Hi,
I'm experiencing a significant discrepancy between the performance reported in the paper and my own experimental results.
I noticed that if I run the sh scripts/etth1.sh script provided in the GitHub repository, the parameters used are quite different from those described in the paper, which leads to a major difference in performance.
Therefore, I tried to set the parameters exactly as specified in the paper for the ETTh1 dataset. Here is the configuration I am using:
if [ ! -d "./logs" ]; then
mkdir ./logs
fi
if [ ! -d "./logs/LongForecasting" ]; then
mkdir ./logs/LongForecasting
fi
model_name=EntroPE
root_path_name=./dataset/
entropy_model_checkpoint_dir=./entropy_model_checkpoints/
data_path_name=ETTh1.csv
model_id_name=ETTh1
data_name=ETTh1
enc_in=7
seq_len=96
quant_range=4
dim=8
multiple_of=128
heads=2
layers=1
batch_size=64
learning_rate=0.001
dropout=0.05
monotonicity=1
patching_threshold=3
patching_threshold_add=0.25
max_patch_length=24
train_epochs=20
patience=7
for random_seed in 2025
do
for pred_len in 192 336 720
do
python -u run_longExp_gpu1.py \
--random_seed $random_seed \
--is_training 1 \
--root_path $root_path_name \
--entropy_model_checkpoint_dir $entropy_model_checkpoint_dir \
--data_path $data_path_name \
--model_id $model_id_name_$seq_len'_'$pred_len \
--model_id_name $model_id_name \
--model $model_name \
--data $data_name \
--features M \
--seq_len $seq_len \
--pred_len $pred_len \
--enc_in $enc_in \
--vocab_size 256 \
--quant_range $quant_range \
--n_layers_local_encoder $layers \
--n_layers_local_decoder $layers \
--n_layers_global $layers \
--dim_global $dim \
--dim_local_encoder $dim \
--dim_local_decoder $dim \
--n_heads_local_encoder $heads \
--n_heads_local_decoder $heads \
--n_heads_global $heads \
--cross_attn_nheads $heads \
--dropout $dropout \
--multiple_of $multiple_of\
--max_patch_length $max_patch_length\
--patching_threshold $patching_threshold \
--patching_threshold_add $patching_threshold_add \
--monotonicity $monotonicity \
--des 'Exp' \
--train_epochs $train_epochs \
--patience $patience \
--lradj 'TST'\
--pct_start 0.3\
--batch_size $batch_size \
--patching_batch_size $((batch_size * enc_in)) \
--learning_rate $learning_rate \
>logs/LongForecasting/$model_name'_'$model_id_name'_'$seq_len'_'$pred_len.log
done
doneCould you please confirm if these are the correct settings? If there have been any updates or changes, I would greatly appreciate it if you could provide the exact parameter configuration required to reproduce the results reported in the paper.
Thank you
Hi,
I'm experiencing a significant discrepancy between the performance reported in the paper and my own experimental results.
I noticed that if I run the
sh scripts/etth1.shscript provided in the GitHub repository, the parameters used are quite different from those described in the paper, which leads to a major difference in performance.Therefore, I tried to set the parameters exactly as specified in the paper for the ETTh1 dataset. Here is the configuration I am using:
Could you please confirm if these are the correct settings? If there have been any updates or changes, I would greatly appreciate it if you could provide the exact parameter configuration required to reproduce the results reported in the paper.
Thank you