the dimension of pre_status and cur_status NOT equal in leave_one_out_stage2

[husthzy]

I trained with my own dataset.

python leave_one_out_stage2.py -s data/realcap/mario -m output/gs_init/mario_loo -r 8 --sparse_view_num 4 --sh_degree 2 --init_pcd_name dust3r_4 --dust3r_json output/gs_init/mario/refined_cams.json --white_background --random_background --use_dust3r

This is the error info:
`
[ITER 7000] Evaluating train: L1 0.005111759598366916 PSNR 30.6793571472168 [29/11 17:37:02]
Training progress: 100%|██████████████████████████████████████████████| 24000/24000 [04:38<00:00, 86.09it/s, Loss=0.0032718, n=29527]

[ITER 30000] Saving Gaussians [29/11 17:41:22]
key:_xyz [29/11 17:41:22]
pre_c shape:torch.Size([29667, 3]) [29/11 17:41:22]
cur_c shape:torch.Size([29527, 3]) [29/11 17:41:22]
Traceback (most recent call last):
File "/home/code/GaussianObject/leave_one_out_stage2.py", line 328, in
train_3dgs(args, ids)
File "/home/code/GaussianObject/leave_one_out_stage2.py", line 280, in train_3dgs
leave_one_out_training(args,
File "/home/code/GaussianObject/leave_one_out_stage2.py", line 136, in leave_one_out_training
diff = pre_c - cur_c
RuntimeError: The size of tensor a (29667) must match the size of tensor b (29527) at non-singleton dimension 0
`

[7rwang]

Fixed? Sometimes I will also encounter the same error

[husthzy]

Fixed? Sometimes I will also encounter the same error

Not yet. I am trying to reinstall the env

[husthzy]

set iterations = 10_000, it will be ok.
set iterations = 30_000, it will be error.

I have not know the reason.

[7rwang]

set iterations = 10_000, it will be ok. set iterations = 30_000, it will be error.

I have not know the reason.

It's weird. Cuz from my perspective, prune is random, it's difficult to keep the number unchanged in two stages

[husthzy]

set iterations = 10_000, it will be ok. set iterations = 30_000, it will be error.
I have not know the reason.

It's weird. Cuz from my perspective, prune is random, it's difficult to keep the number unchanged in two stages

Since pruning will occur, why does the code directly subtract two statuses?

I added a check to force the dimensions to be consistent, and the training worked fine, but I can't confirm whether this is correct.

[UranusITS]

In the leave-one-out Stage 1, we save the 3DGS checkpoint using the checkpoint_iterations argument, which is fixed to 6000.

GaussianObject/leave_one_out_stage1.py

Line 304 in 91048a5

parser.add_argument("--checkpoint_iterations", nargs="+", type=int, default=[6000])

In Stage 2, we directly load the checkpoint file, and this causes the unexpected issue when changing the number of iterations. Relevant sections:

GaussianObject/leave_one_out_stage2.py

Line 29 in 91048a5

first_iter = 6000 # in this code, we just use the data from 6000 iter

GaussianObject/leave_one_out_stage2.py

Line 275 in 91048a5

args.start_checkpoint = os.path.join(args.model_path, 'chkpnt6000.pth') # load this ckpt

Apologies for not fully considering the case when changing the number of iterations.

[chensjtu]

exactly