3
Ðñß^Ä?  ã               @   sZ   d dl Zdd„ Zdd„ Zdddgd dfd	d
„Zdddgfdd„Zddddgd fdd„ZdS )é    Nc             C   sR  |}t jƒ }�x>|D �]4}|dkr8|jt j| ddƒƒ q|d dkrb|jt j| |d |d ƒƒ q|dkr–|jt j| ||dddƒƒ |jt j|ƒƒ q|d dkrÚ|jt j| ||d dddƒƒ |d }|jt j|ƒƒ q|d dkoðt|ƒdk�r*|jt j| ||d ddƒƒ |d }|jt j|ƒƒ q|d dk�rt|ƒdk�r|jt jƒ jt j| ||d ddƒƒjt jƒ jt j| ||d dddƒƒjt j|d ƒƒjt j| |d |d ddƒƒjt j|d ƒƒjt j	| |d |d dddƒƒƒƒjt j
ƒ ƒ |d |d  }|jt j|ƒƒ q|d dk�r<t|ƒd	k�r<|jt jƒ jt j| ||d ddƒƒjt jƒ jt j| ||d dddƒƒjt j|d ƒƒjt j| |d |d ddƒƒjt j|d ƒƒjt j	| |d |d dddƒƒƒjt jƒ jt j| ||d dddƒƒjt j|d ƒƒjt j| |d |d ddƒƒjt j|d ƒƒjt j| |d |d dddƒƒjt j|d ƒƒjt j| |d |d ddƒƒjt j|d ƒƒjt j	| |d |d dddƒƒjt j|d ƒƒjt j| |d |d ddƒƒjt j|d ƒƒjt j	| |d |d dddƒƒƒƒjt j
ƒ ƒ |d |d  |d  }|jt j|ƒƒ q|d dkrt|ƒd
kr|jt jƒ jt j| ||d ddƒƒjt jƒ jt j| ||d dddƒƒjt j|d ƒƒjt j| |d |d ddƒƒjt j|d ƒƒjt j	| |d |d dddƒƒƒjt jƒ jt j| ||d dddƒƒjt j|d ƒƒjt j| |d |d ddƒƒjt j|d ƒƒjt j| |d |d dddƒƒjt j|d ƒƒjt j| |d |d ddƒƒjt j|d ƒƒjt j	| |d |d dddƒƒjt j|d ƒƒjt j| |d |d ddƒƒjt j|d ƒƒjt j	| |d |d dddƒƒƒjt jƒ jt j| ||d	 dddƒƒjt j|d	 ƒƒjt j| |d	 |d	 ddƒƒjt j|d	 ƒƒjt j| |d	 |d	 dddƒƒjt j|d	 ƒƒjt j| |d	 |d	 ddƒƒjt j|d	 ƒƒjt j| |d	 |d	 dddƒƒjt j|d	 ƒƒjt j| |d	 |d	 ddƒƒjt j|d	 ƒƒjt j	| |d	 |d	 dddƒƒjt j|d	 ƒƒjt j| |d	 |d	 ddƒƒjt j|d	 ƒƒjt j	| |d	 |d	 dddƒƒjt j|d	 ƒƒjt j| |d	 |d	 ddƒƒjt j|d	 ƒƒjt j	| |d	 |d	 dddƒƒƒƒjt j
ƒ ƒ |d |d  |d  |d	  }|jt j|ƒƒ qW |S )zc
    VGG style nets
    Use submanifold convolutions
    Also implements 'Plus'-augmented nets
    ÚMPé   é   r   é   zC3/2FÚCé   é   )ÚscnÚ
SequentialÚaddÚ
MaxPoolingÚConvolutionÚBatchNormReLUÚlenÚSubmanifoldConvolutionÚConcatTableÚDeconvolutionÚ	JoinTable)Ú	dimensionÚnInputPlanesÚlayersÚnPlanesÚmÚx© r   úA/home/yuhw/wc/SparseConvNet/sparseconvnet/networkArchitectures.pyÚSparseVggNet	   sÒ    "
 "
  "
 r   c                sn  |}t jƒ }‡ fdd„}�x>|D �]4\}}}}	�x$t|ƒD �]}
|d dk�r@|
dkrÞ|jt j|ƒƒ |jt jƒ jt jƒ j|	dkr”t jˆ ||ddƒnt jˆ ||d|	dƒƒjt j|ƒƒjt jˆ ||ddƒƒƒj||||	ƒƒƒ nb|jt jƒ jt jƒ jt j|ƒƒjt jˆ ||ddƒƒjt j|ƒƒjt jˆ ||ddƒƒƒjt jƒ ƒƒ |}|jt j	ƒ ƒ q:W q W |jt j|ƒƒ |S )zR
    pre-activated ResNet
    e.g. layers = {{'basic',16,2,1},{'basic',32,2}}
    c                s>   |dkrt jˆ | |d|dƒS | |kr2t j| |dƒS t jƒ S d S )Nr   r   F)r	   r   ÚNetworkInNetworkÚIdentity)ÚnInÚnOutÚstride)r   r   r   ÚresidualŒ   s
    zSparseResNet.<locals>.residualr   Úbr   r   F)
r	   r
   Úranger   r   r   r   r   r   ÚAddTable)r   r   r   r   r   r"   Z	blockTypeÚnÚrepsr!   Úrepr   )r   r   ÚSparseResNet„   sf    

r)   Fr   r   c                s6   ‡‡‡fdd„‰d‡ ‡‡‡‡‡fdd„	‰ ˆ ||ƒ}|S )a~  
    U-Net style network with VGG or ResNet-style blocks.
    For voxel level prediction:
    import sparseconvnet as scn
    import torch.nn
    class Model(nn.Module):
        def __init__(self):
            nn.Module.__init__(self)
            self.sparseModel = scn.Sequential().add(
               scn.SubmanifoldConvolution(3, nInputFeatures, 64, 3, False)).add(
               scn.UNet(3, 2, [64, 128, 192, 256], residual_blocks=True, downsample=[2, 2]))
            self.linear = nn.Linear(64, nClasses)
        def forward(self,x):
            x=self.sparseModel(x).features
            x=self.linear(x)
            return x
    c          	      sÆ   ˆr�| j tjƒ j ||kr tjƒ ntj||dƒƒj tjƒ j tj|ˆd�ƒj tjˆ ||ddƒƒj tj|ˆd�ƒj tjˆ ||ddƒƒƒƒj tjƒ ƒ n2| j tjƒ j tj|ˆd�ƒj tjˆ ||ddƒƒƒ d S )NF)Ú	leakinessr   )	r   r	   r   r   r   r
   ÚBatchNormLeakyReLUr   r%   )r   Úar#   )r   r*   Úresidual_blocksr   r   ÚblockÝ   s     zUNet.<locals>.blockr   c          
      s@  t jƒ }x4tˆƒD ](}ˆ||dkr&|n| d | d ƒ d}qW t| ƒdk�r<|jt jƒ jt jƒ ƒjt jƒ jt j| d ˆd�ƒjt jˆ| d | d ˆd ˆd dƒƒjˆ | dd … ƒƒjt j| d ˆd�ƒjt j	ˆ| d | d ˆd ˆd dƒƒƒƒ |jt j
ƒ ƒ x8tˆƒD ],}ˆ|| d |dk�r(dnd | d ƒ �qW |S )Nr   r   )r*   Fr   éÿÿÿÿr/   )r	   r
   r$   r   r   r   r   r+   r   r   r   )r   Ún_input_planesr   Úi)ÚUr.   r   Ú
downsampler*   r'   r   r   r2   ë   s(     
,zUNet.<locals>.Ur/   )r/   r   )r   r'   r   r-   r3   r*   r0   r   r   )r2   r.   r   r3   r*   r'   r-   r   ÚUNetË   s    
r4   c                s.   ‡‡fdd„‰‡ ‡‡‡‡fdd„‰ ˆ |ƒ}|S )a©  
    Fully-convolutional style network with VGG or ResNet-style blocks.
    For voxel level prediction:
    import sparseconvnet as scn
    import torch.nn
    class Model(nn.Module):
        def __init__(self):
            nn.Module.__init__(self)
            self.sparseModel = scn.Sequential().add(
               scn.SubmanifoldConvolution(3, nInputFeatures, 64, 3, False)).add(
               scn.FullyConvolutionalNet(3, 2, [64, 128, 192, 256], residual_blocks=True, downsample=[2, 2]))
            self.linear = nn.Linear(64+128+192+256, nClasses)
        def forward(self,x):
            x=self.sparseModel(x).features
            x=self.linear(x)
            return x
    c          	      sº   ˆrˆ| j tjƒ j ||kr tjƒ ntj||dƒƒj tjƒ j tj|ƒƒj tjˆ ||ddƒƒj tj|ƒƒj tjˆ ||ddƒƒƒƒj tjƒ ƒ n.| j tjƒ j tj|ƒƒj tjˆ ||ddƒƒƒ d S )NFr   )	r   r	   r   r   r   r
   r   r   r%   )r   r,   r#   )r   r-   r   r   r.     s     z$FullyConvolutionalNet.<locals>.blockc          
      sþ   t jƒ }t| ƒdkr<xätˆƒD ]}ˆ|| d | d ƒ qW n¾t jƒ }x$tˆƒD ]}ˆ|| d | d ƒ qNW |jt jƒ jt jƒ ƒjt jƒ jt j| d ƒƒjt jˆ| d | d ˆd ˆd dƒƒjˆ | dd … ƒƒjt j	ˆˆd ˆd ƒƒƒƒ |jt j
ƒ ƒ |S )Nr   r   F)r	   r
   r   r$   r   r   r   r   r   Ú	UnPoolingr   )r   r   Ú_)r2   r.   r   r3   r'   r   r   r2   #  s$    
z FullyConvolutionalNet.<locals>.Ur   )r   r'   r   r-   r3   r   r   )r2   r.   r   r3   r'   r-   r   ÚFullyConvolutionalNet  s    r7   c                sZ   ˆdkrˆd ‰‡fdd„}‡‡‡‡	fdd„‰‡‡fdd„‰ ‡ ‡‡‡‡‡fd	d
„‰ˆˆƒS )Nr   r   c                s   | ˆ  S )Nr   )r   )r   r   r   Úl=  s    z/FullConvolutionalNetIntegratedLinear.<locals>.lc          
      s¸   x²t ˆƒD ]¦}ˆrˆ| jtjƒ jtjƒ ƒjtjƒ jtj|ˆd�ƒjtjˆ ||ddƒƒjtj|ˆd�ƒjtjˆ ||ddƒƒƒƒjtjƒ ƒ q
| jtj|ˆd�ƒjtjˆ ||ddƒƒ q
W d S )N)r*   r   F)	r$   r   r	   r   r   r
   r+   r   r%   )r   Únpr6   )r   r*   r'   r"   r   r   Úfoo?  s    
z1FullConvolutionalNetIntegratedLinear.<locals>.fooc                s,   | j tj|ˆ d�ƒ | j tj|ˆ|ƒƒ d S )N)r*   )r   r	   r+   r   )r   r   Úbias)r*   ÚnClassesr   r   ÚbarM  s    z1FullConvolutionalNetIntegratedLinear.<locals>.barc          	      sÂ   t jƒ }ˆ|| d ƒ t| ƒdkr4ˆ || d dƒ nŠt jƒ }ˆ || dƒ t jt j| ˆd�t jˆ| d | d ˆd ˆd dƒˆ| dd … ƒt jˆˆd ˆd ƒƒ}|jt||ƒƒ |jt jƒ ƒ d S )Nr   r   TF)r*   )	r	   r
   r   r+   r   r5   r   r   r%   )r   r   r,   r#   )r=   Úbazr   r3   r:   r*   r   r   r>   P  s    "z1FullConvolutionalNetIntegratedLinear.<locals>.bazr/   r   )r   r'   r   r<   r"   r3   r*   r8   r   )
r=   r>   r   r3   r:   r*   r<   r   r'   r"   r   Ú$FullConvolutionalNetIntegratedLinear:  s    r?   r/   r/   )Úsparseconvnetr	   r   r)   r4   r7   r?   r   r   r   r   Ú<module>   s
   {G87