Skip to content

Latest commit

 

History

History
94 lines (71 loc) · 3.05 KB

training.md

File metadata and controls

94 lines (71 loc) · 3.05 KB

ImageNet training

Requirement: torch, torchvision, numpy. As long as the version is not too old, it should be fine.

All possible model names are BN models: resnet50, resnet101, resnet152, resnext50_32x4d, resnext101_32x8d, non-normalization models: rescale50, rescale101, rescale152, rescaleX50_32x4d, rescaleX101_32x8d, fixup50, fixup101.

ResNet50

python imagenet.py --model_name=resnet50 --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --drop_conv=0.03 --drop_fc=0.3 --alpha=0.0 --multi_step=[30,60,90]

Rescale50

python imagenet.py --model_name=rescale50 --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --drop_conv=0.03 --drop_fc=0.3 --alpha=0.0 --multi_step=[30,60,90]

EUNNet50

python imagenet.py --model_name=eunnet50 --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --drop_conv=0.03 --drop_fc=0.3 --alpha=0.0 --multi_step=[30,60,90]

Using Mixup and no Dropout

python imagenet.py --model_name=eunnet50 --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --drop_conv=0.0 --drop_fc=0.0 --alpha=0.5 --multi_step=[30,60,90]

python imagenet.py --model_name=eunnet50 --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --drop_conv=0.0 --drop_fc=0.0 --alpha=0.7 --multi_step=[30,60,90]

No regularization

python imagenet.py --model_name=eunnet50 --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --drop_conv=0.0 --drop_fc=0.0 --alpha=0.0 --multi_step=[30,60,90]

Cosine Learning rate

python imagenet.py --model_name=eunnet50 --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --drop_conv=0.03 --drop_fc=0.3 --alpha=0.0 --multi_step=[]

rescale101

python imagenet.py --model_name=eunnet101 --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --drop_conv=0.03 --drop_fc=0.3 --alpha=0.0 --multi_step=[30,60,90]

rescaleX101_32x8d

python imagenet.py --model_name=rescaleX101_32x8d --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --drop_conv=0.03 --drop_fc=0.3 --alpha=0.0 --multi_step=[30,60,90]

rescaleX101_32x8d + cosline

python imagenet.py --model_name=rescaleX101_32x8d --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --drop_conv=0.03 --drop_fc=0.3 --alpha=0.0 --multi_step=[]

VGG19

python vgg_imgaenet.py --model_name=vgg19_noBN --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --multi_step=[60, 90]

VGG19 + cosine

python vgg_imgaenet.py --model_name=vgg19_noBN --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --multi_step=[] --bs256_lr=0.01

VGG19_BN

python vgg_imgaenet.py --model_name=vgg19_bn --train_path=TRAIN_PATH --val_path=VAL_PATH --batch_size=1024 \
                   --multi_step=[30, 60, 90]