all_exp.sh

#!/bin/bash

# 定义要遍历的seed列表
SEEDS=(10 20 30 40 50)

# 定义数据集列表及其对应的GPU和端口偏移量
declare -A DATASET_CONFIG=(
    ["Rotten_Tomatoes"]="0 0"   # GPU 0, 端口偏移量 0
    ["imdb"]="1 10"             # GPU 1, 端口偏移量 10
    ["ag_news"]="2 20"          # GPU 2, 端口偏移量 20
    ["20_newsgroups"]="3 30"    # GPU 3, 端口偏移量 30
)

# 定义模型配置（只保留base模型），每个元素是模型偏移量、配置文件后缀
MODELS=(
    "0 transformer_base.yaml"   # 模型偏移量 0
    "1 conv_base.yaml"          # 模型偏移量 1
)

# 基础端口号
BASE_PORT=20700

# 遍历每个seed
for SEED in "${SEEDS[@]}"; do
    echo "开始运行seed: $SEED"
    
    # 存储当前seed批次的所有进程ID
    PIDS=()
    
    # 遍历每个数据集
    for DATASET in "${!DATASET_CONFIG[@]}"; do
        echo "  开始处理数据集: $DATASET"
        
        # 获取该数据集对应的GPU和端口偏移量
        read GPU DATASET_PORT_OFFSET <<< "${DATASET_CONFIG[$DATASET]}"
        
        # 遍历所有base模型配置
        for MODEL in "${MODELS[@]}"; do
            # 解析配置参数
            read MODEL_OFFSET CFG_SUFFIX <<< "$MODEL"
            
            # 根据数据集和模型配置构建完整路径和唯一端口
            CFG="cfgs/${DATASET}/${CFG_SUFFIX}"
            # 计算唯一端口：基础端口 + 数据集偏移 + 模型偏移 + seed
            PORT=$((BASE_PORT + DATASET_PORT_OFFSET + MODEL_OFFSET + SEED))
            
            # 构造标签
            TAG="seed_${SEED}"
            
            # 检查配置文件是否存在
            if [ ! -f "$CFG" ]; then
                echo "    警告: 配置文件 $CFG 不存在，跳过..."
                continue
            fi
            
            # 后台运行命令并记录PID
            nohup env CUDA_VISIBLE_DEVICES=$GPU torchrun --nproc_per_node=1 --master-port=$PORT main.py \
                --cfg "$CFG" --seed $SEED --tag "$TAG" > /dev/null 2>&1 &
            
            # 保存进程ID
            PID=$!
            PIDS+=($PID)
            
            echo "    启动: GPU=$GPU, PORT=$PORT, CFG=$CFG, TAG=$TAG, PID=$PID"
        done
    done
    
    # 等待当前seed的所有任务完成
    if [ ${#PIDS[@]} -gt 0 ]; then
        echo "  等待seed $SEED 的所有任务完成..."
        for PID in "${PIDS[@]}"; do
            wait $PID
            echo "  进程 $PID 已完成"
        done
    fi
    
    echo "seed $SEED 的所有任务执行完毕"
    echo "========================================"
done

echo "所有seed的任务都已完成！"




#!/bin/bash

# Define datasets
DATASETS=("CIFAR-10" "CIFAR-100" "SVHN" "FashionMNIST")

# Define models (一组：vit + conv vit)
MODELS=("vit_base.yaml" "conv_base.yaml")

# Configuration: 种子列表
SEEDS=(20 30 40)
BASE_PORT=29500

# Create logs directory
mkdir -p logs

echo "Starting image dataset experiments..."
echo "Datasets: ${DATASETS[*]}"
echo "Models: ${MODELS[*]} (一组并行运行)"
echo "Seeds: ${SEEDS[*]}"
echo "========================================"

# 外层循环：遍历每个种子
for SEED in "${SEEDS[@]}"; do
    echo "========================================"
    echo "Processing with SEED: $SEED"
    echo "========================================"
    
    for DATASET in "${DATASETS[@]}"; do
        echo "Processing dataset: $DATASET (Seed: $SEED)"
        
        # Ensure logs directory for dataset exists
        mkdir -p "logs/${DATASET}"
        
        PIDS=()  # 存储当前数据集下的模型PID
        
        # 内层循环：并行启动一组模型（vit + conv vit）
        for MODEL_CFG in "${MODELS[@]}"; do
            MODEL_NAME=$(basename "$MODEL_CFG" .yaml)
            CFG="cfgs/${DATASET}/${MODEL_CFG}"
            
            # Check if config exists
            if [ ! -f "$CFG" ]; then
                echo "  Warning: Config $CFG not found, skipping..."
                continue
            fi
            
            # Tag包含种子
            TAG="seed_${SEED}"
            
            # Unique port for each run
            PORT=$BASE_PORT
            BASE_PORT=$((BASE_PORT + 1))
            
            # 日志文件名包含种子
            LOG_FILE="logs/${DATASET}/${MODEL_NAME}_seed${SEED}.log"
            
            echo "  Starting $MODEL_NAME on $DATASET (Seed: $SEED, Port: $PORT)..."
            
            # 并行启动模型（后台运行）
            nohup torchrun --nproc_per_node=4 --master-port=$PORT main.py \
                --cfg "$CFG" --seed $SEED --tag "$TAG" > "$LOG_FILE" 2>&1 &
            
            # 记录PID
            PID=$!
            PIDS+=($PID)
            echo "  $MODEL_NAME started with PID: $PID"
        done
        
        # 等待当前数据集的一组模型全部完成
        if [ ${#PIDS[@]} -gt 0 ]; then
            echo "  Waiting for all models on $DATASET (Seed: $SEED) to complete..."
            for PID in "${PIDS[@]}"; do
                wait $PID
                if [ $? -eq 0 ]; then
                    echo "  PID $PID completed successfully"
                else
                    echo "  PID $PID exited with error"
                fi
            done
        fi
        
        echo "Completed all models for $DATASET (Seed: $SEED)"
        echo "--------------------------------"
    done
    
    echo "Completed all datasets for SEED: $SEED"
done

echo "All experiments completed!"