Skip to content

Latest commit

 

History

History
72 lines (58 loc) · 2.56 KB

lsq.md

File metadata and controls

72 lines (58 loc) · 2.56 KB
Raw

LSQ

cifar10 fp32 w4a4 w3a3 w2a2
vggsmall 93.8 93.7 92.5 86 
./run_cli.sh examples/classifier_cifar10/prototxt/vggsmall_lsq_w4a4_single_gpu.prototxt
./run_cli.sh examples/classifier_cifar10/prototxt/vggsmall_lsq_w3a3_single_gpu.prototxt
./run_cli.sh examples/classifier_cifar10/prototxt/vggsmall_lsq_w2a2_single_gpu.prototxt

The last layer is quantized.

LSQ fp32 w4a4 w3a3
AlexNet 56.55, 79.09 56.36 48.21

LSQ fp32 w4a4 w3a3 w2a2 w8a8(1epoch, quantize data)
AlexNet 56.55, 79.09 56.96, 79.46 √ 55.31, 78.59 51.18, 75.38
ResNet18 69.76, 89.08 70.26, 89.34 √ 69.45, 88.85 69.68 88.92 √

bash

AlexNet_LSQ_w4a4

python examples/classifier_imagenet/main.py ~/datasets/data.imagenet \
    -a alexnet_lsq -j 10 --pretrained -b 2048 --log-name $2 \
    --lr 0.01 --wd 1e-4 --warmup-epoch -1 \
    --gpu $1 --epochs 90 --lr-scheduler CosineAnnealingLR \
    --qw 4 --qa 4 --q-mode layer_wise \
    --debug

AlexNet_LSQ_w3a3

python examples/classifier_imagenet/main.py ~/datasets/data.imagenet \
    -a alexnet_lsq -j 10 --pretrained -b 2048 --log-name $2 \
    --lr 0.01 --wd 1e-4 --warmup-epoch -1 \
    --gpu $1 --epochs 90 --lr-scheduler CosineAnnealingLR \
    --qw 3 --qa 3 --q-mode layer_wise \
    --debug

AlexNet_LSQ_w2a2

python examples/classifier_imagenet/main.py ~/datasets/data.imagenet \
    -a alexnet_lsq -j 10 --pretrained -b 2048 --log-name $2 \
    --lr 0.01 --wd 1e-4 --warmup-epoch -1 \
    --gpu $1 --epochs 90 --lr-scheduler CosineAnnealingLR \
    --qw 3 --qa 3 --q-mode layer_wise \
    --debug

ResNet18_LSQ_w4a4

python examples/classifier_imagenet/main.py ~/datasets/data.imagenet \
    -a resnet18_lsq -j 10 -b 512 --pretrained \
    --lr 0.01 --wd 1e-4 --warmup-epoch -1 \
    --gpu $1 --log-name $2 --epochs 90 --lr-scheduler CosineAnnealingLR \
    --qw 4 --qa 4 --q-mode layer_wise \
    --debug