时间重要性剪枝问题

[dccmr123]

你好！在时间重要性剪枝的相关代码里
def weight_prune(self, weight_threshold): prune_mask = (self._max_weight < weight_threshold)
在这句代码后面我添加了
print(f"[WeightPrune] prune={prune_mask.sum().item()} / {prune_mask.numel()} | " )
本意是想查看剪掉了多少漂浮点，却发现几次剪枝都输出[WeightPrune] prune=0 / 244622 | [18/05 13:53:16]这样的，显示是没有剪掉任何点，为什么？使用的数据集是N3DV的cut_roasted_beef，即论文里展示剪枝效果所用的那个数据集

[dccmr123]

我想着可能得调整下阈值，于是又添加如下代码

print("max_weight min/mean/median/max:",
            self._max_weight.min().item(),
            self._max_weight.mean().item(),
            self._max_weight.median().item(),
            self._max_weight.max().item())
print("below 0.005:", (self._max_weight < 0.005).float().mean().item())
print("below 0.01 :", (self._max_weight < 0.01).float().mean().item())
print("below 0.02 :", (self._max_weight < 0.02).float().mean().item())
print("below 0.05 :", (self._max_weight < 0.05).float().mean().item())

结果打印出来是这样的[WeightPrune] prune=0 / 221642 | [18/05 14:41:01]
max_weight min/mean/median/max: 0.9900000095367432 0.9900000691413879 0.9900000095367432 0.9900000095367432 [18/05 14:41:01]
below 0.005: 0.0 [18/05 14:41:01]
below 0.01 : 0.0 [18/05 14:41:01]
below 0.02 : 0.0 [18/05 14:41:01]
below 0.05 : 0.0 [18/05 14:41:01]

[WuJH2001]

Hi @dccmr123 ,非常感谢您的关注、按照代码设置会周期性的根据max_weight 进行裁剪、然后max_weight 会置0，开启新的一轮计算最大权重，应该不至于所有高斯点的权重都很大（也就是贡献度很高），您是否可以提供一下当前的训练轮次是多少，当前代码设置的是
weight_prune_from_iter = 3000,
weight_prune_interval = 3000,
weight_prune_threshold = 0.02,
这个裁剪策略在一开始的时候才会裁掉很多点、到后期冗余点裁剪完了可能就不会进行裁剪了

[dccmr123]

Hi @dccmr123 ,非常感谢您的关注、按照代码设置会周期性的根据max_weight 进行裁剪、然后max_weight 会置0，开启新的一轮计算最大权重，应该不至于所有高斯点的权重都很大（也就是贡献度很高），您是否可以提供一下当前的训练轮次是多少，当前代码设置的是 weight_prune_from_iter = 3000, weight_prune_interval = 3000, weight_prune_threshold = 0.02, 这个裁剪策略在一开始的时候才会裁掉很多点、到后期冗余点裁剪完了可能就不会进行裁剪了
你好!
weight_prune_from_iter = 3000,
weight_prune_interval = 3000,
weight_prune_threshold = 0.02,这几个参数并未改动，迭代次数也没有改动，事实上相较于原始代码，只添加了这些打印输出的内容，
完整的输出如下

Output folder: ./output/dynerf/cut_roasted_beef [18/05 14:27:33]
Calculating std: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:00<00:00, 335.09it/s]
meta data loaded, total image:5700 [18/05 14:27:35]
Calculating std: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 20/20 [00:00<00:00, 346.05it/s]
meta data loaded, total image:300 [18/05 14:27:35]
100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 5700/5700 [00:00<00:00, 255629.92it/s]
300it [00:00, 48590.18it/s]
origin points, 37207 [18/05 14:27:35]
after points, 37207 [18/05 14:27:35]
Loading Training Cameras [18/05 14:27:35]
Loading Test Cameras [18/05 14:27:35]
Loading Video Cameras [18/05 14:27:35]
Number of points at initialisation :  37207 [18/05 14:27:35]
data loading done [18/05 14:27:35]
select index: [0, 300, 600, 900, 1200, 1500, 1800, 2100, 2400, 2700, 3000, 3300, 3600, 3900, 4200, 4500, 4800, 5100, 5400] [18/05 14:27:36]
Training progress:  60%|████████████████████████████████████████████████                                | 3000/4999 [03:01<02:04, 16.05it/s, Loss=0.0172354, psnr=34.04, point=162534]
[ITER 3000] Saving Gaussians [18/05 14:30:40]
[WeightPrune] prune=0 / 162534 |  [18/05 14:30:42]
max_weight min/mean/median/max: 0.9900000095367432 0.9900000691413879 0.9900000095367432 0.9900000095367432 [18/05 14:30:42]
below 0.005: 0.0 [18/05 14:30:42]
below 0.01 : 0.0 [18/05 14:30:42]
below 0.02 : 0.0 [18/05 14:30:42]
below 0.05 : 0.0 [18/05 14:30:42]
reset opacity [18/05 14:30:42]
Training progress: 100%|███████████████████████████████████████████████████████████████████████████████▊| 4990/4999 [05:23<00:00, 15.42it/s, Loss=0.0140132, psnr=37.27, point=179590]
The num of dynamic point: 0 [18/05 14:33:02]
select index: [0, 300, 600, 900, 1200, 1500, 1800, 2100, 2400, 2700, 3000, 3300, 3600, 3900, 4200, 4500, 4800, 5100, 5400] [18/05 14:33:02]
Dynamic Training progress: 3000it [01:20, 37.17it/s, Loss=0.0533851]                                                                                                                  
data loading done [18/05 14:34:24]
Training progress:  22%|█████████████████▎                                                             | 2850/13000 [06:09<21:56,  7.71it/s, Loss=0.0203348, psnr=33.75, point=220576]reset dataloader into random dataloader. [18/05 14:40:34]
Training progress:  23%|██████████████████▏                                                            | 3000/13000 [06:30<21:57,  7.59it/s, Loss=0.0179475, psnr=32.90, point=221642]
[ITER 3000] Evaluating test: L1 0.014460309627739821 PSNR 33.15666580200195 [18/05 14:40:58]

[ITER 3000] Evaluating train: L1 0.012190045734100482 PSNR 33.673935160917395 [18/05 14:40:58]

[ITER 3000] Saving Gaussians [18/05 14:40:58]
[WeightPrune] prune=0 / 221642 |  [18/05 14:41:01]
max_weight min/mean/median/max: 0.9900000095367432 0.9900000691413879 0.9900000095367432 0.9900000095367432 [18/05 14:41:01]
below 0.005: 0.0 [18/05 14:41:01]
below 0.01 : 0.0 [18/05 14:41:01]
below 0.02 : 0.0 [18/05 14:41:01]
below 0.05 : 0.0 [18/05 14:41:01]
Training progress:  44%|██████████████████████████████████▋                                            | 5700/13000 [12:28<15:33,  7.82it/s, Loss=0.0175905, psnr=34.15, point=241639]reset dataloader into random dataloader. [18/05 14:46:53]
Training progress:  46%|████████████████████████████████████▍                                          | 6000/13000 [13:09<15:15,  7.65it/s, Loss=0.0156357, psnr=35.43, point=243470][WeightPrune] prune=0 / 243470 |  [18/05 14:47:34]
max_weight min/mean/median/max: 0.9900000095367432 0.9900000691413879 0.9900000095367432 0.9900000095367432 [18/05 14:47:34]
below 0.005: 0.0 [18/05 14:47:34]
below 0.01 : 0.0 [18/05 14:47:34]
below 0.02 : 0.0 [18/05 14:47:34]
below 0.05 : 0.0 [18/05 14:47:34]
Training progress:  54%|██████████████████████████████████████████▌                                    | 7000/13000 [15:20<12:53,  7.76it/s, Loss=0.0148238, psnr=34.75, point=247982]
[ITER 7000] Evaluating test: L1 0.014770066639518036 PSNR 33.220141018138214 [18/05 14:49:47]

[ITER 7000] Evaluating train: L1 0.010020437979084603 PSNR 35.575673271627984 [18/05 14:49:47]

[ITER 7000] Saving Gaussians [18/05 14:49:47]
Training progress:  66%|███████████████████████████████████████████████████▉                           | 8550/13000 [18:54<09:56,  7.46it/s, Loss=0.0159899, psnr=35.64, point=253621]reset dataloader into random dataloader. [18/05 14:53:20]
Training progress:  69%|██████████████████████████████████████████████████████▋                        | 9000/13000 [19:55<08:46,  7.60it/s, Loss=0.0143371, psnr=36.41, point=255345][WeightPrune] prune=0 / 255345 |  [18/05 14:54:20]
max_weight min/mean/median/max: 0.9900000095367432 0.9900000095367432 0.9900000095367432 0.9900000095367432 [18/05 14:54:20]
below 0.005: 0.0 [18/05 14:54:20]
below 0.01 : 0.0 [18/05 14:54:20]
below 0.02 : 0.0 [18/05 14:54:20]
below 0.05 : 0.0 [18/05 14:54:20]
Training progress:  77%|████████████████████████████████████████████████████████████                  | 10000/13000 [22:09<06:39,  7.51it/s, Loss=0.0142120, psnr=36.30, point=258011]
[ITER 10000] Evaluating test: L1 0.018403940779321334 PSNR 31.385351293227252 [18/05 14:56:36]

[ITER 10000] Evaluating train: L1 0.00971271308577236 PSNR 35.959260155172906 [18/05 14:56:36]

[ITER 10000] Saving Gaussians [18/05 14:56:36]
Training progress:  88%|████████████████████████████████████████████████████████████████████▍         | 11400/13000 [25:18<03:27,  7.73it/s, Loss=0.0157355, psnr=35.15, point=258011]reset dataloader into random dataloader. [18/05 14:59:43]
Training progress: 100%|██████████████████████████████████████████████████████████████████████████████| 13000/13000 [28:47<00:00,  7.53it/s, Loss=0.0143746, psnr=36.00, point=258011]
The num of dynamic point: 48879 [18/05 15:03:12]

[ITER 13000] Evaluating test: L1 0.01815791252781363 PSNR 31.613924699671127 [18/05 15:03:14]

[ITER 13000] Evaluating train: L1 0.008795289432301241 PSNR 36.91127844417797 [18/05 15:03:15]

[ITER 13000] Saving Gaussians [18/05 15:03:15]

[WuJH2001]

真的很奇怪、我检查一下代码，为什么他们的最大最小权重都是 0.9900000095367432