Skip to content

Latest commit

 

History

History
48 lines (37 loc) · 2.32 KB

TRAINING.md

File metadata and controls

48 lines (37 loc) · 2.32 KB
Raw

RDNet for Image Classification

backbone/rdnet.py for detection and segmentation is the exact same file as rdnet.py for classification.

Tested Environment

  • python 3.11.5
  • pytorch 1.13.1
  • timm 0.9.6
  • deepspeed 0.12.6 (For flops calculation)

Training Graphs

We provide the graphs of the training procedure. The graph is generated by the Weights & Biases service. You can check the graph by clicking the link below.

https://api.wandb.ai/links/dhkim0225/822w2zsj

Train

# RDNet-T
torchrun --nproc_per_node=4 train.py --pin-mem --channels-last --config train_configs/rdnet.yaml --drop-path 0.15 --model rdnet_tiny

# RDNet-S
torchrun --nproc_per_node=4 train.py --pin-mem --channels-last --config train_configs/rdnet.yaml --drop-path 0.35 --model rdnet_small

# RDNet-B
torchrun --nproc_per_node=4 train.py --pin-mem --channels-last --config train_configs/rdnet.yaml --drop-path 0.4 --model rdnet_base

# RDNet-L
torchrun --nproc_per_node=4 train.py --pin-mem --channels-last --config train_configs/rdnet.yaml --drop-path 0.5 --model rdnet_large

# RDNet-L 384 Finetuning
torchrun --nproc_per_node=4 train.py --pin-mem --channels-last --config train_configs/rdnet_finetune.yaml --model rdnet_large --initial-checkpoint [checkpoint path]

Validation

python validate.py --pin-mem --data-dir [Path to ilsvrc2012] --batch-size 64 --channels-last --amp --img-size 224 --model rdnet_tiny --pretrained
python validate.py --pin-mem --data-dir [Path to ilsvrc2012] --batch-size 64 --channels-last --amp --img-size 224 --model rdnet_small --pretrained
python validate.py --pin-mem --data-dir [Path to ilsvrc2012] --batch-size 64 --channels-last --amp --img-size 224 --model rdnet_base --pretrained
python validate.py --pin-mem --data-dir [Path to ilsvrc2012] --batch-size 64 --channels-last --amp --img-size 224 --model rdnet_large --pretrained

Benchmark

python benchmark.py -b 128 --num-bench-iter 1000 --img-size=224 --fuser te --channels-last --bench infer -m rdnet_tiny
python benchmark.py -b 128 --num-bench-iter 1000 --img-size=224 --fuser te --channels-last --bench infer -m rdnet_small
python benchmark.py -b 128 --num-bench-iter 1000 --img-size=224 --fuser te --channels-last --bench infer -m rdnet_base
python benchmark.py -b 128 --num-bench-iter 1000 --img-size=224 --fuser te --channels-last --bench infer -m rdnet_large