ValueError: Window (25) cannot be longer than the sequences

[Aria-D]

Hi,

I am trying to run tfmodisco-lite on my one-hot encoded sequences and calculated attributions but am getting this error with "-w" argument to "modisco motifs".

In this error message, I tried it with a window size of 25, but even with a size of 1 the error message persists. The size of my attrs and seqs are:
SEQ: (46542, 4, 200)
ATTR: (46542, 4, 200)

which I believe is consistent with the tutorial? Even if I try transposing or swapping the 4-channel dimension, I get the same error. I was able to run modisco-lite with no issues a couple of months ago. I started getting this error ~2 weeks ago.

Any help with this issue would be great!

[isaacvock]

I think this error only occurs if you choose an odd window size (as in not evenly divisible by 2). If you set the window size to an even number, this error should go away.

I think the problem is due to some quirks in the following code in the CLI:

    if h5_path:
        f = h5py.File(h5_path, "r")
        try:
            center = f["hyp_scores"].shape[1] // 2
            start, end = calculate_window_offsets(center, window)
            attributions = f["hyp_scores"][..., :][..., start:end, :]
            sequences = f["input_seqs"][..., :][..., start:end, :]
        except KeyError:
            center = f["shap"]["seq"].shape[2] // 2
            start, end = calculate_window_offsets(center, window)
            attributions = f["shap"]["seq"][..., :, start:end].transpose(0, 2, 1)
            sequences = f["raw"]["seq"][..., :, start:end].transpose(0, 2, 1)
        f.close()
    else:
        seq_path = Path(seq_path)
        attr_path = Path(attr_path)
        sequences = (
            np.load(seq_path)["arr_0"]
            if seq_path.suffix == ".npz"
            else np.load(seq_path)
        )
        attributions = (
            np.load(attr_path)["arr_0"]
            if attr_path.suffix == ".npz"
            else np.load(attr_path)
        )
        center = sequences.shape[2] // 2
        start, end = calculate_window_offsets(center, window)
        sequences = sequences[:, :, start:end].transpose(0, 2, 1)
        attributions = attributions[:, :, start:end].transpose(0, 2, 1)

This interacts poorly with the calculate_window_offsets function:

def calculate_window_offsets(center: int, window_size: int) -> tuple:
	return (center - window_size // 2, center + window_size // 2)

This causes the array slice to have one less element than the window size if the window size is odd.