如何理解TensorFlow中的静态形状和动态形状? [英] How to understand static shape and dynamic shape in TensorFlow?

问题描述

在

In TensorFlow FAQ, it says:

In TensorFlow, a tensor has both a static (inferred) shape and a dynamic (true) shape. The static shape can be read using the tf.Tensor.get_shape() method: this shape is inferred from the operations that were used to create the tensor, and may be partially complete. If the static shape is not fully defined, the dynamic shape of a Tensor t can be determined by evaluating tf.shape(t).

But I still cannot fully understand the relationship between static shape and dynamic shape. Are there any examples showing their differences? Thanks.

解决方案

Sometimes the shape of a tensor depends on a value that is computed at runtime. Let's take the following example, where x is defined as a tf.placeholder() vector with four elements:

x = tf.placeholder(tf.int32, shape=[4])
print x.get_shape()
# ==> '(4,)'

The value of x.get_shape() is the static shape of x, and the (4,) means that it is a vector of length 4. Now let's apply the tf.unique() op to x

y, _ = tf.unique(x)
print y.get_shape()
# ==> '(?,)'

The (?,) means that y is a vector of unknown length. Why is it unknown? tf.unique(x) returns the unique values from x, and the values of x are unknown because it is a tf.placeholder(), so it doesn't have a value until you feed it. Let's see what happens if you feed two different values:

sess = tf.Session()
print sess.run(y, feed_dict={x: [0, 1, 2, 3]}).shape
# ==> '(4,)'
print sess.run(y, feed_dict={x: [0, 0, 0, 0]}).shape
# ==> '(1,)'

Hopefully this makes it clear that a tensor can have a different static and dynamic shape. The dynamic shape is always fully defined—it has no ? dimensions—but the static shape can be less specific. This is what allows TensorFlow to support operations like tf.unique() and tf.dynamic_partition(), which can have variable-sized outputs, and are used in advanced applications.

Finally, the tf.shape() op can be used to get the dynamic shape of a tensor and use it in a TensorFlow computation:

z = tf.shape(y)
print sess.run(z, feed_dict={x: [0, 1, 2, 3]})
# ==> [4]
print sess.run(z, feed_dict={x: [0, 0, 0, 0]})
# ==> [1]

Here's a schematic image showing both:

这篇关于如何理解TensorFlow中的静态形状和动态形状?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！