Tensor attributes

Data type

A tensor’s data type (dtype) states how each element is stored and what range of values it can hold. The common dtypes are exposed as constants:

import tensorplay as tp

print(tp.float32)   # the 32-bit floating-point dtype
print(tp.int64)     # the 64-bit integer dtype
print(tp.bool)      # a boolean dtype

Pass a dtype to a creation function or Tensor to control it:

x = tp.tensor([1, 2, 3], dtype=tp.float32)
print(x.dtype)      # tensorplay.float32

There are integer, floating-point, and complex dtypes (tensorplay.complex64, tensorplay.complex128). Operations that mix dtypes promote according to TensorPlay’s type-promotion rules, so a float32 tensor times an int64 tensor yields a float32 tensor.

Device

A tensor lives on a device — cpu, or a GPU such as cuda. The device determines where the data is physically stored and which kernels can run on it:

x = tp.tensor([1.0, 2.0])
print(x.device)      # cpu

Use .to(device) to move a tensor (and model.to(device) to move a model) to another device. See the CUDA page for working with accelerators.

Defaults

set_default_dtype and set_default_device change the dtype and device used for tensors created without an explicit one, and get_default_dtype / get_default_device read them back.

Where to go next

• Tensors — the full tensor creation and math API.

• Type information — finfo and iinfo, the per-dtype limits.