Delete viscop_qwen2.5_7b_viscop-lora_egocentric-expert/train_viscop.sh

viscop_qwen2.5_7b_viscop-lora_egocentric-expert/train_viscop.sh

@@ -1,124 +0,0 @@
-#!/bin/bash
-# Environment Variables
-ARG_WORLD_SIZE=${1:-1}
-ARG_NPROC_PER_NODE=${2:-8}
-
-if [[ -v MASTER_ADDR_PASSED ]]; then
-    ARG_MASTER_ADDR=$MASTER_ADDR_PASSED # passed via slurm submission script
-else
-    ARG_MASTER_ADDR=127.0.0.1 # for dev environments
-fi
-ARG_MASTER_PORT=12355
-# ARG_RANK=$SLURM_NODEID
-ARG_RANK=0
-
-# Multiple conditions
-if [ ! -n "$WORLD_SIZE" ] || [ ! -n "$NPROC_PER_NODE" ]; then
-    WORLD_SIZE=$ARG_WORLD_SIZE
-    NPROC_PER_NODE=$ARG_NPROC_PER_NODE
-fi
-
-if [ ! -n "$MASTER_ADDR" ] || [ ! -n "$MASTER_PORT" ] || [ ! -n "$RANK" ]; then
-    MASTER_ADDR=$ARG_MASTER_ADDR
-    MASTER_PORT=$ARG_MASTER_PORT
-    RANK=$ARG_RANK
-fi
-
-echo "MASTER_ADDR: $MASTER_ADDR. MASTER_PORT: $MASTER_PORT. RANK: $RANK"
-echo "WORLD_SIZE: $WORLD_SIZE"
-echo "NPROC_PER_NODE: $NPROC_PER_NODE"
-
-# Training Arguments
-GLOBAL_BATCH_SIZE=128 # aka effective batch size
-LOCAL_BATCH_SIZE=8 # batch size per GPU
-GRADIENT_ACCUMULATION_STEPS=$[$GLOBAL_BATCH_SIZE/($WORLD_SIZE*$NPROC_PER_NODE*$LOCAL_BATCH_SIZE)]
-echo $GRADIENT_ACCUMULATION_STEPS
-
-INIT_MODEL=/home/dreilly1/Projects/VisCoP/base_vlm/videollama3-image_7b_local # path to base VLM (for ViSCoP we use VideoLLaMA3 as the base VLM)
-
-NUM_DATA_WORKERS=8
-NUM_TRAIN_EPOCHS=3
-LORA_TRAINING=True
-
-# ViSCoP Arguments
-NUM_VISUAL_PROBES=16
-INTERACTION_MODULE_POS=all
-PASS_PROBES_TO_LLM=True
-PASS_VIS_FEATURES_TO_LLM=True
-
-# Logging Arguments
-export WANDB_PROJECT=sony26_mm_viscop
-REPORT_TO=wandb
-OUTP_DIR=work_dirs/egoexo
-RUN_NAME=viscop_qwen2.5_7b_EgoExo4D-EGOvideos_train-VisCoP_projector-LLM_LoRA_CorrectedOptimizer
-
-# Data Arguments
-# DATA_DIR=/home/dreilly1/Projects/paired_egoexo_dl/paired_videos_DATA/
-DATA_DIR=/data/dreilly1/EgoExo4D_symlink/
-TRAINING_JSON="/home/dreilly1/Projects/viscop_sony/training_jsons/train-instr_viscop_egoview_ALLDATA.json"
-
-# if [[ $TRAINING_JSON == *"egoview"* ]]; then
-#     MAX_FRAMES=40 # use 40 frames for training on ego
-# else
-#     MAX_FRAMES=180
-# fi
-MAX_FRAMES=180
-
-# Optional Arguments. Set TESTING to 1 to quickly test the training script without logging or data workers, useful for debugging
-TESTING=0
-if [ $TESTING -eq 1 ]; then
-    NUM_DATA_WORKERS=0
-    REPORT_TO=none
-    RUN_NAME=TESTING
-fi
-
-mkdir -p "${OUTP_DIR}/${RUN_NAME}/"
-cp "$0" "${OUTP_DIR}/${RUN_NAME}/"
-
-torchrun --nnodes $WORLD_SIZE \
-    --nproc_per_node $NPROC_PER_NODE \
-    --master_addr=$MASTER_ADDR \
-    --master_port=$MASTER_PORT \
-    --node_rank $RANK \
-    viscop/train_viscop.py \
-    --interaction_module_layers $INTERACTION_MODULE_POS \
-    --lora_enable $LORA_TRAINING \
-    --num_train_epochs $NUM_TRAIN_EPOCHS \
-    --deepspeed scripts/zero2.json \
-    --model_type viscop_qwen2 \
-    --model_path $INIT_MODEL \
-    --vision_encoder DAMO-NLP-SG/SigLIP-NaViT \
-    --mm_projector_type mlp2x_gelu \
-    --data_path $TRAINING_JSON \
-    --data_folder $DATA_DIR \
-    --image_merge_size 2 \
-    --video_merge_size 2 \
-    --fps 1 \
-    --max_frames $MAX_FRAMES \
-    --model_max_length 16384 \
-    --mm_max_length 10240 \
-    --bf16 True \
-    --tf32 True \
-    --fp16 False \
-    --output_dir ${OUTP_DIR}/${RUN_NAME} \
-    --per_device_train_batch_size $LOCAL_BATCH_SIZE \
-    --per_device_eval_batch_size 2 \
-    --gradient_accumulation_steps $GRADIENT_ACCUMULATION_STEPS \
-    --evaluation_strategy "no" \
-    --save_strategy "no" \
-    --save_steps 5000 \
-    --save_total_limit 1 \
-    --mm_projector_lr 1e-5 \
-    --llm_lr 1e-5 \
-    --weight_decay 0. \
-    --warmup_ratio 0.03 \
-    --lr_scheduler_type "cosine" \
-    --logging_steps 1 \
-    --gradient_checkpointing True \
-    --dataloader_num_workers $NUM_DATA_WORKERS \
-    --report_to $REPORT_TO \
-    --run_name $RUN_NAME \
-    --dataset_cache_dir /home/dreilly1/.cache/viscop_datasetcache \
-    --include_visual_tokens $PASS_VIS_FEATURES_TO_LLM \
-    --include_visual_probes $PASS_PROBES_TO_LLM \
-    --num_visual_probes $NUM_VISUAL_PROBES