hyperparameters.md

Please also check our extensive experiments page for better results and other network architectures.

Some arguments explained:

• --data-path should point to a directory where cifar10 is downloaded to (the first time), and later loaded from. If you get a corrupt ZIP-File Error you might have cancelled a previous download, simply delete the ZIP and restart. (only needed for CIFAR)
• --data-dir should point to a directory with ImageNet record files, i.e. imagenet-val.rec and imagenet-train.rec (only needed for ImageNet); use the official guide or the guide from Gluon-CV to create these files
• --resume and --start-epoch are to continue training from a previous checkpoint, i.e. --resume checkpoint_epoch_10.params --start-epoch 10 lets you restart training at epoch 10 from that checkpoint. Note that both arguments are independent, and epoch must be set correctly manually.

This should train a binary model with ~90% Accuracy on CIFAR10:

python image_classification.py \
    --data-path "/path/to/cifar/" \
    --augmentation low \
    --batch-size 128 \
    --clip-threshold 1.3 \
    --dataset cifar10 \
    --epochs 150 \
    --gpus 0 \
    --log-interval 50 \
    --lr 0.02 \
    --lr-mode cosine \
    --mode hybrid \
    --model resnet18_v1 \
    --optimizer adam \
    --warmup-epochs 5

This should train a binary model with ~59% Accuracy on ImageNet:

python image_classification.py \
    --data-dir ~/.mxnet/datasets/imagenet/ \
    --augmentation medium \
    --batch-size 32 \
    --clip-threshold 1.3 \
    --dataset imagenet \
    --epochs 120 \
    --fp-downsample-sc \
    --gpus 0,1 \
    --lr 0.002 \
    --lr-mode cosine \
    --mode hybrid \
    --model resnet18_e1 \
    --optimizer adam \
    --pool-downsample-sc \
    --warmup-epochs 5