使用python实现3d卷积

代码：

import numpy as np

def conv_per_kernel(feat,kernel,s,p):
    c, h, w = feat.shape
    c, k, _ = kernel.shape
    newh = int((h + 2 * p - k) / s) + 1
    neww = int((w + 2 * p - k) / s) + 1

    res = np.zeros([newh, neww])
    temp=np.zeros([h+2*p,w+2*p])
    temp[p:-p,p:-p]=feat
    
    for i in range(newh):
        for j in range(neww):
            center_i=i+k/2+s*i
            center_j=j+k/2+s*j
            start_i=center_i-k/2
            end_i=center_i+k/2
            start_j=center_j-k/2
            end_j=center_j+k/2
            sum=0
            for m in range(start_i,end_i+1):
                for n in range(start_j,end_j+1):
                   for u in range(k):
                     sum+=temp[m,n,u]*kernel[m-start_i,n-start_j,u]        
            res[i,j]=sum
    return res

def conv_per_batch(feat,kernel,s,p):
    c, h, w = feat.shape
    o, c, k, _ = kernel.shape
    newh=int((h+2*p-k)/s)+1
    neww=int((w+2*p-k)/s)+1

    res=np.zeros([o,newh,neww])

    for i in range(o):
        res[i,:,:]=conv_per_kernel(feat,kernel[i,:,:,:],s,p)
    
    return res
    
def conv3d(featuremap,kernel,stride,padding):
    b,c,h,w=featuremap.shape
    o,c,ks,_=kernel.shape
    s=stride
    p=padding

    res=[]

    for i in range(b):
        out=conv_per_batch(featuremap[i,:,:,:],kernel,s,p)
        res.append(out)

    res=np.asarray(res)
    return res

if __name__=='__main__':
    pass

思路：

把整个过程梳理了一下，如有不对的地方，请指正；