File size: 5,915 Bytes
fbc94ef | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 | #!/usr/bin/env bash
set -euo pipefail
# One-GPU ablation queue (default GPU1).
# This script runs experiments sequentially and stores all artifacts under:
# results/ablation_study/
ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
cd "$ROOT_DIR"
GPU_ID="${GPU_ID:-1}"
EPOCHS_ABL="${EPOCHS_ABL:-5}"
MAX_SAMPLES="${MAX_SAMPLES:-10000}"
TOPK="${TOPK:-30}"
ABL_ROOT="results/ablation_study"
CKPT_DIR="$ABL_ROOT/checkpoints"
RUN_DIR="$ABL_ROOT/runs"
LOG_DIR="$ABL_ROOT/logs"
mkdir -p "$CKPT_DIR" "$RUN_DIR" "$LOG_DIR"
run() {
local name="$1"
shift
echo
echo "============================================================"
echo "[$(date '+%F %T')] START: $name"
echo "CMD: $*"
echo "============================================================"
"$@"
echo "[$(date '+%F %T')] DONE: $name"
}
export PYTHONPATH=.
# 0) Figure-2 continuation: resume mix_d2c to epoch20 (if possible)
if [[ -f checkpoints/optical_mix_d2c/latest.pt ]]; then
run "figure2_resume_to_e20" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/train_compressor.py \
--output_dir checkpoints/optical_mix_d2c \
--resume checkpoints/optical_mix_d2c/latest.pt \
--epochs 20 \
--max_samples "$MAX_SAMPLES" \
--mix_root data \
--mix_images_subdir ref_screenshots \
--mix_gt_subdir gt_html \
--max_html_tokens 8192 \
--eval_after_epoch \
--eval_output_dir results/clip_per_epoch/optical_mix_d2c \
--eval_clip_device cuda
fi
# 1) Remove LoRA ablation
run "train_no_lora_256" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/train_compressor.py \
--output_dir "$CKPT_DIR/no_lora_256" \
--disable_lora \
--target_tokens 256 \
--epochs "$EPOCHS_ABL" \
--max_samples "$MAX_SAMPLES" \
--mix_root data \
--mix_images_subdir ref_screenshots \
--mix_gt_subdir gt_html \
--max_html_tokens 8192
run "eval_no_lora_256" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
--method uipress \
--checkpoint "$CKPT_DIR/no_lora_256/latest.pt" \
--target_tokens 256 \
--max_samples 50 \
--data_dir data \
--output_dir "$RUN_DIR/no_lora_256"
run "clip_no_lora_256" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
--method_dir "$RUN_DIR/no_lora_256/uipress_256" \
--ref_dir data/ref_screenshots \
--clip_device cuda
# 2) Token sensitivity ablation
for tok in 64 128 512; do
run "train_token_${tok}" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/train_compressor.py \
--output_dir "$CKPT_DIR/token_${tok}" \
--target_tokens "$tok" \
--epochs "$EPOCHS_ABL" \
--max_samples "$MAX_SAMPLES" \
--mix_root data \
--mix_images_subdir ref_screenshots \
--mix_gt_subdir gt_html \
--max_html_tokens 8192
run "eval_token_${tok}" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
--method uipress \
--checkpoint "$CKPT_DIR/token_${tok}/latest.pt" \
--target_tokens "$tok" \
--max_samples 50 \
--data_dir data \
--output_dir "$RUN_DIR/token_${tok}"
run "clip_token_${tok}" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
--method_dir "$RUN_DIR/token_${tok}/uipress_${tok}" \
--ref_dir data/ref_screenshots \
--clip_device cuda
done
# 3) Learning-rate scan (compressor lr)
for lr in 1e-4 2e-4 4e-4; do
safe_lr="${lr//./p}"
run "train_lr_${safe_lr}" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/train_compressor.py \
--output_dir "$CKPT_DIR/lr_${safe_lr}" \
--target_tokens 256 \
--lr_compressor "$lr" \
--epochs "$EPOCHS_ABL" \
--max_samples "$MAX_SAMPLES" \
--mix_root data \
--mix_images_subdir ref_screenshots \
--mix_gt_subdir gt_html \
--max_html_tokens 8192
run "eval_lr_${safe_lr}" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
--method uipress \
--checkpoint "$CKPT_DIR/lr_${safe_lr}/latest.pt" \
--target_tokens 256 \
--max_samples 50 \
--data_dir data \
--output_dir "$RUN_DIR/lr_${safe_lr}"
run "clip_lr_${safe_lr}" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
--method_dir "$RUN_DIR/lr_${safe_lr}/uipress_256" \
--ref_dir data/ref_screenshots \
--clip_device cuda
done
# 4) Cross-domain validation (WebSight screenshots as eval set)
TMP_WEBSIGHT_DIR="$ABL_ROOT/tmp_websight_eval"
mkdir -p "$TMP_WEBSIGHT_DIR"
if [[ ! -e "$TMP_WEBSIGHT_DIR/ref_screenshots" ]]; then
ln -s "$(realpath data/ref_screenshots_websight)" "$TMP_WEBSIGHT_DIR/ref_screenshots"
fi
run "cross_domain_qwen3_full" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
--method baseline \
--max_samples 50 \
--data_dir "$TMP_WEBSIGHT_DIR" \
--output_dir "$RUN_DIR/cross_domain"
run "cross_domain_uipress_latest" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/eval_all.py \
--method uipress \
--checkpoint checkpoints/optical_mix_d2c/latest.pt \
--target_tokens 256 \
--max_samples 50 \
--data_dir "$TMP_WEBSIGHT_DIR" \
--output_dir "$RUN_DIR/cross_domain"
run "cross_domain_clip_qwen3_full" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
--method_dir "$RUN_DIR/cross_domain/qwen3_full" \
--ref_dir data/ref_screenshots_websight \
--clip_device cuda
run "cross_domain_clip_uipress_256" \
env CUDA_VISIBLE_DEVICES="$GPU_ID" python scripts/step_clip_batch.py \
--method_dir "$RUN_DIR/cross_domain/uipress_256" \
--ref_dir data/ref_screenshots_websight \
--clip_device cuda
# 5) Build Top-K report from current available methods
run "build_topk_report" \
python scripts/ablation_topk_report.py \
--topk "$TOPK" \
--out_root "$ABL_ROOT"
echo
echo "All queue steps completed at $(date '+%F %T')."
|