File size: 20,512 Bytes
8cd1f2e |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 |
#!/bin/bash -ex
source ./scripts/env.sh
# 1: model
# 2: dataset
# 3: size
# 4: loss
# 5: num_seed
# 6: val_check_interval
# 7: train_dir
# 8: early_stopping
# 9: train_list
# 10: val_list
# 11: pred_list
run_lora() {
local model=$1
local dataset=$2
local size=$3
local loss=$4
local num_seed=$5
local seed=$((base_seed + num_seed))
local val_check_interval=$6
local train_dir=$7
local early_stopping=$8
local train_list=$9
local val_list=${10}
local pred_list=${11}
local lora_args="--lora_r=8 --lora_alpha=16 --lora_dropout=0.05 --lora_query --lora_key --lora_value --lora_projection --lora_mlp --lora_head"
local global_batch_size=8
local micro_batch_size=1
local learning_rate=0.0001
local optimizer="adamw"
local weight_decay=0.0
local patience=10
local precision="bf16-true"
if [ $early_stopping = true ] && [ $train_list = "0.0-1.0" ]; then
local max_steps=$(python -c "import torch; print(torch.load('$train_dir/../../0.0-0.7/0.7-1.0/best.ckpt', weights_only=False)['step_count'],end='')")
else
local max_steps=-1
fi
# TRAIN
local model_dir="$CHECKPOINTS_DIR/${model2checkpoint[$model]}"
local log_dir="$train_dir/logs"
local output_checkpoint_dir="$train_dir/checkpoint"
if [ ! -f $train_dir/train_args.yaml ]; then
mkdir -p $train_dir $log_dir $output_checkpoint_dir
for file in config.json generation_config.json model_config.yaml tokenizer.json tokenizer.model tokenizer_config.json; do
if [ -f $model_dir/$file ]; then
cp $model_dir/$file $output_checkpoint_dir
fi
done
ln -sf $(readlink -f $model_dir/lit_model.pth) $output_checkpoint_dir/lit_model.pth
python -m llmcal.scripts.train_lora \
--base_checkpoint_dir $model_dir \
--data_paths outputs/prompts/$model/$dataset/all.jsonl \
--train_lists lists/$dataset/size=$size/seed=$num_seed/$train_list.txt \
--val_lists lists/$dataset/size=$size/seed=$num_seed/$val_list.txt \
--output_dir $train_dir \
--output_checkpoint_dir $output_checkpoint_dir \
--log_dir $log_dir \
--precision $precision \
--devices 1 \
--num_nodes 1 \
--global_batch_size $global_batch_size \
--micro_batch_size $micro_batch_size \
--val_check_interval $val_check_interval \
--learning_rate $learning_rate \
--optimizer $optimizer \
--weight_decay $weight_decay \
--loss $loss \
--patience $patience \
--max_steps $max_steps \
--seed $seed \
$lora_args
fi
# PREDICT
local output_dir="$train_dir/test=$dataset/list=$pred_list"
if [ ! -f $output_dir/logits.csv ]; then
mkdir -p $output_dir
python -m llmcal.scripts.run_posteriors \
--base_checkpoint_dir $model_dir \
--checkpoint_dir $output_checkpoint_dir \
--peft "lora" \
--data_path outputs/prompts/$model/$dataset/all.jsonl \
--output_dir $output_dir \
--prediction_lists lists/$dataset/size=$size/seed=$num_seed/$pred_list.txt \
--precision $precision \
--devices 1 \
--num_nodes 1 \
--batch_size 1 \
--max_seq_length $max_seq_length \
$lora_args
fi
}
train_and_run_calibration() {
local method=$1
local num_seed=$2
local train_dir=$3
local pred_dir=$4
local output_dir=$5
local seed=$((base_seed + num_seed))
if [ ! -f $output_dir/logits.csv ]; then
mkdir -p $output_dir
python -m llmcal.scripts.affine_calibration \
--output_dir $output_dir \
--log_dir "$output_dir/../../logs" \
--checkpoint_dir "$output_dir/../../" \
--train_logits "$train_dir/logits.csv" \
--train_labels "$train_dir/labels.csv" \
--predict_logits "$pred_dir/logits.csv" \
--predict_labels "$pred_dir/labels.csv" \
--method $method \
--learning_rate 1e-3 \
--tolerance 1e-5 \
--max_ls 40
fi
}
run_calibration() {
local method=$1
local checkpoint_path=$2
local predict_dir=$3
local output_dir=$4
if [ ! -f $output_dir/logits.csv ]; then
mkdir -p $output_dir
python -m llmcal.scripts.affine_prediction \
--checkpoint_path $checkpoint_path \
--method $method \
--predict_logits $predict_dir/logits.csv \
--predict_labels $predict_dir/labels.csv \
--output_dir $output_dir
fi
}
# 1: model
# 2: sizes
# 3: val_check_interval
run_lora_vs_samples() {
local model=$1
local val_check_interval=$2
for size in ${FACTORS[@]}; do
for dataset in "${DATASETS[@]}"; do
local test_list="test_${dataset2testsize[$dataset]}"
local num_seeds=${dataset2nseeds[$dataset]}
for num_seed in $(seq 0 $(($num_seeds - 1))); do
# Train lora-ans without early stopping on 70% of the data and calibrate on 30% of the data
train_list="0.0-0.7"
val_list="0.0-0.3"
pred_list="0.7-1.0"
train_dir="outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list"
run_lora $model $dataset $size ans $num_seed $val_check_interval $train_dir false $train_list $val_list $pred_list
train_and_run_calibration "dp_calibration" $num_seed \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$pred_list" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_dpcal/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
train_and_run_calibration "temp_scaling" $num_seed \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$pred_list" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_tempscaling/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
train_and_run_calibration "vector_scaling" $num_seed \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$pred_list" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_vectorscaling/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
train_and_run_calibration "bias_shift" $num_seed \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$pred_list" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_biasshift/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
# Train lora-ans without early stopping on 100% of the data and calibrate using the above calibrated model
train_list="0.0-1.0"
val_list="0.7-1.0"
pred_list="0.7-1.0"
train_dir="outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list"
mkdir -p $train_dir
run_lora $model $dataset $size ans $num_seed $val_check_interval $train_dir false $train_list $val_list $test_list
run_calibration "dp_calibration" \
"outputs/lora_plus_dpcal/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/0.0-0.7/0.0-0.3/$pred_list/state.ckpt" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_dpcal/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
run_calibration "temp_scaling" \
"outputs/lora_plus_tempscaling/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/0.0-0.7/0.0-0.3/$pred_list/state.ckpt" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_tempscaling/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
run_calibration "vector_scaling" \
"outputs/lora_plus_vectorscaling/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/0.0-0.7/0.0-0.3/$pred_list/state.ckpt" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_vectorscaling/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
run_calibration "bias_shift" \
"outputs/lora_plus_biasshift/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/0.0-0.7/0.0-0.3/$pred_list/state.ckpt" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_biasshift/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
# Train lora-ans without early stopping on 100% of the data + naive calibration
train_list="0.0-1.0"
val_list="0.7-1.0"
pred_list="0.0-1.0"
train_dir="outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list"
mkdir -p $train_dir
run_lora $model $dataset $size ans $num_seed $val_check_interval $train_dir false $train_list $val_list $pred_list
train_and_run_calibration "dp_calibration" $num_seed \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$pred_list" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_dpcal_naive/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
train_and_run_calibration "temp_scaling" $num_seed \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$pred_list" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_tempscaling_naive/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
# Train lora-ans without early stopping on 100% of the data + calibration train on test
train_list="0.0-1.0"
val_list="0.7-1.0"
pred_list="0.7-1.0"
train_dir="outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list"
mkdir -p $train_dir
run_lora $model $dataset $size ans $num_seed $val_check_interval $train_dir false $train_list $val_list $test_list
train_and_run_calibration "dp_calibration" $num_seed \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_dpcal_trainontest/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/$test_list/test=$dataset/list=$test_list"
train_and_run_calibration "temp_scaling" $num_seed \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_tempscaling_trainontest/$model/$dataset/size=$size/seed=$num_seed/lora_ans_no_es/$train_list/$val_list/$test_list/test=$dataset/list=$test_list"
# Train lora-ans with early stopping on 70% of the data and calibrate on 30% of the data
train_list="0.0-0.7"
val_list="0.7-1.0"
pred_list="0.7-1.0"
train_dir="outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list"
mkdir -p $train_dir
run_lora $model $dataset $size ans $num_seed $val_check_interval $train_dir true $train_list $val_list $pred_list
train_and_run_calibration "dp_calibration" $num_seed \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$pred_list" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_dpcal/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
train_and_run_calibration "temp_scaling" $num_seed \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$pred_list" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_tempscaling/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
# train_and_run_calibration "vector_scaling" $num_seed \
# "outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$pred_list" \
# "outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$test_list" \
# "outputs/lora_plus_vectorscaling/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
# train_and_run_calibration "bias_shift" $num_seed \
# "outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$pred_list" \
# "outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$test_list" \
# "outputs/lora_plus_biasshift/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
# Train lora-ans with early stopping on 100% of the data and calibrate using the above calibrated model
train_list="0.0-1.0"
val_list="0.7-1.0"
pred_list="0.7-1.0"
train_dir="outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list"
mkdir -p $train_dir
run_lora $model $dataset $size ans $num_seed $val_check_interval $train_dir true $train_list $val_list $test_list
run_calibration "dp_calibration" \
"outputs/lora_plus_dpcal/$model/$dataset/size=$size/seed=$num_seed/lora_ans/0.0-0.7/0.7-1.0/$pred_list/state.ckpt" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_dpcal/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
run_calibration "temp_scaling" \
"outputs/lora_plus_tempscaling/$model/$dataset/size=$size/seed=$num_seed/lora_ans/0.0-0.7/0.7-1.0/$pred_list/state.ckpt" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_tempscaling/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
# run_calibration "vector_scaling" \
# "outputs/lora_plus_vectorscaling/$model/$dataset/size=$size/seed=$num_seed/lora_ans/0.0-0.7/0.7-1.0/$pred_list/state.ckpt" \
# "outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$test_list" \
# "outputs/lora_plus_vectorscaling/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
# run_calibration "bias_shift" \
# "outputs/lora_plus_biasshift/$model/$dataset/size=$size/seed=$num_seed/lora_ans/0.0-0.7/0.7-1.0/$pred_list/state.ckpt" \
# "outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$test_list" \
# "outputs/lora_plus_biasshift/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/$pred_list/test=$dataset/list=$test_list"
# Train lora-ans with early stopping on 100% of the data + calibration on test set
train_list="0.0-1.0"
val_list="0.7-1.0"
train_dir="outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list"
mkdir -p $train_dir
run_lora $model $dataset $size ans $num_seed $val_check_interval $train_dir true $train_list $val_list $test_list
train_and_run_calibration "dp_calibration" $num_seed \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_dpcal_trainontest/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/$test_list/test=$dataset/list=$test_list"
train_and_run_calibration "temp_scaling" $num_seed \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/finetune_lora/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/test=$dataset/list=$test_list" \
"outputs/lora_plus_tempscaling_trainontest/$model/$dataset/size=$size/seed=$num_seed/lora_ans/$train_list/$val_list/$test_list/test=$dataset/list=$test_list"
done
done
done
}
run_lora_vs_samples $model 16
|