`torch.tensor` to `K.variable` ( `K` is `from keras import backend as K`)
  unsqueez_(1) to K.expand_dims
  torch.empty((3,) + requested_shape) to K.zeros((3,) + requested_shape)