011. tf.constant
EXECUTIVE_SUMMARY // AEO_OPTIMIZED
[Answer Engine Overview: What, Why & How]
The most basic unit in TensorFlow is tf.constant(). It creates a tensor that cannot be changed. This immutability allows TensorFlow to optimize the memory allocation perfectly on the GPU. You use constants for input data, like the pixels of an image or the text of a book, because the input data shouldn't change during training.
022. The Importance of dtype
When you pass a Python list like [1, 2, 3] into tf.constant, TensorFlow automatically infers the dtype as tf.int32. This is dangerous! Neural networks require decimal numbers for gradient descent. Always be explicit: tf.constant([1, 2, 3], dtype=tf.float32). If you mix an int32 tensor with a float32 tensor in a math operation, TensorFlow will crash.
033. NumPy Interoperability
TensorFlow was designed to play incredibly nicely with NumPy. If you pass a NumPy array into tf.constant(), it works perfectly. If you have a Tensor and want to plot it using Matplotlib (which requires NumPy arrays), you simply call .numpy() on the tensor. This bridge makes data preprocessing very easy.
?Frequently Asked Questions
Why do we use `tf.float32` instead of `tf.float64`?
Float64 has twice the precision, but uses twice the GPU VRAM and takes twice as long to process. In Deep Learning, we don't need 15 decimal places of accuracy; `float32` (7 decimal places) is the perfect industry standard.
Are Tensors just arrays?
Conceptually, yes. But physically, Tensors are backed by highly optimized C++ code that allows them to run natively on graphics cards (GPUs) and Google's custom hardware (TPUs), whereas NumPy arrays only run on the CPU.
