Skip to content

Latest commit

 

History

History
141 lines (126 loc) · 3.96 KB

NAT.md

File metadata and controls

141 lines (126 loc) · 3.96 KB
Raw

NAT - Semantic Segmentation

Training on ADE20K

NAT-Mini + UPerNet 
./dist_train.sh configs/nat/upernet_nat_mini_512x512_160k_ade20k.py $NUM_GPUS --cfg-options data.samples_per_gpu=$((16/$NUM_GPUS)) data.workers_per_gpu=$((16/$NUM_GPUS))
NAT-Tiny + UPerNet 
./dist_train.sh configs/nat/upernet_nat_tiny_512x512_160k_ade20k.py $NUM_GPUS --cfg-options data.samples_per_gpu=$((16/$NUM_GPUS)) data.workers_per_gpu=$((16/$NUM_GPUS))
NAT-Small + UPerNet 
./dist_train.sh configs/nat/upernet_nat_small_512x512_160k_ade20k.py $NUM_GPUS --cfg-options data.samples_per_gpu=$((16/$NUM_GPUS)) data.workers_per_gpu=$((16/$NUM_GPUS))
NAT-Base + UPerNet 
./dist_train.sh configs/nat/upernet_nat_base_512x512_160k_ade20k.py $NUM_GPUS --cfg-options data.samples_per_gpu=$((16/$NUM_GPUS)) data.workers_per_gpu=$((16/$NUM_GPUS))

Validation

NAT-Mini + UPerNet

Single scale:

./dist_test.sh \
    configs/nat/upernet_nat_mini_512x512_160k_ade20k.py \
    https://shi-labs.com/projects/nat/checkpoints/SEG/nat_mini_upernet.pth \
    $NUM_GPUS \
    --eval mIoU

Multi scale:

./dist_test.sh \
    configs/nat/upernet_nat_mini_512x512_160k_ade20k.py \
    https://shi-labs.com/projects/nat/checkpoints/SEG/nat_mini_upernet.pth \
    $NUM_GPUS \
    --eval mIoU --aug-test
NAT-Tiny + UPerNet

Single scale:

./dist_test.sh \
    configs/nat/upernet_nat_tiny_512x512_160k_ade20k.py \
    https://shi-labs.com/projects/nat/checkpoints/SEG/nat_tiny_upernet.pth \
    $NUM_GPUS \
    --eval mIoU

Multi scale:

./dist_test.sh \
    configs/nat/upernet_nat_tiny_512x512_160k_ade20k.py \
    https://shi-labs.com/projects/nat/checkpoints/SEG/nat_tiny_upernet.pth \
    $NUM_GPUS \
    --eval mIoU --aug-test
NAT-Small + UPerNet

Single scale:

./dist_test.sh \
    configs/nat/upernet_nat_small_512x512_160k_ade20k.py \
    https://shi-labs.com/projects/nat/checkpoints/SEG/nat_small_upernet.pth \
    $NUM_GPUS \
    --eval mIoU

Multi scale:

./dist_test.sh \
    configs/nat/upernet_nat_small_512x512_160k_ade20k.py \
    https://shi-labs.com/projects/nat/checkpoints/SEG/nat_small_upernet.pth \
    $NUM_GPUS \
    --eval mIoU --aug-test
NAT-Base + UPerNet

Single scale:

./dist_test.sh \
    configs/nat/upernet_nat_base_512x512_160k_ade20k.py \
    https://shi-labs.com/projects/nat/checkpoints/SEG/nat_base_upernet.pth \
    $NUM_GPUS \
    --eval mIoU

Multi scale:

./dist_test.sh \
    configs/nat/upernet_nat_base_512x512_160k_ade20k.py \
    https://shi-labs.com/projects/nat/checkpoints/SEG/nat_base_upernet.pth \
    $NUM_GPUS \
    --eval mIoU --aug-test

Checkpoints

Backbone Network # of Params FLOPs mIoU mIoU (multi-scale) Checkpoint Config
NAT-Mini UPerNet 50M 900G 45.1 46.4 Download config.py
NAT-Tiny UPerNet 58M 934G 47.1 48.4 Download config.py
NAT-Small UPerNet 82M 1010G 48.0 49.5 Download config.py
NAT-Base UPerNet 123M 1137G 48.5 49.7 Download config.py