Skip to content

Latest commit

 

History

History
153 lines (148 loc) · 5.66 KB

DP.md

File metadata and controls

153 lines (148 loc) · 5.66 KB
Raw

Constructing look-up tables & Solving DP algorithm

We provide the instruction that can construct the look-up tables for latency and importance here. The latency measurements are obtained using a single RTX2080 Ti GPU, while the importance are measured using two RTX3090 GPUs.

ResNet-34

  1. Move to HALP directory. 
    cd HALP
  2. Construct the latency look-up table (we used single RTX2080 Ti GPU in the paper). 
    lymg_kim24_dp -a resnet34 --mode time \
-d LUT_kim24/time/rn34/ \
-t fish_1206_batch128
  3. Measure the importance look-up table entries (we used two RTX3090 GPU per entry). This process can be parallelized across multiple GPUs by setting --from_blk and --to_blk for each device. 
    lymg_kim24_imp -a resnet34 \
--pretrained pretrained/resnet34_full.pth \
--save_path LUT_kim24/imp/rn34/par \
--data /ssd_data/imagenet/ --from_blk 0 --to_blk 150
  4. Aggregate importance table entries into a single file. 
    lymg_agg -d LUT_kim24/imp/rn34 -n 150
  5. Normalize the importance value. 
    lymg_kim24_dp -a resnet34 --mode normalize \
--imp_path LUT_kim24/imp/rn34/importance.csv
  6. To solve DP problem with $T_0$ time budget, run 
    lymg_kim24_dp -a resnet34 --mode solve \
-d LUT_kim24/solve/rtx2080/rn34/ \
--time_path LUT_kim24/time/rn34/time_fish_1206_batch128.csv \
--imp_path LUT_kim24/imp/rn34/normalized_importance.csv \
--time_limit {T_0}
    • To reproduce the numbers in Table 1, try {T_0} value from {146.0, 123.0, 103.0} and proceed to next step.

MobileNetV2-(1.0/1.4)

  1. Move to Efficient-CNN-Depth-Compression directory. 
    cd Efficient-CNN-Depth-Compression
  2. Construct the latency look-up table (we used single RTX2080 Ti GPU in the paper). 
    # MobileNetV2-1.0
lymg_kim24_dp -a mobilenetv2 --mode time \
-d LUT_kim24/time/mbv2/ \
-t fish_1211_batch128
    # MobileNetV2-1.4
lymg_kim24_dp -a mobilenetv2_w1.4 --mode time \
-d LUT_kim24/time/mbv2_w1.4/ \
-t fish_0112_batch128
  3. Measure the importance look-up table entries (we used two RTX3090 GPU per entry). This process can be parallelized across multiple GPUs by setting --from_blk and --to_blk for each device. 
    # MobileNetV2-1.0
lymg_kim24_imp -a mobilenetv2 \
--pretrained pretrained/mobilenetv2_100_ra-b33bc2c4.pth \
--save_path LUT_kim24/imp/mbv2/par \
--data /ssd_data/imagenet/ --from_blk 0 --to_blk 391
    # MobileNetV2-1.4
lymg_kim24_imp -a mobilenetv2_w1.4 \
--pretrained pretrained/mobilenetv2_140_ra-21a4e913.pth \
--save_path LUT_kim24/imp/mbv2_w1.4/par \
--data /ssd_data/imagenet/ --from_blk 0 --to_blk 391
  4. Aggregate importance table entries into a single file. 
    # MobileNetV2-1.0
lymg_agg -d LUT_kim24/imp/mbv2 -n 391
    # MobileNetV2-1.4
lymg_agg -d LUT_kim24/imp/mbv2_w1.4 -n 391
  5. Normalize the importance value. 
    # MobileNetV2-1.0
lymg_kim24_dp -a mobilenetv2 --mode normalize \
--imp_path LUT_kim24/imp/mbv2/importance.csv
    # MobileNetV2-1.4
lymg_kim24_dp -a mobilenetv2_w1.4 --mode normalize \
--imp_path LUT_kim24/imp/mbv2_w1.4/importance.csv
  6. To solve DP problem with $T_0$ time budget, run 
    # MobileNetV2-1.0
lymg_kim24_dp -a mobilenetv2 --mode solve \
-d LUT_kim24/solve/rtx2080/mbv2/ \
--time_path LUT_kim24/time/mbv2/time_fish_1211_batch128.csv \
--imp_path LUT_kim24/imp/mbv2/normalized_importance.csv \
--time_limit {T_0}
    • To reproduce the numbers in Table 2, try {T_0} value from {22.3, 18.5, 15.6, 13.4} and proceed to next step.
    # MobileNetV2-1.4
lymg_kim24_dp -a mobilenetv2_w1.4 --mode solve \
-d LUT_kim24/solve/rtx2080/mbv2_w1.4/ \
--time_path LUT_kim24/time/mbv2_w1.4/time_fish_0112_batch128.csv \
--imp_path LUT_kim24/imp/mbv2_w1.4/normalized_importance.csv \
--time_limit {T_0}
    • To reproduce the numbers in Table 3, try {T_0} value from {26.1, 25.0, 21.0, 18.0} and proceed to next step.

DDPM

  1. Move to Diff-Pruning/exp_code directory. 
    cd Diff-Pruning/exp_code
  2. Construct the latency look-up table (we used single RTX2080 Ti GPU in the paper). 
    lymg_kim24_dp -a ddpm_cifar10 --mode time
-d LUT_kim24/time/ddpm_cifar10 \
-t fish_0116_batch128 \
  3. Measure the importance look-up table entries (we used two RTX3090 GPU per entry). This process can be parallelized across multiple GPUs by setting --from_blk and --to_blk for each device. 
    slmg_kim24_imp -a ddpm_cifar10 \
--imp_epoch 50 \
--pretrained run/cache/diffusion_models_converted/ema_diffusion_cifar10_model/model-790000.ckpt \
--save_path LUT_kim24/imp/ddpm_cifar10/par \
--data /data_large/readonly/ --from_blk 0 --to_blk 98
  4. Aggregate importance table entries into a single file. 
    lymg_agg -d LUT_kim24/imp/ddpm_cifar10 -n 98
  5. Normalize the importance value. 
    lymg_kim24_dp -a ddpm_cifar10 --mode normalize \
--imp_path LUT_kim24/imp/ddpm_cifar10/importance.csv
  6. To solve DP problem with $T_0$ time budget, run 
    lymg_kim24_dp -a ddpm_cifar10 --mode solve \
-d LUT_kim24/solve/rtx2080/ddpm_cifar10/ \
--time_path LUT_kim24/time/ddpm_cifar10/time_fish_0116_batch128.csv \
--imp_path LUT_kim24/imp/ddpm_cifar10/normalized_importance.csv \
--time_limit {T_0}
    • To reproduce the numbers in Table 4, try {T_0} value from {48.0, 46.0, 38.0} and proceed to next step.