Do you think the _meta should be removed as Batch?

Since you are asking, yes I think so. It makes everything pretty difficult to understand and maintain without adding any feature at the end. However, I'm afraid I don't really understand how this mechanism is working right now, so I'm not able to refactor this submodule by myself. It would be awesome to take advantage of the newly added flexibility of Batch to simplify the implement of this module.

It is almost the same as the previous Batch implementation. The difference includes automatically create the np.array with given data. Nothing special.

Ok I'll have a look at it.

I think I should be able to handle it finally. I let you know.

Doing batch.obs[indices] = ... (item assignment) triggers this broadcasting, while batch.obs = ... (direct assignment) does not. It is important, since item assignment must not break item access, which means doing afterward batch.obs[indices]must be valid under any condition, which is not the case if you do not broadcast, while direct assignment has not such guarantee.

Another solution is that this new feature can be only appeared in Buffer instead of Batch.

Finally I will do this, since it is impossible to get a reliable behaviour using broadcasting (lower level nested Batch instances have no idea of the size of the top Batch, making it impossible to perform broadcast), and it was pretty hard to maintain.

@Trinkle23897 PR ready to be merged ! All tests passing successfully finally !

By the way, I found that in #92, the overhead of Batch.__getitem__ grows 20% because of calling _valid_bounds. Do you have any idea?

YEs, I'm on it !

@Trinkle23897 Can you check this branch ? #95