tensorflow 更新部分参数或参数分开更新

思路：

  1.根据变量名称过滤要更新的权重：

  2.如果参数分开更新，还需要设置多个优化器

 

代码示例：

def Net_1(input):

    with tf.variable_scope('Net_1'):

        fmap_input = tf.layers.conv2d(input,32,32,(1,1),padding='same',name='conv1')

        _, xh, xw, xc = fmap_input.get_shape().as_list()

        gap = tf.layers.max_pooling2d(fmap_input,(xh,xw),strides=(1,1),name='gap')

        _,h,w,c = gap.get_shape().as_list()

        gap = tf.reshape(gap,(-1,c))

        cls_logit = tf.layers.dense(gap,3,name='fc')

        cls_probs_soft = tf.nn.softmax(cls_logit, axis=1)

        cls_probs = tf.clip_by_value(cls_probs_soft, 1e-7, 1.0)

　　。。。。

　　return tmp

 

def Net_2(input_fm):

    with tf.variable_scope('Net_2'):

        _, xh, xw, xc = input_fm.get_shape().as_list()

        gap = tf.layers.max_pooling2d(input_fm, (xh, xw),strides=(1,1), name='gap_cls_head')

        _,h,w,c = gap.get_shape().as_list()

        gap = tf.reshape(gap,(-1,c))

        fc_cls_logits = tf.layers.dense(gap,3)

        cls_probs_soft = tf.nn.softmax(fc_cls_logits, axis=1)

        cls_probs = tf.clip_by_value(cls_probs_soft, 1e-8, 1.0)

        return cls_probs

 

net1的输出作为net2的输入

input_placeholder = tf.placeholder(dtype=tf.float32, shape=[None, 128, 128, 3], name='input_')

gt = tf.placeholder(dtype=tf.int32, shape=[None], name='label')

global_step = tf.Variable(0, name='globel_step', trainable=False)

 

output1 = Net_1(input_placeholder)

output2 = Net_2(output1)

loss=损失（output1,gt）

loss1=损失（output2,gt）

#net1更新
optimizer_anet = tf.train.AdamOptimizer(0.01)
#法1
net1_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='Net_1')
#法2
#tvars1 = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
#net1_vars = [vr for vr in tvars1 if 'Net_1' in vr.name]
for tmp in net1_vars:
print('net1--->',tmp.name)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
train_op_net1 = optimizer_anet.minimize(loss, global_step=global_step, var_list=net1_vars)

#net2更新
optimizer = tf.train.AdamOptimizer(0.01)
#法1
other_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='Net_2')
#法2
#tvars1 = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES)
#other_vars = [vr for vr in tvars1 if 'Net_2' in vr.name]
for tmp in other_vars:
print('net_2-->',tmp.name)
update_ops1 = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops1):
train_op_net2 = optimizer.minimize(loss2,global_step=global_step,var_list=other_vars)

    

with tf.Session() as sess:

    init_op = tf.group(tf.global_variables_initializer(),

                       tf.local_variables_initializer())

    sess.run(init_op)

　#只更新net1

   _=sess.run(train_op_net1,feed_dict={input_placeholder:数据，label:数据})

　#只更新net2

 _=sess.run(train_op_net2,feed_dict={input_placeholder:数据，label:数据})