What 's the difference between tf.placeholder and tf. Variable ?

In the TensorFlow framework, tf.placeholder and tf.Variable are two distinct types of constructs that serve different purposes in building neural networks.

tf.Variable

tf.Variable is primarily used to store and update parameters that the network learns during training. For example, weights and biases in the network are typically defined as tf.Variable because these parameters must be continuously updated to optimize network performance.

Example:

python
weights = tf.Variable(tf.random_normal([784, 200], stddev=0.35), name="weights")
biases = tf.Variable(tf.zeros([200]), name="biases")

In the above example, weights and biases represent learnable parameters defined as tf.Variable to enable updates during training.

tf.placeholder

tf.placeholder is used to define the input data structure for computations, which must be explicitly filled when TensorFlow executes a calculation. Typically, during neural network training, tf.placeholder is employed to pass input data and labels.

Example:

python
x = tf.placeholder(tf.float32, shape=[None, 784], name="x")
y = tf.placeholder(tf.float32, shape=[None, 10], name="y")

In this example, x and y denote input image data and corresponding labels, which are populated with actual data during training.

Summary

In summary, tf.Variable is used to store model parameters that are updated during learning, whereas tf.placeholder is used to define the structure of input data, which must be filled when the model is executed. Both are essential components in TensorFlow-based neural network construction, but they serve fundamentally different roles.