model/utils/tensor.py

import torch
import torch.nn.functional as F


def max_shape(data):
    """Gets the maximum length along all dimensions in a list of Tensors"""
    shapes = torch.Tensor([_.shape for _ in data])
    return torch.max(
        shapes.transpose(0, 1),
        dim=1,
    )[0].int()


def pad_data_to_same_shape(
    tensor_list,
    pad_value=0,
):
    target_shape = max_shape(tensor_list)

    padded_dataset_shape = [len(tensor_list)] + list(target_shape)
    padded_dataset = torch.Tensor(*padded_dataset_shape).type_as(
        tensor_list[0])

    for i, data in enumerate(tensor_list):
        # Get how much padding is needed per dimension
        padding = reversed(target_shape - torch.Tensor(list(data.shape)).int())

        # Add 0 every other index to indicate only right padding
        padding = F.pad(
            padding.unsqueeze(0).t(),
            (1, 0, 0, 0),
        ).view(-1, 1)
        padding = padding.view(1, -1)[0].tolist()

        padded_data = F.pad(
            data,
            padding,
            value=pad_value,
        )
        padded_dataset[i] = padded_data

    return padded_dataset