3D Convolution

a=torch.Tensor
([[[[[1, 0, 0],
     [0, 0, 0],
     [0, 0, 0]],
    
    [[0, 0, 0],
     [0, 0, 0],
     [0, 0, 0]],
    
    [[0, 0, 0],
     [0, 0, 0],
     [0, 0, 0]]],
     
     
                  
   [[[1, 0, 0],
     [0, 0, 0],
     [0, 0, 0]],
     
    [[0, 0, 0],
     [0, 0, 0],
     [0, 0, 0]],
     
    [[0, 0, 0],
     [0, 0, 0],
     [0, 0, 0]]],
                  
                  
                  
                  
   [[[0, 0, 0], 
     [0, 0, 0], 
     [0, 0, 0]], 
     
    [[0, 0, 0], 
     [0, 0, 0], 
     [0, 0, 0]], 
     
    [[0, 0, 0], 
     [0, 0, 0], 
     [0, 0, 0]]]]])
     
     
     
     
print(a.size())

c=nn.Conv3d(3,1,kernel_size=3,stride=1,padding=0,bias=False)
#print(c.kernel_size)
print('parameter:',list(c.parameters())[0].size())
print('parameter:',list(c.parameters()))

print('output:',c(a).size())
print(c(a))             
                 
torch.Size([1, 3, 3, 3, 3])
parameter: torch.Size([1, 3, 3, 3, 3])
parameter: [Parameter containing:
tensor([[[[[ 0.0963, -0.0892, -0.0216],
           [-0.0956, -0.0681, -0.0771],
           [-0.0806,  0.0087, -0.1094]],

          [[ 0.1024,  0.0459, -0.0247],
           [ 0.0777, -0.0914,  0.0320],
           [ 0.0076, -0.1092, -0.1107]],

          [[-0.0586, -0.1091,  0.0917],
           [ 0.0980,  0.0100,  0.0645],
           [-0.1089,  0.0556, -0.0833]]],



         [[[-0.0283,  0.0584, -0.0582],
           [-0.0206, -0.0328,  0.0369],
           [ 0.0903,  0.1025,  0.0495]],

          [[-0.1022,  0.0692, -0.0019],
           [ 0.0744, -0.0196, -0.0499],
           [ 0.0618, -0.0375, -0.0653]],

          [[ 0.0642,  0.0721,  0.0559],
           [ 0.0326, -0.0983,  0.0013],
           [ 0.0132, -0.0825,  0.0183]]],





         [[[-0.0428, -0.0781, -0.1067],
           [ 0.0123, -0.0801,  0.1090],
           [-0.0895,  0.0606, -0.0247]],

          [[-0.0524, -0.0748,  0.0461],
           [-0.1102, -0.0550, -0.0648],
           [ 0.0374, -0.1109,  0.0591]],

          [[ 0.0328, -0.0591,  0.0898],
           [-0.0086, -0.1068,  0.0611],
           [ 0.0645, -0.0241, -0.0358]]]]], requires_grad=True)]
           
           
output: torch.Size([1, 1, 1, 1, 1])
tensor([[[[[0.0680]]]]], grad_fn=<ThnnConv3DBackward>)

 

0.0963-0.0283=0.0680

 

a=torch.Tensor(
[[[[[0, 0, 0],
    [0, 0, 0],
    [0, 0, 0],
    [0, 0, 0]],
    
   [[0, 0, 0],
    [0, 0, 0],
    [0, 0, 0],
    [0, 0, 0]],
    
   [[0, 0, 0],
    [0, 0, 0],
    [0, 0, 0],
    [0, 0, 0]],
    
   [[1, 0, 0],
    [0, 0, 0],
    [0, 0, 0],
    [0, 0, 0]]],
    
    
    
[[[0, 0, 0],
  [0, 0, 0],
  [0, 0, 0],
  [0, 0, 0]],
  
 [[0, 0, 0],
  [0, 0, 0],
  [0, 0, 0],
  [0, 0, 0]],
  
 [[0, 0, 0],
  [0, 0, 0],
  [0, 0, 0],
  [0, 0, 0]],
  
 [[0, 0, 0],
  [0, 0, 0],
  [0, 0, 0],
  [0, 0, 0]]],
  
  
  
  
[[[0, 0, 0], 
  [0, 0, 0], 
  [0, 0, 0], 
  [0, 0, 0]], 
  
 [[0, 0, 0], 
  [0, 0, 0], 
  [0, 0, 0], 
  [0, 0, 0]],
  
 [[0, 0, 0], 
  [0, 0, 0], 
  [0, 0, 0], 
  [0, 0, 0]],
   
 [[0, 0, 0],
  [0, 0, 0],
  [0, 0, 0],
  [0, 0, 0]]]]])
print(a.size())
#print(a)


c=nn.Conv3d(3,1,kernel_size=3,stride=1,padding=0,bias=False)
#print(c.kernel_size)
print('parameter:',list(c.parameters())[0].size())
print('parameter:',list(c.parameters()))

print('output:',c(a).size())
print(c(a))


torch.Size([1, 3, 4, 4, 3])
parameter: torch.Size([1, 3, 3, 3, 3])
parameter: [Parameter containing:
tensor([[[[[-0.0322, -0.0171,  0.0772],
           [ 0.0457, -0.0895,  0.0420],
           [-0.0041, -0.0943,  0.0544]],

          [[-0.0073, -0.0972,  0.0987],
           [ 0.0156,  0.0337,  0.0930],
           [ 0.0870, -0.0117,  0.0135]],

          [[ 0.0989,  0.0122,  0.0637],
           [-0.0081, -0.0978, -0.0976],
           [-0.0183, -0.0734,  0.0018]]],


         [[[ 0.1081,  0.0143,  0.0393],
           [ 0.1051, -0.1046,  0.0843],
           [ 0.0247,  0.0990, -0.0842]],

          [[-0.0996,  0.0797, -0.0786],
           [ 0.0521,  0.0205, -0.0817],
           [ 0.0725,  0.0032, -0.0549]],

          [[-0.0191,  0.0530,  0.0610],
           [ 0.0903,  0.0611,  0.1045],
           [ 0.0710,  0.1000, -0.1012]]],


         [[[ 0.0978,  0.0958,  0.0387],
           [-0.0788,  0.0464, -0.0106],
           [-0.0678, -0.0065,  0.0845]],

          [[-0.0376, -0.0325, -0.0785],
           [ 0.1079,  0.0447,  0.0608],
           [ 0.0972, -0.0460,  0.0304]],

          [[-0.0019,  0.0226, -0.1037],
           [ 0.0025,  0.0582,  0.0162],
           [-0.1090, -0.0908, -0.0908]]]]], requires_grad=True)]
output: torch.Size([1, 1, 2, 2, 1])
tensor([[[[[0.0000],
           [0.0000]],

          [[0.0989],
           [0.0000]]]]], grad_fn=<ThnnConv3DBackward>)

1.depth도 height/weight 계산하는 것 처럼 계산

2.depth는 맞춰주지 않음 4 였을때 첫번째 filter 재사용 (Conv3D 의 정의:Conv를 3D 방향으로 한다.)