TopolearnEvery value flowing through a TensorFlow program is a tensor, so this is the vocabulary for everything else.
A `tf.Tensor` is an n-dimensional array with a shape and a dtype, much like a NumPy array. Create one with tf.constant(value) or from factories like tf.zeros, tf.ones, tf.range and tf.random.normal.
Since TensorFlow 2, code runs eagerly: each operation executes immediately and returns a concrete result, so you can print tensors and debug with normal Python tools. (Graphs still exist, but you opt into them with tf.function.)
Things that differ from NumPy:
- Tensors are immutable. You cannot assign to
t[0, 0]; you create a new tensor instead. Mutable state lives intf.Variable. - Python ints become
int32and Python floatsfloat32by default; NumPy arrays keep their dtype (oftenint64orfloat64). t.numpy()returns a NumPy array, and most NumPy functions accept tensors directly.
Inspect a tensor with t.shape (a TensorShape, which may contain None for unknown dimensions in graphs), t.dtype, tf.rank(t) and tf.size(t).
Operations are placed on a GPU automatically when one is available, so the same code runs on CPU and GPU without changes.
Example
import tensorflow as tf
t = tf.constant([[1, 2, 3], [4, 5, 6]])
print(t.shape, t.dtype) # (2, 3) <dtype: 'int32'>
print(t.numpy()) # [[1 2 3]
# [4 5 6]]
f = tf.constant([1.0, 2.0])
print(f.dtype) # <dtype: 'float32'>
print(t[1, 2].numpy()) # 6
print(tf.rank(t).numpy(), tf.size(t).numpy()) # 2 6
print(tf.zeros([2, 2]).shape) # (2, 2)
# t[0, 0] = 9 -> TypeError: tensors are immutable; use tf.Variable for stateRemember
- Tensors are typed, immutable n-d arrays
- TF 2 runs eagerly;
.numpy()gets the values - Python ints default to int32 and floats to float32
Watch out for
- Trying to assign into a tensor; build a new one or use a
tf.Variable.