empty elements were found in the dataset

[yiwang454]

Hi authors,
Thanks for providing this abx tool. I met an error of "fastabx.verify.EmptyDataPointsError: 4 empty elements were found in the dataset (with indices [170498, 264684, 397790, 505072])". I would like to ask what could be the usual reason for encountering this issue, and any preprocessing approach is recommended to avoid this error.

I'm now using it with my data with the following codes:

    item, frequency = "./largeclass_tenth.item", 50
    features = f"./data/hubert_feature/large_l{layer}_mic1"
    dataset = Dataset.from_item(item, features, frequency)
    subsampler = Subsampler(max_size_group=10, max_x_across=2)
    task = Task(dataset, on="#phone", by=["next-phone", "prev-phone"], across=["speaker"], subsampler=subsampler,)

The codes worked for part of my data and obtained abx error rate at expected scale. But when using some other parts of the data, I got the following error (hubert_abx.py is the file containing the above scripts):

Building dataset: 100%|
Traceback (most recent call last):
  File "/mnt/ceph_rbd/workspace/fastabx/scripts/hubert_abx.py", line 38, in <module>
    main()
  File "/mnt/ceph_rbd/workspace/fastabx/scripts/hubert_abx.py", line 17, in main
    dataset = Dataset.from_item(item, features, frequency)
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/mnt/ceph_rbd/workspace/fastabx/src/fastabx/dataset.py", line 274, in from_item
    return Dataset(labels=labels, accessor=InMemoryAccessor(indices, data))
                                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/mnt/ceph_rbd/workspace/fastabx/src/fastabx/dataset.py", line 65, in __init__
    verify_empty_datapoints(self.indices)
  File "/mnt/ceph_rbd/workspace/fastabx/src/fastabx/verify.py", line 61, in verify_empty_datapoints
    raise EmptyDataPointsError(empty)
fastabx.verify.EmptyDataPointsError: 4 empty elements were found in the dataset (with indices [170498, 264684, 397790, 505072])

Meanwhile, I have already modified the load_data_from_item function in the dataset module, by padding the feature to the slicing end index:

            if end > features.size(0):
                missing_frame = end - features.size(0)
                features = torch.cat([features] + [features[-1:] for _ in range(missing_frame)], dim=0)

So it shouldn't be caused by that features are short in length.

[mxmpl]

Hi!
When computing the ABX in your setting, this exception is raised if some entries of the item file correspond to empty features.
In your case, that would mean that the 170498, 264684, 397790, and 505072-th entries of your item file
are less than 20ms, the size of one feature frame.

You can check your item file with this:

import polars as pl
from fastabx.dataset import dummy_dataset_from_item

dataset = dummy_dataset_from_item("./largeclass_tenth.item", frequency=50)
print(dataset.labels.with_row_index().filter(pl.col("end") <= pl.col("start")))

This should print the four entries.

The first and last indices (both included) used to slice the input features are:

$$\begin{align} i_\text{start} = \left\lceil \frac{t_\text{on}}{\Delta t} - \frac{1}{2} \right\rceil, \\\ i_\text{end} = \left\lfloor \frac{t_\text{off}}{\Delta t} - \frac{1}{2} \right\rfloor. \end{align}$$

where $\Delta t = 0.02$ seconds in your case (50Hz framerate), and $t_\text{on}$ and $t_\text{off}$ are the onset and offset times in your item file (cf https://docs.cognitive-ml.fr/fastabx/advanced/slicing.html).

Please tell me if this works for you.
I will improve the exception message in the next release!

[mxmpl]

This commit adds a check for empty features in load_data_from_item: 681b49f
It is available if you upgrade to version 0.5.0.