def backward(self, y_pred, y_train, cache):
X, h1_cache, h2_cache, score_cache, nl_cache1, nl_cache2, u1, u2, bn1_cache, bn2_cache = cache
# Output layer
grad_y = self.dloss_funs[self.loss](y_pred, y_train)
# Third layer
dh2, dW3, db3 = l.fc_backward(grad_y, score_cache)
dW3 += reg.dl2_reg(self.model['W3'], self.lam)
dh2 = self.backward_nonlin(dh2, nl_cache2)
dh2 = l.dropout_backward(dh2, u2)
dh2, dgamma2, dbeta2 = l.bn_backward(dh2, bn2_cache)
# Second layer
dh1, dW2, db2 = l.fc_backward(dh2, h2_cache)
dW2 += reg.dl2_reg(self.model['W2'], self.lam)
dh1 = self.backward_nonlin(dh1, nl_cache1)
dh1 = l.dropout_backward(dh1, u1)
dh1, dgamma1, dbeta1 = l.bn_backward(dh1, bn1_cache)
# First layer
_, dW1, db1 = l.fc_backward(dh1, h1_cache)
dW1 += reg.dl2_reg(self.model['W1'], self.lam)
grad = dict(
W1=dW1, W2=dW2, W3=dW3, b1=db1, b2=db2, b3=db3, gamma1=dgamma1,
gamma2=dgamma2, beta1=dbeta1, beta2=dbeta2
)
return grad
def backward(self, y_pred, y_train, cache):
X, h1_cache, h2_cache, score_cache, nl_cache1, nl_cache2, u1, u2, bn1_cache, bn2_cache = cache
# Output layer
grad_y = self.dloss_funs[self.loss](y_pred, y_train)
# Third layer
dh2, dW3, db3 = l.fc_backward(grad_y, score_cache)
dW3 += reg.dl2_reg(self.model['W3'], self.lam)
dh2 = self.backward_nonlin(dh2, nl_cache2)
dh2 = l.dropout_backward(dh2, u2)
dh2, dgamma2, dbeta2 = l.bn_backward(dh2, bn2_cache)
# Second layer
dh1, dW2, db2 = l.fc_backward(dh2, h2_cache)
dW2 += reg.dl2_reg(self.model['W2'], self.lam)
dh1 = self.backward_nonlin(dh1, nl_cache1)
dh1 = l.dropout_backward(dh1, u1)
dh1, dgamma1, dbeta1 = l.bn_backward(dh1, bn1_cache)
# First layer
_, dW1, db1 = l.fc_backward(dh1, h1_cache)
dW1 += reg.dl2_reg(self.model['W1'], self.lam)
grad = dict(
W1=dW1, W2=dW2, W3=dW3, b1=db1, b2=db2, b3=db3, gamma1=dgamma1,
gamma2=dgamma2, beta1=dbeta1, beta2=dbeta2
)
return grad
def backward(self, y_pred, y_train, cache):
(X, h1_cache, h2_cache, h4_cache, h5_cache, score_cache,
hpool1_cache, hpool1, hpool2_cache, hpool2,
nl_cache1, nl_cache2, nl_cache4, nl_cache5,
bn4_cache,bn5_cache
) = cache
'''Output layer'''
grad_y = self.dloss_funs[self.loss](y_pred, y_train)
dh5, dW6, db6 = l.fc_backward(grad_y, score_cache)
'''FC-2'''
dh5 = self.backward_nonlin(dh5, nl_cache5)
dh5, dgamma5, dbeta5 = l.bn_backward(dh5, bn5_cache)
dh4, dW5, db5 = l.fc_backward(dh5, h5_cache)
'''FC -1'''
dh4 = self.backward_nonlin(dh4, nl_cache4)
dh4, dgamma4, dbeta4 = l.bn_backward(dh4,bn4_cache)
dhpool3_, dW4, db4 = l.fc_backward(dh4, h4_cache)
'''reshape'''
dhpool3 = dhpool3_.ravel().reshape(hpool2.shape)
'''Pool -2'''
dpool2 = l.maxpool_backward(dhpool3, hpool2_cache)
'''Conv -2'''
dh2 = self.backward_nonlin(dpool2, nl_cache2)
dh1, dW2, db2 = l.conv_backward(dh2, h2_cache)
'''pool -1'''
dpool1 = l.maxpool_backward(dh1, hpool1_cache)
'''conv -1'''
dh1 = self.backward_nonlin(dpool1, nl_cache1)
dX, dW1, db1 = l.conv_backward(dh1, h1_cache)
grad = dict(W1=dW1, W2=dW2, W4=dW4, W5=dW5, W6=dW6,
b1=db1, b2=db2, b4=db4, b5=db5, b6=db6,
gamma4=dgamma4,gamma5=dgamma5,
beta4=dbeta4,beta5=dbeta5
)
return grad