The code of the paper MDMLP.

Thanks to TIMM library!

How to use

It's used in the same way as timm.

We use fvcore to measure params and flops.

We changed timm/data/dataset_factory.py to be able to train on Flowers102 and Food101.

# For cifar10
python3 train.py /path-to-cifar10 -c ymls/cifar10_sgd.yml --model mdmlp_patch4_lap2_dim64_depth8_32
# For cifar100
python3 train.py /path-to-cifar100 -c ymls/cifar100_sgd.yml --model mdmlp_patch4_lap2_dim64_depth8_32

# For Flowers102
python3 train.py /path-to-flowers102 -c ymls/flowers102_sgd.yml --model mdmlp_patch14_lap7_dim64_depth8_224
# For Food101
python3 train.py /path-to-food101 -c ymls/food101_sgd.yml --model mdmlp_patch14_lap7_dim64_depth8_224

# For MDAttnTool
python3 train.py /path-to-cifar10 -c ymls/cifar10_sgd.yml --model mdmlp_attn_32