Skip to content

Latest commit

 

History

History
64 lines (58 loc) · 7.29 KB

Leaderboard.md

File metadata and controls

64 lines (58 loc) · 7.29 KB
Raw

Leaderboard

模型名 acc (%) 训练时间 权重及训练日志下载地址 备注
ResNet50-E5 86.01 6:50:28 百度网盘，谷歌网盘
ResNet50-E5+Stochastic Depth 86.12 6:53:32 dp=0.1
ResNet50-E5+Stochastic Depth 86.06 7:57:22 dp=0.2
ResNet50-E5+Stochastic Depth 86.34 3:42:15 4卡RTX3090，混合精度训练，单卡批量大小64，总批量大小256，学习率0.1，dp=0.1
ResNet50-E5+Drop Block 85.82 8:12:54 db=0.1
WRN-50-2-E5 85.74 12:35:44
ResNet-101-E5 85.50 10:02:03
ResNet-152-E5 86.38 14:57:37
ResNet-152-E5+Stochastic Depth 86.53 18:46:47 dp=0.5（跑实验的时候，还有其他的实验也占了大量显存，整张卡显存几乎占满了。训练时间仅供参考）
ResNet50-E6-tiny 85.49 10:16:27
ResNet50-E6-small 86.32 19:47:55
WRN-18-10-E6 (resnet18_10_E6) 85.03 10:45:00
WRN-18-10-E6-v2 (resnet18_10_E6_v2) 85.22 7:48:08
WRN-26-10-E6-v2 (resnet26_10_E6_v2) 85.00 16:04:39
WRN-34-10-E6-v2 (resnet34_10_E6_v2) 85.55 19:06:49
ResNet272-E6-v2 (resnet272_E6_v2) 86.39 18:14:55 双卡RTX3090
下面是WRN原文中定义的模型（未修改架构）
WRN-16-8 81.27 3:16:37
WRN-28-10 85.29 11:45:23
WRN-40-10 85.65 16:37:49
下面是ResNet原文中定义的处理CIFAR数据集的模型（未修改架构）
ResNet20 66.87 1:31:27
ResNet32 71.18 2:08:42
ResNet44 73.56 2:48:14
ResNet56 75.27 3:24:39
ResNet110 77.90 6:23:43
resnet1202 81.06 2 days, 15:42:18 4卡RTX3090

上述结果默认训练脚本：

#!/bin/bash

for model in 'model_name'
do
    torchrun --nproc_per_node=1  --master_port="29429" classification/train.py \
        --model ${model} \
        --model_lib custom \
        --data_name cifar100 \
        --batch-size 128 \
        --lr 0.1 \
        --lr-scheduler cosineannealinglr \
        --epochs 300 \
        --lr-warmup-epochs 20 \
        --lr-min 1e-6 \
        --wd 5e-4 \
        --auto_augment \
        --random_erase 0.25 \
        --mixup-alpha 1 \
        --cutmix-alpha 1 \
        --act_layer relu \
        --loss_type ce \
        --print-freq 100 \
        --output-dir ./work_dir/aa-re_0.25-mixup-cutmix \
        --data-path /path/to/cifar100

    wait
done