evn - demo

July 21, 2026 · View on GitHub

conda create -n LTAP 3.8.11   
source activate LTAP 
pip install -r requirements.txt
python main.py --save_all_epoch --hesso --target_group_sparsity 0.1 --isc LT-vote --imb_ratio 100  --bs --gpu 1 > bs_1.log  2>&1 &

All the command templates have been placed in 1.sh; you can run it directly or execute individual commands from it.

About the large-scale datasets (ImageNet-LT & iNaturalist) / 关于两个大数据集

Due to copyright restrictions, we are unable to provide the data and the large-scale training code for the two larger datasets (ImageNet-LT and iNaturalist) in this repository. The missing pieces can, however, be traced back to the two upstream projects this work is built on:

Pruning framework — Only Train Once (OTO) (now maintained at microsoft/only_train_once): the birthplace of the only_train_once / HESSO / sanity_check code included here. The pruning itself (HESSO) is dataset-agnostic — it acts on the model, not the data — so it plugs into any training loop regardless of dataset size. Large-scale examples (e.g. ResNet50 on ImageNet) live in the OTOv2 branch's main experiments and the distributed-training (DDP) tutorial.
Long-tailed training pipeline — NCL (Nested Collaborative Learning), CVPR 2022: the earliest source of the long-tailed multi-expert code here (multi_Network_MOCO, Combiner, MoCo, the ncl loss, etc.). It provides full, runnable configs and training scripts for the large-scale long-tailed datasets — ImageNet-LT (config/ImageNet_LT), iNaturalist 2018 (config/iNat18), and Places-LT (config/Places_LT). This is the piece that is missing here due to copyright, and it is where you should look first.

Applying the method described in this paper to the large-scale setting only requires minor modifications on top of these two repositories.

由于版权限制，本仓库无法提供两个较大数据集（ImageNet-LT 与 iNaturalist）的数据与大规模训练代码。缺失的这部分可以追溯到本工作所依赖的两个上游项目：

剪枝框架 —— Only Train Once (OTO)： 本仓库中 only_train_once / HESSO / sanity_check 代码的“发源地”，其本身已包含适用于大数据集（如 ImageNet）的代码。

长尾训练流程 —— NCL（Nested Collaborative Learning），CVPR 2022： 本仓库长尾多专家代码（multi_Network_MOCO、Combiner、MoCo、ncl 损失等）最早的出处。这份“极早时期”的长尾代码提供了 ImageNet-LT、iNaturalist 2018、Places-LT 等大规模长尾数据集的完整配置与训练脚本。

只需在上述两个仓库的基础上稍作修改，即可将本文提出的方法应用到大规模数据集上。

Pseudocode

ir =50

Method	F	Head	Medium	Tail	All	C	C/F
ce	100.0	68.0	38.0	13.2	46.0	100.0	1.0
ce + ato	84.7	46.7	17.3	6.83	29.1	63.2	0.7
ce + regg	52.1	43.8	14.8	0.83	24.5	53.2	1.0
ce + ours w.o. $\kappa$	23.3	64.5	30.0	4.5	39.5	85.8	3.6
ce + ours	23.3	64.8	31.7	7.2	41.1	89.3	3.8
la	100.0	59.9	46.7	41.3	51.3	100.0	1.0
la + ato	84.7	34.5	33.9	29.1	34.2	66.7	0.7
la + regg	52.1	31.0	30.2	25.8	29.9	58.3	1.1
la + ours w.o. $\kappa$	22.8	53.5	42.7	30.0	44.8	87.3	3.8
la + ours	22.8	54.0	43.4	38.4	47.1	91.8	4.0

ir = 100

Method	F	Head	Medium	Tail	All	C	C/F
ce	100.0	70.7	40.0	7.2	41.0	100.0	1.0
ce + ato	84.7	50.4	16.5	6.6	25.2	61.5	0.7
ce + regg	52.1	47.7	13.6	0.6	21.9	53.4	1.0
ce + ours w.o. $\kappa$	23.3	67.1	30.8	0.5	34.4	83.9	3.6
ce + ours	23.3	66.1	31.7	2.5	35.1	85.6	3.6
la	100.0	62.9	47.7	29.6	47.9	100.0	1.0
la + ato	84.7	42.1	30.6	18.4	31.4	65.6	0.7
la + regg	52.1	38.5	27.0	14.6	27.6	57.6	1.1
la + ours w.o. $\kappa$	22.8	55.1	42.0	18.6	39.5	82.4	3.6
la + ours	22.8	56.1	45.3	22.6	42.6	88.9	3.9