Skip to content

zhengxiawu/pytorch_image_classification

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

69 Commits

data

data

 
 

models

models

 
 

script

script

 
 

.DS_Store

.DS_Store

 
 

.gitignore

.gitignore

 
 

README.md

README.md

 
 

config.py

config.py

 
 

flops_counter.py

flops_counter.py

 
 

test.py

test.py

 
 

timm_test.py

timm_test.py

 
 

timm_train.py

timm_train.py

 
 

train.py

train.py

 
 

utils.py

utils.py

 
 

Repository files navigation

Requirements

  • Pytorch 1.0.1.post2
  • Python 3.6+
  • DALI

Train

Step1: Go into your project path

cd /userhome/project/pytorch_image_classification;

Step2: Move data to memory

./script/data_to_memory.sh cifar10
./script/data_to_memory.sh imagenet

Step3: Start training

# Training with cifar10 DALI on different neural networks

python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --model_method manual --model_name MobileNetV2 --data_loader_type dali
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --model_method manual --model_name MobileNetV3Large --data_loader_type dali
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --model_method manual --model_name Resnet18 --data_loader_type dali

python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --model_method proxyless_NAS --model_name proxyless_gpu --data_loader_type dali 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --model_method proxyless_NAS --model_name proxyless_cpu --data_loader_type dali 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --model_method proxyless_NAS --model_name proxyless_mobile --data_loader_type dali 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --model_method proxyless_NAS --model_name proxyless_mobile_14 --data_loader_type dali 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --model_method proxyless_NAS --model_name ofa_595 --data_loader_type dali 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --model_method proxyless_NAS --model_name ofa_482 --data_loader_type dali
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --model_method proxyless_NAS --model_name ofa_398 --data_loader_type dali

# init channel 44 epoch 1800 dropout 0.7 will have a higher performance
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type dali --drop_path_prob 0.2 --aux_weight 0.4 --init_channels 36 --layers 20 --epochs 600 --model_method darts_NAS --model_name MDENAS
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type dali --drop_path_prob 0.2 --aux_weight 0.4 --init_channels 36 --layers 20 --epochs 600 --model_method darts_NAS --model_name DDPNAS_V1
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type dali --drop_path_prob 0.2 --aux_weight 0.4 --init_channels 36 --layers 20 --epochs 600 --model_method darts_NAS --model_name DDPNAS_V2
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type dali --drop_path_prob 0.2 --aux_weight 0.4 --init_channels 36 --layers 20 --epochs 600 --model_method darts_NAS --model_name DARTS_V1
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type dali --drop_path_prob 0.2 --aux_weight 0.4 --init_channels 36 --layers 20 --epochs 600 --model_method darts_NAS --model_name DARTS_V2

# Training with cifar10 Torch on different neural networks for high performance 

python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --model_method manual --model_name MobileNetV2 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --model_method manual --model_name MobileNetV3Large 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --model_method manual --model_name Resnet18 

python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --model_method proxyless_NAS --model_name proxyless_gpu 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --model_method proxyless_NAS --model_name proxyless_cpu 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --model_method proxyless_NAS --model_name proxyless_mobile 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --model_method proxyless_NAS --model_name proxyless_mobile_14 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --model_method proxyless_NAS --model_name ofa_595 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --model_method proxyless_NAS --model_name ofa_482 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --model_method proxyless_NAS --model_name ofa_398 

python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --drop_path_prob 0.2 --aux_weight 0.4 --model_method darts_NAS --model_name MDENAS
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --drop_path_prob 0.2 --aux_weight 0.4 --model_method darts_NAS --model_name DDPNAS_V1
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --drop_path_prob 0.2 --aux_weight 0.4 --model_method darts_NAS --model_name DDPNAS_V2
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --drop_path_prob 0.2 --aux_weight 0.4 --model_method darts_NAS --model_name DARTS_V1
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 600 --drop_path_prob 0.2 --aux_weight 0.4 --model_method darts_NAS --model_name DARTS_V2

python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 1800 --init_channels 44 --batch_size 96 --drop_path_prob 0.2 --aux_weight 0.4 --model_method darts_NAS --model_name MDENAS
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 1800 --init_channels 44 --batch_size 96 --drop_path_prob 0.2 --aux_weight 0.4 --model_method darts_NAS --model_name DDPNAS_V1 
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 1800 --init_channels 44 --batch_size 96 --drop_path_prob 0.2 --aux_weight 0.4 --model_method darts_NAS --model_name DDPNAS_V2
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 1800 --init_channels 44 --batch_size 96 --drop_path_prob 0.2 --aux_weight 0.4 --model_method darts_NAS --model_name DARTS_V1
python train.py --dataset cifar10 --data_path /userhome/temp_data/cifar10 --data_loader_type torch --auto_augmentation --cutout_length 16 --epochs 1800 --init_channels 44 --batch_size 96 --drop_path_prob 0.2 --aux_weight 0.4 --model_method darts_NAS --model_name DARTS_V2

# Training with ImageNet torch on different neural networks

python train.py --dataset imagenet --data_path /userhome/temp_data/ImageNet --data_loader_type dali --drop_path_prob 0.2 --aux_weight 0.4 --init_channels 48 --layers 14 --epochs 300 --model_method darts_NAS --model_name MDENAS
python train.py --dataset imagenet --data_path /userhome/temp_data/ImageNet --data_loader_type dali --model_method proxyless_NAS --model_name proxyless_gpu 

python train.py --dataset imagenet --data_path /userhome/temp_data/ImageNet --data_loader_type torch --epochs 300 --auto_augmentation --drop_path_prob 0.2 --aux_weight 0.4 --init_channels 48 --layers 14 --model_method darts_NAS --model_name MDENAS
python train.py --dataset imagenet --data_path /userhome/temp_data/ImageNet --data_loader_type torch --epochs 300 --auto_augmentation  --model_method proxyless_NAS --model_name proxyless_gpu

Train with Timm and apex

python timm_train.py --data /userhome/temp_data/ImageNet -b 256 --model_method mual --model_name MobileNetV3Large --sched step --epochs 450 --decay-epochs 2.4 --decay-rate .97 --opt rmsproptf --opt-eps .001 -j 8 --warmup-lr 1e-6 --weight-decay 1e-5 --drop 0.2 --drop-connect 0.2 --model-ema --model-ema-decay 0.9999 --aa rand-m9-mstd0.5 --remode pixel --reprob 0.2 --amp --lr .016

Results

ImageNet

train.py

Model Epoch Dropout LabelSmooth FLOPs Result
MobileNetV2 150 0.0 0.1 300.774 71.67
MobileNetV3 150 0.0 0.1 216.590 72.93
proxyless_mobile_14 150 0.0 0.1 580.883 75.28
proxyless_mobile 150 0.0 0.1 320.428 73.41
proxyless_gpu 150 0.0 0.1 465.260 73.93
proxyless_cpu 150 0.0 0.1 439.244 74.15
ofa_595 150 0.0 0.1 512.862 75.59
ofa_482 150 0.0 0.1 482.413 75.36
ofa_398 150 0.0 0.1 389.488 74.61
my_600_cifar10 150 0.0 0.1 570.014 75.26
my_500_cifar10 150 0.0 0.1 494.585 74.99
my_400_cifar10 150 0.0 0.1 395.348 73.36
MobileNetV2 300 0.2 0.1 300.774 72.53
MobileNetV3 300 0.2 0.1 216.590 73.94
proxyless_mobile_14 300 0.2 0.1 580.883 76.47
proxyless_mobile 300 0.2 0.1 320.428 74.36
proxyless_gpu 300 0.2 0.1 465.260 75.42
proxyless_cpu 300 0.2 0.1 439.244 75.00
ofa_595 300 0.2 0.1 512.862 76.87
ofa_482 300 0.2 0.1 482.413 76.63
ofa_398 300 0.2 0.1 389.488 75.90
my_600_cifar10 300 0.0 0.1 570.014 75.97
proxyless_mobile_14 300 0.0 0.1 580.883 75.27
proxyless_gpu 300 0.0 0.1 465.260 74.16
ofa_482 300 0.0 0.1 482.413 75.80

timm_train.py

Model Epoch Dropout LabelSmooth FLOPs Result
MobileNetV2 450 0.2 0.1 300.774 73.41
MobileNetV2 250 0.2 0.1 300.774 72.41
MobileNetV3 450 0.2 0.1 216.590 75.492
MobileNetV3 250 0.2 0.1 216.590 75.41
MobileNetV3 200 0.2 0.1 216.590 75.052
MobileNetV3 150 0.2 0.1 216.590 74.426

CIFAR-10

Model Epoch Dropout LabelSmooth FLOPs Result
MobileNetV2 300 0.0 0.1 6.125 82.73
MobileNetV3Large 300 0.0 0.1 7.087 83.00
Resnet18 300 0.0 0.1 555.42 93.59
my_400 300 0.0 0.1 10.641 86.30
my_500 300 0.0 0.1 13.323 86.35
my_600 300 0.0 0.1 15.236 85.82
ofa_398 300 0.0 0.1 12.115 84.02
ofa_482 300 0.0 0.1 14.386 85.37
ofa_595 300 0.0 0.1 15.029 85.72
proxyless_cpu 300 0.0 0.1 8.949 82.85
proxyless_gpu 300 0.0 0.1 9.477 80.99
proxyless_mobile 300 0.0 0.1 6.526 81.28
proxyless_mobile_14 300 0.0 0.1 11.836 82.91

Experiments

CIFAR-10

Model Epoch Dropout LabelSmooth FLOPs Result
ofa__dataset_cifar10_width_multi_1.2_epochs_200_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:100 300 0.0 0.1 4.162 79.74
ofa__dataset_cifar10_width_multi_1.2_epochs_200_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:200 300 0.0 0.1 7.889 80.94
ofa__dataset_cifar10_width_multi_1.2_epochs_200_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:300 300 0.0 0.1 9.920 84.08
ofa__dataset_cifar10_width_multi_1.2_epochs_200_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:400 300 0.0 0.1 12.271 85.55
ofa__dataset_cifar10_width_multi_1.2_epochs_200_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:500 300 0.0 0.1 14.274 85.37
ofa__dataset_cifar10_width_multi_1.2_epochs_200_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:600 300 0.0 0.1 14.984 85.67
proxyless__dataset_cifar10_width_multi_1.3_epochs_200_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:100 300 0.0 0.1 1.965 79.62
proxyless__dataset_cifar10_width_multi_1.3_epochs_200_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:200 300 0.0 0.1 4.083 83.75
proxyless__dataset_cifar10_width_multi_1.3_epochs_200_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:300 300 0.0 0.1 6.124 82.54
proxyless__dataset_cifar10_width_multi_1.3_epochs_200_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:400 300 0.0 0.1 8.127 84.87
proxyless__dataset_cifar10_width_multi_1.3_epochs_200_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:500 300 0.0 0.1 10.148 84.70
proxyless__dataset_cifar10_width_multi_1.3_epochs_200_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:600 300 0.0 0.1 12.123 85.51
ofa_cifar10_width_multi_1.2_epochs_1000_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:600 300 0.0 0.1 13.719 85.55
ofa_cifar10_width_multi_1.2_epochs_1000_data_split_10_warm_up_epochs_0_lr_0.1_pruning_step_3:600 300 0.0 0.1 16.544 86.37
ofa_cifar10_width_multi_1.2_epochs_1000_data_split_10_warm_up_epochs_0_lr_0.1_pruning_step_6:600 300 0.0 0.1 12.259 86.11
ofa_cifar10_width_multi_1.2_epochs_1000_data_split_10_warm_up_epochs_0_lr_0.1_pruning_step_9:600 300 0.0 0.1 15.372 86.76
ofa_cifar10_width_multi_1.2_epochs_1000_data_split_10_warm_up_epochs_10_lr_0.1_pruning_step_3:600 300 0.0 0.1 12.846 85.76
ofa_cifar10_width_multi_1.2_epochs_1000_data_split_10_warm_up_epochs_10_lr_0.1_pruning_step_3_1:600 300 0.0 0.1 15.458 86.11
ofa_cifar10_width_multi_1.2_epochs_1000_data_split_10_warm_up_epochs_10_lr_0.1_pruning_step_3_2:600 300 0.0 0.1 15.444 86.48
ofa_cifar10_width_multi_1.2_epochs_1000_data_split_10_warm_up_epochs_10_lr_0.1_pruning_step_3_3:600 300 0.0 0.1 15.878 86.28
ofa_cifar10_width_multi_1.2_epochs_1000_data_split_2_warm_up_epochs_0_lr_0.1_pruning_step_3:600 300 0.0 0.1 13.883 86.25
ofa_cifar10_width_multi_1.2_epochs_1000_data_split_5_warm_up_epochs_0_lr_0.1_pruning_step_3:600 300 0.0 0.1 16.372 86.22
ofa_cifar10_width_multi_1.3_epochs_1000_data_split_10_warm_up_epochs_0_lr_0.1_pruning_step_3:600 300 0.0 0.1 17.203 86.37
ofa_cifar10_width_multi_1.4_epochs_1000_data_split_10_warm_up_epochs_0_lr_0.1_pruning_step_3:600 300 0.0 0.1 16.431 86.48
EfficientNet_b0 300 0.0 0.1 8.475 80.93
FBNet-C 300 0.0 0.1 7.836 79.3
proxyless__dataset_imagenet_width_multi_1.3_epochs_1000_data_split_10_warm_up_epochs_0_lr_0.01_pruning_step_3:600 300 0.0 0.1 12.129 84.48

Test

MIGONAS_400

Download the model at here and the password is ur0t

Test the model on ImageNet by

 python test.py --structure_path models/model_zoo/MIGONAS_400.json --pretrained path_to_the_model_file

MIGONAS_600

Download the model at here and the password is v8rh

Test the model on ImageNet by

 python test.py --structure_path models/model_zoo/MIGONAS_600.json --pretrained path_to_the_model_file

About

image classification by pytorch

Topics

pytorch image-classification

Resources

Readme
Activity

Stars

9 stars

Watchers

2 watching

Forks

5 forks
Report repository

Releases

No releases published

Packages

No packages published

Contributors 2

  •  
  •  

Languages