@dpj135 we have, training was done using the horribly unreadable code in modeling_iter.py, I reworked the file and missed that bit. It should be there now, if you have anything else, feel free to ask.

num_left_to_nest_depth_index = unsqueeze_and_expand_dims(
                # [B, N, nest_depth] -> [B, N, nest_depth, num_right, num_left, T]
                closest_left_indices,
                (3, 4, 5),
                (3, num_right), (4, num_left), (5, self.num_types),
            )
            rr_pair_flag = torch.gather(
                rr_pair_flag,
                dim=2,
                index=num_left_to_nest_depth_index
            )
            # [B, N, nest_depth, num_right, nest_depth, T]
            # here is the problem
            rr_pair_flag = torch.gather(
                rr_pair_flag,
                dim=4,
                index=unsqueeze_and_expand_dims(
                    is_right_closest_left_indices,
                    (1, 2, 5),
                    (1, seq_len), (2, nest_depth), (5, self.num_types)
                )
            )

Thank you！I find some problems with your code in iter/modeling_iter2.py:line:780&795. I think self.num_types may be self.num_links.
I found that there is not enough video memory during training. I use 3090 24GB. May I ask if you have any suggestions？

Prombles about datasets #1

Description

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions