如何冻结张量流模型？

Question

I have a model.py file that save a class of rnn. for example:

class TextRNN:
    def __init__(self, hidden_size, num_classes, learning_rate...):

        self.input_data = tf.placeholder(tf.int32,[batch_size,None],name="input_data")
        self.output_label = tf.placeholder(tf.float32,[batch_size,num_classes],name="output_label")
        self.dropout_rate = tf.placeholder(tf.float32,name="dropout_rate")

        with tf.name_scope("embedding"):
            ...

        with tf.name_scope("hidden"):
            ...

        with tf.name_scope("output"):
            ...

In my main.py, I used the following code to use this model.

with tf.Graph().as_default():
        sess = tf.Session()
        with sess.as_default():
            rnn = TextRNN(...)
            ...

            #training step
            def train_step(x_batch, y_batch...):
                feed_dict = {rnn.input_data:x_batch,rnn.output_label:y_batch,rnn.dropout_rate:0.5}
                sess.run(feed_dict)
                ...
            #testing step
            def test_step(x_batch, y_batch...):
                feed_dict = {rnn.input_data:x_batch,rnn.output_label:y_batch,rnn.dropout_rate:0.5}
                sess.run(feed_dict)

My question is how to freeze the model in the testing step? I know if I run training step, the weights of the model would be updated. But when I run testing step, I don't want to update the weights anymore, I only want to get the predicted results? How should I modify my code to make this happen?

Answer 1

To train your model, you need to define a training op an run it, eg train_op = tf.train.AdamOptimizer(1e-4).minimize(loss_op) would mean that each time you run train_op it will proceed one step of the AdamOptimizer on the loss_op and it will update the variables.

But if you want to evaluate your model, just run you output_op and you'll get your result. TensorFlow will only run the part of the graph that is required and not more. So when you ask for output_op or loss_op it won't modify the variables.

You might find this page useful to understand more how things work in TensorFlow.