Tensor

Classes for transferring point cloud (NumPy) -> PyTorch Tensor

ToTensor

Recursively convert NumPy arrays, scalars, and container structures to PyTorch tensors.

This transform is designed to work on:

Individual values:
torch.Tensor: returned as-is.
int → LongTensor([value]).
float → FloatTensor([value]).
str: returned as-is (strings are not converted).
np.ndarray with:
- boolean dtype → torch.from_numpy(arr) (bool tensor),
- integer dtype → torch.from_numpy(arr).long(),
- float dtype → torch.from_numpy(arr).float().
Containers:
Mapping (e.g. dict): converts each value recursively, preserving keys.
Sequence (e.g. list, tuple): converts each element recursively and returns a Python list of tensors/converted items.

Any unsupported type will raise a TypeError.

Typical usage is at the end of a preprocessing pipeline, to convert a nested sample dictionary (coords, features, labels) from NumPy to tensors.

`call(data)`

Convert input data (and nested contents) to PyTorch tensors. Args: data: Arbitrary input to convert. Can be a scalar, NumPy array, tensor, mapping (e.g. dict), or sequence (e.g. list/tuple).

Returns:

Type	Description
`Any`	Converted object where all supported leaves are PyTorch tensors,
`Any`	and the original container structure (dict/list) is preserved.

Raises:

Type	Description
`TypeError`	If `data` (or some nested leaf) has a type that cannot be converted to a tensor.

FinalFeatures

Assemble a final feature tensor and manage bookkeeping fields in a sample dict.

This transform is typically used at the end of a preprocessing pipeline to:

Build a unified feature array under the key "feat" by selecting and concatenating one or more existing fields from data_dict.
Optionally remove some intermediate fields (e.g., "norm", auxiliary features) to save memory.
Optionally add offset fields ("offset", "fps_offset") that are useful when batching variable-length point clouds.

Behavior:

Feature construction:
If feat is a string, data_dict["feat"] is set to data_dict[feat].
If feat is a list/tuple of strings, the corresponding arrays are concatenated along the last dimension:
```
feat = np.concatenate([data_dict[name] for name in feat], axis=-1)
```

All specified feature names must exist in data_dict.

Field removal:
If remove is a string, that key is deleted from data_dict.
If remove is a list/tuple of strings, each corresponding key is deleted.
All specified keys must exist in data_dict.
Offsets:
If add_offset is True and "offset" is not already present, then:
```
data_dict["offset"] = len(data_dict["coord"])
```
This is often interpreted as the number of points in this sample. - If add_fps_offset is True and "fps_offset" is not present but "fps_index" exists, then:
```
data_dict["fps_offset"] = len(data_dict["fps_index"])
```
This is typically the number of FPS (farthest-point sampling) indices for this sample.

Parameters:

Name	Type	Description	Default
`feat`	`str or sequence of str`	Name(s) of fields in `data_dict` to be used as final features. If a sequence, the corresponding arrays are concatenated along the last dimension and stored in `data_dict["feat"]`. If `None`, no feature tensor is constructed. Defaults to `"coord"`.	`'coord'`
`remove`	`str or sequence of str`	Name(s) of fields to delete from `data_dict` after (optional) feature construction. This is useful to drop intermediate fields (e.g., `"norm"`) that are no longer needed. If `None`, no keys are removed. Defaults to `"norm"`.	`'norm'`
`add_offset`	`bool`	If True and `"offset"` is not already present, add an integer offset equal to `len(data_dict["coord"])`. Defaults to True.	`True`
`add_fps_offset`	`bool`	If True, `"fps_offset"` is not present, and `"fps_index"` exists, add an integer offset equal to `len(data_dict["fps_index"])`. Defaults to True.	`True`

`call(data_dict)`

Construct the final feature array and update bookkeeping fields.