kernel_size=5, stride=1, padding=2, dilation=1, ceil_mode=False)
    )
    两种池化后求和，再view后与scores3x3相乘得到feature3x3[16, 128, 128, 30]
    和feature3x1求和后最大池化
    得到ret[16, 128, 128]再转置出out[128, 128, 16]
  (Head): SeparateFCs()
  得到embs[128, 128, 16]
  )
  (loss_aggregator): LossAggregator(
    (losses): ModuleDict(
      (triplet): TripletLoss()
    )
  )
)

{'training_feat': {
	'triplet': {
		'embeddings': embs[128, 128, 16]，
		'labels': labs128维}}, 
'visual_summary': {
	'image/sils': sils给view成[3840, 1, 64, 44]}, 
'inference_feat': {
	'embeddings': embs[128, 128, 16]}}