def backProp_bpLayer(dfm2,fm1,fm2,sfm,W1,b1,W2,b2):
"return [dw,db,din] 权值的梯度，偏置值的梯度，输入数据的梯度"
dfm2=dfm2*fun.dsigmoid(fm2)
db2=np.zeros(b2.shape)
dw2=np.zeros(W2.shape)
dfm1=np.zeros(fm1.shape)
for i in xrange(0,db2.shape[0]):
this_fm2=fm2[i]
this_dfm2=dfm2[i]
db2[i]=np.sum(this_dfm2)
this_w2=W2[i]
this_dfm1,dw2[i]=CNN.dMultiply(this_dfm2,this_fm2,fm1,this_w2)
dfm1=dfm1+this_dfm1
dfm1=dfm1*fun.dsigmoid(fm1)
#dfm1=dfm1.reshape(fm1.shape)
db1=np.zeros(b1.shape)
dsfm=np.zeros(sfm.shape)
dw1=np.zeros(W1.shape)
for i in xrange(b1.shape[0]):
this_dfm1=dfm1[i]
this_fm1=fm1[i]
db1[i]=np.sum(this_dfm1)
this_dsfm,dw1[i]=fun.dconv3d(this_dfm1,this_fm1,sfm,W1[i])
dsfm=dsfm+this_dsfm
return [dsfm,dw1,db1,dw2,db2]
def backProp_conv(dcfm,cfm,fm,ct,w,b):
'将卷积层反向传播'
'dcfm 误差关于cfm的偏导数'
'cfm 卷基层的输出'
'fm 卷基层的输入'
'ct fm和cfm之间的连接表'
'w 卷积核'
'b 偏置'
dfm=np.zeros(fm.shape)
dw=np.zeros(w.shape)
db=np.zeros(b.shape)
dcfm=dcfm*fun.dsigmoid(cfm)
for i in xrange(0,b.shape[0]):
this_dcfm=dcfm[i]
db[i]=np.sum(this_dcfm)
for i in xrange(0,w.shape[0]):
this_fm=fm[ct[0,i]]
this_w=w[i]
this_dcfm=dcfm[ct[1,i]]
'误差关于dfm的偏导数'
this_dfm=fun.dconv2_in(this_dcfm, this_w,this_w)
dfm[ct[0,i]]=dfm[ct[0,i]]+this_dfm
'误差关于dw的偏导数'
dw[i]=fun.dconv2_kernel(this_dcfm, this_fm,this_w)
return [dfm,dw,db]
def backPropSubsampling(dsfm,sfm,fm,sw,sb,pool_size,stride):
dims,height,width=fm.shape
dfm=np.zeros((dims,height,width))
dsw=np.zeros(sw.shape)
dsb=np.zeros(sb.shape)
dsfm=dsfm*fun.dsigmoid(sfm)
kernel=np.ones(pool_size)
for i in xrange(0,dims):
this_dsfm=dsfm[i]
this_kernel=kernel*sw[i]
dsb[i]=np.sum(this_dsfm[:])
dsfm_beforeSubsampling=np.zeros((height-pool_size[0]+1,width-pool_size[1]+1))
dx=0
dy=0
for y in xrange(0,dsfm_beforeSubsampling.shape[0],stride):
dx=0
for x in xrange(0,dsfm_beforeSubsampling.shape[1],stride):
dsfm_beforeSubsampling[y,x]=this_dsfm[dy,dx]
dx+=1
dy+=1
dfm[i]=fun.dconv2_in(dsfm_beforeSubsampling, fm[i],this_kernel)
dthis_kernel=fun.dconv2_kernel(dsfm_beforeSubsampling,fm[i],this_kernel)
dsw[i]=np.sum(dthis_kernel)
return [dfm,dsw,dsb]
