config modification for mar-base training on cifar-100

[cyboTiger]

Great work! I'm trying to conduct a verification experiment on the pretraining on cifar dataset (img_size 32x32) and evaluate the trained mar-base model.

So I'm wondering if you have any suggestions on modifications on eval config? Here's mine:

torchrun --nproc_per_node=8 --nnodes=1 --node_rank=${NODE_RANK} --master_addr=${MASTER_ADDR} --master_port=${MASTER_PORT} \
main_mar.py \
--img_size 32 --vae_path pretrained_models/vae/kl16.ckpt --vae_embed_dim 16 --vae_stride 16 --patch_size 1 \
--model ${MAR_SIZE} --diffloss_d 6 --diffloss_w 1024 \
--epochs 400 --warmup_epochs 100 --batch_size 64 --blr 1.0e-4 --diffusion_batch_mul 4 \
--output_dir ${OUTPUT_DIR} --wandb \
--data_path ${DATA_PATH} \
# below is eval-related configs
--online_eval --eval_bsz 256 \
--cfg 2.9 --cfg_schedule linear  --temperature 1.0 \
--num_iter 4 --num_sampling_steps 100 \
--eval_freq 20 --save_last_freq 5 \
--use_cached --cached_path ${CACHED_PATH} \
# --resume ${OUTPUT_DIR}

In particular, i think num_sampling_steps is most likely to be modified because cifar img size is only 32x32. Should I change it or stay the same?

Also, it seems that to evaluate FID and IS score on 32x32 generated images I'll need some fid_statistics_file other than fid_stats/adm_in256_stats.npz in this repo, where can I get it?

Thanks for your response in advance!