imgtools.utils

Support tools for programming, e.g., type and dtype conversion.

Links

Helpers

Tensorlike	Represent a union type
_to_channel_coeff	Convert shape of coefficeints such that coeff (op) img is valid, where op may be one of arithmetic operations or other operator.
align_device_type	Mathes the device and dtype of the source tensor to the target tensor.
arrayize	Converts an item to a np.ndarray.
tensorize	Converts an item to a contiguous torch.Tensor.

---

Documents

imgtools.utils.align_device_type(source: Tensor, target: Tensor)

Mathes the device and dtype of the source tensor to the target tensor.

Parameters:

sourcetorch.Tensor

A tensor need to align.

targettorch.Tensor

A tensor provides device and dtype.

Returns:

torch.Tensor

Source tensor with

• same device as target tensor

• same dtype as target tensor if target.dtype is a floating point

• float32 dtype if target.dtype is not a floating point

imgtools.utils.arrayize(img: Tensor | ndarray | list[int | float | bool] | int | float | bool) → ndarray

Converts an item to a np.ndarray.

If input is a torch.Tensor:

1. Moves -3-axis to the last for ndim >= 3.

2. Without any handling for the other cases.

For other types, convert to a ndarray by np.array.

This function is not jit-able.

Parameters:

imgTensorlike

An item to be converted to tensor.

imgtools.utils.tensorize(img: Tensor | ndarray | list[int | float | bool] | int | float | bool) → Tensor

Converts an item to a contiguous torch.Tensor.

If input is a np.ndarray:

1. Moves -1-axis to -3-axis if ndim >= 3.

2. Reshape to (1, H, W) if ndim = 2.

3. Otherwise, without any handling.

For other types, convert to a tensor by torch.tensor.

This function is not jit-able.

Parameters:

imgTensorlike

An item to be converted to tensor.

Examples

>>> import numpy as np
>>> from imgtools.utils import tensorize
>>> from PIL import Image
>>>
>>> img = np.asarray(Image.open(file))
>>> img.shape  # (H, W, C)
>>> img_t = tensorize(img)
>>> img_t.shape  # (C, H, W)