Reduced-Order Model of Time-Projection Chamber

These datasets are generated by a reduced-order model (ROM) of the sPHENIX Time-projection chamber (TPC), inner layer group. We have two datasets:

`dataset.zip` and

`dataset_with_label.zip`.

Each dataset contains 8000 training examples and 2000 test examples. Each sample in the dataset contains ~50 to 100 trajectories. Samples in both dataset contain fields: `response`, `tracks`, and `params `. Samples in the `dataset_with_label.zip` has an additional field `label`. 

Here are the description of fields:

`response` (shared): float tensor of shape (16, 256, 1152) in radial (layer), axial (horizontal), and azimuthal directions. 

`tracks` (shared): float tensor of shape (N, 2, 16), where N is the number of trajectories. Each trajectory is recorded as a tensor of shape (2, 16). The two numbers for each layer (last axis) is the angle (in the azimuthal direction) and axial location of the trajectory. 

`params` (shared): float tensor of shape (N, 4). The four numbers for the n-th entry are initial momentum (3D) and mass-charge ratio (scalar) of the particle that generates the n-th trajectory.

`label` (`dataset_with_label` only): integer tensor of shape (16, 256, 1152), the value for voxels with positive response is the trajectory id that contribute the most to the voxel. The value for voxels with zero response is random and should not be used.