Full control on training loop
95% use case tf.keras and tf.data(Higher level API)
What can we do?
Build custom everything from loss function initializers, metrics, layers, models, regularizations
What is TF?
Numeric computation library build by google brain team
Features?
- GPU support(very fast)
- lower level very similar to NumPy
- Computation graph can be exported(to java, Linux cool bro!)
TF kernels
many operations have many kernels(for specific devices GPU, TPU)
Why it is called TensorFlow?
This API revolves around tensors(Kinda of vector). multiple operations of tensor it is called tensor flow.
TF and NumPy
Numpy 64-bit; TF 32 bit
easy convertion .numpy()