text_contexts = self.manifold.weighted_midpoint

[ZhouL2001]

Great Work！

When I'm training, I keep getting the following error: RuntimeError: The size of tensor a (256) must match the size of tensor b (22) at non-singleton dimension 3 in

Geo-Sign/models.py

Line 336 in 5cacfe5

text_contexts = self.manifold.weighted_midpoint(

. Could you please provide some assistance?

[ed-fish]

Hi, sorry you are having issues with the code - can you please provide me the command you are using to run the code. Thanks

[ZhouL2001]

Hi, sorry you are having issues with the code - can you please provide me the command you are using to run the code. Thanks
Command:
torchrun ${DISTRIBUTED_LAUNCH_ARGS[@]} fine_tuning.py
--batch_size 8
--gradient_accumulation_steps 8
--epochs 40
--opt AdamW
--lr 1e-4
--output_dir $output_dir
--finetune $ckpt_path
--dataset CSL_Daily
--task SLT
--use_hyperbolic
--wandb
--init_c 1.0
--hyp_text_emb_src decoder
--hyp_text_cmp token
--alpha 0.7 \

hyp_text_tokens.shape: B, T, D
weights: B, K, D

[ZhouL2001]

Hi, sorry you are having issues with the code - can you please provide me the command you are using to run the code. Thanks

Thank you for your reply.
Before this error, I encountered another error in https://github.com/ed-fish/Geo-Sign/blob/5cacfe5b6cfc560fea5af0eb48f1544882a5844a/models.py#L322：
[rank0]: Traceback of TorchScript (most recent call last):
[rank0]: File "/usr/local/lib/python3.10/dist-packages/geoopt/manifolds/stereographic/math.py", line 1238, in _mobius_matvec
[rank0]: x_norm = x.norm(dim=dim, keepdim=True, p=2).clamp_min(1e-15)
[rank0]: if dim != -1 or m.dim() == 2:
[rank0]: mx = torch.tensordot(x, m, ([dim], [1]))
[rank0]: ~~~~~~~~~~~~~~~ <--- HERE
[rank0]: else:
[rank0]: mx = torch.matmul(m, x.unsqueeze(-1)).squeeze(-1)
[rank0]: File "/usr/local/lib/python3.10/dist-packages/torch/functional.py", line 1382, in tensordot

[rank0]: if out is None:
[rank0]: return _VF.tensordot(a, b, dims_a, dims_b) # type: ignore[attr-defined]
[rank0]: ~~~~~~~~~~~~~ <--- HERE
[rank0]: else:
[rank0]: return _VF.tensordot(a, b, dims_a, dims_b, out=out) # type: ignore[attr-defined]
[rank0]: RuntimeError: both inputs should have same dtype

So, I changed this line to self.manifold.mobius_matvec(self.hyp_attn_W.float(), hyp_text_tokens.unsqueeze(1))

Looking forward to your reply. Thanks.

[ed-fish]

Okay thanks for letting me know. Please give me a couple of days to look into the issue.