Prevent the image dimensions in the encoder and the decoder from being different due to the odd image dimension, which will lead to sk

In the decoder for basis, the features are pooled to 1*1 size and then enlarged by replication before skip concatenation.

return size: (batch_size, burst_length, kernel_size ** 2, color_channel, height, width)

compute the pred image according to core and frames :param data: [batch_size, burst_length, color_channel, height, width] :pa

forward and obtain pred image directly :param data_with_est: if not blind estimation, it is same as data :param data:

forward function of loss_func :param pred_burst: shape [batch_size, burst_length, color_channel, height, width] :param pred:

:param global_step: int :param pred_burst: [batch_size, burst_length, color_channel, height, width] :param ground_truth: [bat