DALI uses short strings (Python str type) to describe data layout in tensors, by assigning a character to each of the dimensions present in the tensor shape. For example, shape=(400, 300, 3), layout="HWC" means that the data is an image with 3 interleaved channels, 400 pixels of height and 300 pixels of width.
For TensorLists, the index in the list is not treated as a dimension (the number of sample in the batch) and is not included in the layout.
DALI allows you to process data of different nature (e.g. image, video, audio, volumetric images) as well as different formats (e.g. RGB image in planar configuration vs. interleaved channels). Typically, DALI operators can deal with different data formats and will behave in different way depending on the nature of the input.
While we do not restrict the valid characters to be used in a tensor layout, DALI operators assume a certain naming convention. Here is a list of commonly used dimension names:
| Name | Meaning |
|---|---|
| H | Height |
| W | Width |
| C | Channels |
| F | Frames |
| D | Depth |
Here are some examples of typically used layouts:
| Layout | Description |
|---|---|
| HWC | Image (interleaved) |
| CHW | Image (planar) |
| DHWC | Volumetric Image (interleaved) |
| CDHW | Volumetric Image (planar) |
| FHWC | Video |
For instance, a crop operation (Crop operator) receiving an input with interleaved layout ("HWC") will infer that it should crop on the first and second dimensions (H, W). On the other hand, if the input has a planar layout ("CHW") the crop will take place on the second and third dimensions instead.
Some operators inherently modify the layout of the data (e.g. Transpose), while others propagate the same data layout to the output (e.g. Normalize).
The layout restrictions (if any) for each operator are available through the operator's documentation.
It is worth to note that the user is responsible to explicitly fill in the layout information when using ExternalSource API.