Awesome

Mega for Image Classification

Implementation of Mega on Image Classification. This folder is based on the repos of Pytorch Image Models and DeiT

Requirements

Python 3.7, Pytorch 1.11.0 Cuda11.3

Installation

pip install --editable pytorch-image-models/  --user

Model Checkpoints

Training Mega on ImageNet-1k

To train Mega-base on ImageNet on a single node with 8 gpus for 300 epochs with slurm:

srun --label python -u deit/main.py \
        --model mega_base_patch16_224 \
        --batch-size 128 \
        --lr 2e-3 \
        --seed 42 \
        --norm-type 'layernorm' \
        --drop-path 0.3 \
        --epochs 300 \
        --warmup-epochs 20 \
        --weight-decay 0.05 \
        --warmup-lr 1e-5 \
        --clip-grad 1.0 \
        --opt-betas 0.9 0.98 \
        --world_size 8 \
        --reprob 0.25 \
        --repeated-aug 3 \
        --data-path ${DATA_PATH} \
        --output_dir ${MODEL_PATH}