File size: 3,432 Bytes
33569f9 | 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 | #!/bin/bash
# ActivityForensics GRPO — v10 full bundle.
#
# v9_full eval (cot=false, max_new_tokens=64, 1565 test videos):
# stage1_decomp mIoU=43.28 F1@0.5=45.94 F1@0.7=27.07
# v8_ckpt200 mIoU=41.66 F1@0.5=44.48 F1@0.7=25.18
# v9_ckpt240 mIoU=45.99 F1@0.5=51.02 F1@0.7=29.75 (epoch 1 end)
# v9_ckpt478 mIoU=49.16 F1@0.5=56.63 F1@0.7=35.29 (epoch 2 end, +5.9pt vs stage1)
#
# v9 beat stage1 by ~6pt and was still climbing (240→478: +3.17 mIoU). Curve
# did not saturate, so v10 doubles training to 4 epochs.
#
# Augmentation bump: SPI 0.2 → 0.25, FBR 0.1 → 0.15 (P[any aug] 28% → 36%).
# Motivation: with 2x training steps, forward-distribution overfit risk
# rises; a modest aug bump regularises without reversing fix (1) — 36% is
# still well below v8's 62.5% that caused the original regression.
# Expected capacity cost vs v9: <0.5pt.
#
# Other config (reward stack, taus, gradient accumulation, decomp/boundary)
# unchanged from v9. See v9 script comment block for the original three-fix
# root-cause analysis (capacity / strict_boundary bipolar reward / batch).
#
export WANDB_PROJECT=Forensics-GRPO
export WANDB_NAME="${WANDB_NAME:-AF_v10_full}"
export PYTHONPATH=".:$PYTHONPATH"
export DEBUG_MODE="true"
export LOG_PATH="./qwen2.5_7b_vl_forensics_v10_full.txt"
export PATH="/mnt/local-fast/zhangt/torch_env/bin:$PATH"
export LD_LIBRARY_PATH="/opt/conda/lib:${LD_LIBRARY_PATH}"
export FORENSICS_COT=false
export FORENSICS_FORGERY_AWARE=false
export FORENSICS_HUNGARIAN_DECOMP=true
export FORENSICS_HUNGARIAN_MAXK=8
export FORENSICS_BOUNDARY_WEIGHT=0.3
export FORENSICS_BOUNDARY_TAU=2.0
# === v10 changes vs v9 ===
# Augmentation bumped: P[any aug] 28% → 36%.
export FORENSICS_SPI_AUG=true
export FORENSICS_SPI_PROB=0.25
export FORENSICS_SPI_CHUNK_S=2.5
export FORENSICS_SPI_SAFETY_S=0.5
export FORENSICS_FBR_AUG=true
export FORENSICS_FBR_PROB=0.15
# Same wider thresholds with built-in curriculum (kept from v9).
export FORENSICS_STRICT_TAUS="0.3,0.5,0.7"
unset FORENSICS_COT_CONSIS_WEIGHT
unset FORENSICS_COT_CF_WEIGHT
unset FORENSICS_COT_ALIGN_WEIGHT
unset FORENSICS_K_REQUIRE_TIMESTEP
unset FORENSICS_IOU_REQUIRE_THINK
OUTDIR="${OUTDIR:-outputs_forensics/v10_full}"
MODEL_PATH="${MODEL_PATH:-/mnt/local-fast/zhangt/Qwen2.5-VL-7B-Instruct}"
PREPROCESSED="${PREPROCESSED:-/mnt/local-fast/zhangt/forensics_grpo_cache_uniform3584_fps2.0}"
cd "$(dirname "$0")/.."
torchrun --nproc_per_node="8" \
--nnodes="1" \
--node_rank="0" \
--master_addr="127.0.0.1" \
--master_port="12370" \
src/open_r1/grpo_forensics.py \
--deepspeed scripts/zero3_offload.json \
--output_dir "$OUTDIR" \
--model_name_or_path "$MODEL_PATH" \
--annot_dir /mnt/local-fast/zhangt/annot/annot \
--video_root /mnt/local-fast/zhangt/video \
--preprocessed_data_path "$PREPROCESSED" \
--reward_funcs iou strict_boundary \
--dataset_name activityforensics \
--max_prompt_length 8192 \
--max_completion_length 512 \
--num_generations 4 \
--per_device_train_batch_size 1 \
--gradient_accumulation_steps 2 \
--logging_steps 1 \
--bf16 \
--torch_dtype bfloat16 \
--data_seed 42 \
--gradient_checkpointing true \
--attn_implementation flash_attention_2 \
--num_train_epochs 4 \
--run_name "$WANDB_NAME" \
--report_to none \
--save_steps 30 \
--learning_rate 1e-6 \
--save_only_model true
|