Note
Sentis is now called Inference Engine. The documentation has moved to https://docs.unity3d.com/Packages/com.unity.ai.inference@latest. Refer to the new location for the latest updates and guidance. Make sure to update your bookmarks and references accordingly.
Create and modify tensors
Tensor methods in Sentis are similar to methods found in frameworks like NumPy, TensorFlow, and PyTorch.
Create a tensor
Create a basic tensor using the methods in the Tensor API.
Refer to Create input for a model for more information.
Get and set values of a tensor
If your tensor data backendType is BackendType.CPU and has finished being computed IsReadbackRequestDone you can directly set and get values.
var tensor = new Tensor<float>(new TensorShape(1, 2, 3));
tensor[0, 1, 2] = 5.2f; // set value at index 0 of dim0 = 1, index 1 of dim1 = 2 and index 2 of dim2 = 3
float value = tensor[0, 1, 2];
Assert.AreEqual(5.2f, value);
Reshape a tensor
You can reshape a tensor directly, for example:
var tensor = new Tensor<float>(new TensorShape(10));
tensor.Reshape(new TensorShape(2, 5));
The new shape of the tensor must fit in the allocated data on the backend. You can use the length property of a tensor shape and the maxCapacity property of the tensor data to check the number of elements.
var tensor = new Tensor<float>(new TensorShape(10));
Assert.AreEqual(10, tensor.count);
Assert.AreEqual(10, tensor.dataOnBackend.maxCapacity);
// Reshaping the tensor with a smaller shape
tensor.Reshape(new TensorShape(2, 3));
Assert.AreEqual(6, tensor.count);
Assert.AreEqual(10, tensor.dataOnBackend.maxCapacity);
// The underlying dataOnBackend still contains 10 elements
// reshape to match dataOnBackend.maxCapacity
tensor.Reshape(new TensorShape(1, 10));
When you reshape a tensor, Sentis does not modify the data or capacity of the underlying dataOnBackend.
Note
If you use BackendType.GPUPixel, tensors aren't stored in a linear format. Consequently, you will not be able to reshape a tensor if the data is on the GPU.
Download values of a tensor
You can do a blocking download to get a copy of the data of a tensor to a NativeArray or Array as follows:
var nativeArray = tensor.DownloadToNativeArray();
var array = tensor.DownloadToArray();
Note: These methods return copies of your tensor data. Editing an array returned from one of these methods doesn't edit the tensor.
This download is a blocking call and will force a wait if ReadbackRequest hasn't been called or IsReadbackRequestDone is false. Refer to Read Outputs Asynchronously for details.