verify downsampling

.gitignore

@@ -8,4 +8,4 @@ results/
 dataset/
 checkpoints/
 testCuda.py
-mamba_ssm/wandb/*
+*wandb/*

TimeLLM/data_provider/data_loader.py

@@ -279,10 +279,21 @@ def __init__(self, root_path, flag='train', size=None,
         self.freq = freq
         self.percent = percent
 
+        # New parameters for dynamic downsampling
+        self.period_of_interest = period_of_interest
+        print(f"period of interest in dataset_ett_hour: {self.period_of_interest}")
+        self.timesteps = self.pred_len
+
         self.root_path = root_path
         self.data_path = data_path
         self.__read_data__()
 
+        self.downsampling_factor = calculate_downsampling_factor(self.root_path, self.data_path, self.period_of_interest, self.timesteps)
+        print("downsampling factor: ", self.downsampling_factor)
+        self.data_x = self.data_x[::self.downsampling_factor]
+        self.data_y = self.data_y[::self.downsampling_factor]
+        self.data_stamp = self.data_stamp[::self.downsampling_factor]
+
         self.enc_in = self.data_x.shape[-1]
         self.tot_len = len(self.data_x) - self.seq_len - self.pred_len + 1
 

TimeLLM/scripts/ALL_fulllsample_combine.sh

@@ -0,0 +1,108 @@
+#!/bin/bash
+
+model_name="TimeLLM"
+train_epochs=3
+learning_rate=0.01
+llm_layers=6
+
+num_params='2.8b'
+batch_size=16
+d_model=32
+d_ff=128
+num_process=1  # Ensure this variable is defined
+
+# Function to display usage information
+usage() {
+  echo "Usage: $0 -l <llm_layers> -d <d_model> -e <train_epochs> -n <num_params> -c <save_checkpoints> -m <llm_model> -p <period_of_interest> -g <gpu_id>"
+  exit 1
+}
+
+# Parse command-line arguments
+while getopts "l:d:e:n:c:m:p:g:" opt; do
+  case $opt in
+    l) llm_layers=$OPTARG ;;
+    d) d_model=$OPTARG ;;
+    e) train_epochs=$OPTARG ;;
+    n) num_params=$OPTARG ;;
+    c) save_checkpoints=$OPTARG ;;
+    m) llm_model=$OPTARG ;;
+    p) period_of_interest=$OPTARG ;;
+    g) gpu_id=$OPTARG ;;
+    *) usage ;;
+  esac
+done
+
+# Check if required arguments are provided
+if [ -z "$llm_layers" ] || [ -z "$d_model" ] || [ -z "$train_epochs" ] || [ -z "$num_params" ] || [ -z "$save_checkpoints" ] || [ -z "$llm_model" ] || [ -z "$period_of_interest" ] || [ -z "$gpu_id" ]; then
+  usage
+fi
+
+# Export GPU ID and adjust master_port
+export CUDA_VISIBLE_DEVICES=$((gpu_id % 4))
+master_port_base=01180
+master_port="${master_port_base}${gpu_id}"
+
+# Print the values to verify
+echo "Setting llm_layers to $llm_layers"
+echo "Setting d_model to $d_model"
+echo "Setting train_epochs to $train_epochs"
+echo "Setting num_params to $num_params"
+echo "Setting save_checkpoints to $save_checkpoints"
+echo "Setting llm_model to $llm_model"
+echo "Setting period_of_interest to $period_of_interest"
+echo "Setting master_port to $master_port"  # Print master_port for debugging
+
+og_tag="l${llm_layers}_d${d_model}_e${train_epochs}_m${llm_model}_n${num_params}_p${period_of_interest// /_}"
+
+llm_dim=0
+if [ "$num_params" = "130m" ]; then
+  llm_dim=768
+fi
+if [[ "$num_params" == "2.7b" || "$num_params" == "2.8b" ]]; then
+  llm_dim=2560
+fi
+if [ "$num_params" = "7b" ]; then
+  llm_dim=4096
+fi
+
+# Define trials for different pred_len values
+for seed in {1..3}; do
+  tag="ETTh2_${og_tag}_pred${pred_len}_seed${seed}"
+  comment="checkpoints/${tag}"
+  log_file="results/${tag}.txt"
+  exec > "$log_file" 2>&1
+
+  accelerate launch --mixed_precision bf16 --num_processes $num_process --main_process_port $master_port seed_process.py \
+    --task_name long_term_forecast \
+    --is_training 1 \
+    --root_path ./dataset/ETT-small/ \
+    --data_path ETTh2.csv \
+    --model_id "ETTh2_512_96" \
+    --model $model_name \
+    --data ETTh2 \
+    --features M \
+    --seq_len 512 \
+    --label_len 48 \
+    --pred_len 96 \
+    --factor 3 \
+    --enc_in 7 \
+    --dec_in 7 \
+    --c_out 7 \
+    --des 'Exp' \
+    --itr 1 \
+    --d_model $d_model \
+    --d_ff $d_ff \
+    --batch_size $batch_size \
+    --learning_rate $learning_rate \
+    --llm_layers $llm_layers \
+    --train_epochs $train_epochs \
+    --model_comment "$comment" \
+    --llm_model $llm_model \
+    --llm_dim $llm_dim \
+    --num_params $num_params \
+    --period_of_interest "None" \
+    --seed $seed
+
+  echo "ETTh2 with seed $seed completed, saved to $comment"
+done
+

TimeLLM/scripts/clean_combine.sh

@@ -66,8 +66,8 @@ if [ "$num_params" = "7b" ]; then
 fi
 
 # Define trials for different pred_len values
-for pred_len in 48 64; do
-  for seed in {1..10}; do
+for pred_len in 64; do
+  for seed in {1..3}; do
     tag="ETTh1_${og_tag}_pred${pred_len}_seed${seed}"
     comment="checkpoints/${tag}"
     log_file="results/${tag}.txt"

TimeLLM/scripts/combineGPU2.sh

@@ -105,7 +105,7 @@ done
 '
 
 tag="ETTm1_${og_tag}"
-for seed in {3..10}; do
+for seed in {1..10}; do
   comment="checkpoints/${tag}_seed${seed}"
   log_file="results/${tag}_seed${seed}.txt"
   exec > "$log_file" 2>&1

TimeLLM/scripts/smalltest.sh

@@ -12,7 +12,7 @@ d_ff=128
 comment='checkpoints/smallTest'
 gpu_id=3
 export CUDA_VISIBLE_DEVICES=$gpu_id
-
+: '
  accelerate launch --mixed_precision bf16 --num_processes $num_process --main_process_port $master_port seed_process.py \
       --task_name long_term_forecast \
       --is_training 1 \
@@ -40,9 +40,10 @@ export CUDA_VISIBLE_DEVICES=$gpu_id
       --model_comment "$comment" \
       --llm_model Mamba2 \
       --llm_dim 768 \
-      --period_of_interest "1 month" \
+      --period_of_interest "None" \
       --use_wandb 0
-: ' 
+      '
+
 accelerate launch --mixed_precision bf16 --num_processes 1 --gpu_ids $gpu_id --main_process_port $master_port seed_process.py \
   --task_name long_term_forecast \
   --is_training 1 \
@@ -52,9 +53,9 @@ accelerate launch --mixed_precision bf16 --num_processes 1 --gpu_ids $gpu_id --m
   --model $model_name \
   --data Traffic \
   --features M \
-  --seq_len 512 \
+  --seq_len 96 \
   --label_len 48 \
-  --pred_len 96 \
+  --pred_len 128\
   --e_layers 2 \
   --d_layers 1 \
   --factor 3 \
@@ -72,7 +73,7 @@ accelerate launch --mixed_precision bf16 --num_processes 1 --gpu_ids $gpu_id --m
   --period_of_interest "1 month" \
   --use_wandb 0
 
-
+: '
 accelerate launch --mixed_precision bf16 --num_processes 1 --gpu_ids $gpu_id --main_process_port $master_port seed_process.py \
   --task_name long_term_forecast \
   --is_training 1 \

TimeLLM/scripts/traffic_combine.sh

@@ -0,0 +1,109 @@
+#!/bin/bash
+
+model_name="TimeLLM"
+train_epochs=3
+learning_rate=0.01
+llm_layers=6
+
+num_params='2.8b'
+batch_size=16
+d_model=32
+d_ff=128
+num_process=1  # Ensure this variable is defined
+
+# Function to display usage information
+usage() {
+  echo "Usage: $0 -l <llm_layers> -d <d_model> -e <train_epochs> -n <num_params> -c <save_checkpoints> -m <llm_model> -p <period_of_interest> -g <gpu_id>"
+  exit 1
+}
+
+# Parse command-line arguments
+while getopts "l:d:e:n:c:m:p:g:" opt; do
+  case $opt in
+    l) llm_layers=$OPTARG ;;
+    d) d_model=$OPTARG ;;
+    e) train_epochs=$OPTARG ;;
+    n) num_params=$OPTARG ;;
+    c) save_checkpoints=$OPTARG ;;
+    m) llm_model=$OPTARG ;;
+    p) period_of_interest=$OPTARG ;;
+    g) gpu_id=$OPTARG ;;
+    *) usage ;;
+  esac
+done
+
+# Check if required arguments are provided
+if [ -z "$llm_layers" ] || [ -z "$d_model" ] || [ -z "$train_epochs" ] || [ -z "$num_params" ] || [ -z "$save_checkpoints" ] || [ -z "$llm_model" ] || [ -z "$period_of_interest" ] || [ -z "$gpu_id" ]; then
+  usage
+fi
+
+# Export GPU ID and adjust master_port
+export CUDA_VISIBLE_DEVICES=$((gpu_id % 4))
+master_port_base=01180
+master_port="${master_port_base}${gpu_id}"
+
+# Print the values to verify
+echo "Setting llm_layers to $llm_layers"
+echo "Setting d_model to $d_model"
+echo "Setting train_epochs to $train_epochs"
+echo "Setting num_params to $num_params"
+echo "Setting save_checkpoints to $save_checkpoints"
+echo "Setting llm_model to $llm_model"
+echo "Setting period_of_interest to $period_of_interest"
+echo "Setting master_port to $master_port"  # Print master_port for debugging
+
+og_tag="l${llm_layers}_d${d_model}_e${train_epochs}_m${llm_model}_n${num_params}_p${period_of_interest// /_}"
+
+llm_dim=0
+if [ "$num_params" = "130m" ]; then
+  llm_dim=768
+fi
+if [[ "$num_params" == "2.7b" || "$num_params" == "2.8b" ]]; then
+  llm_dim=2560
+fi
+if [ "$num_params" = "7b" ]; then
+  llm_dim=4096
+fi
+
+# Define trials for different pred_len values
+for pred_len in 96 128; do
+  for seed in {1..3}; do
+    tag="Traffic_${og_tag}_pred${pred_len}_seed${seed}"
+    comment="checkpoints/${tag}"
+    log_file="results/${tag}.txt"
+    exec > "$log_file" 2>&1
+
+    accelerate launch --mixed_precision bf16 --num_processes 1 --gpu_ids $gpu_id --main_process_port $master_port seed_process.py \
+      --task_name long_term_forecast \
+      --is_training 1 \
+      --root_path ./dataset/traffic/ \
+      --data_path traffic.csv \
+      --model_id "traffic_96_${pred_len}" \
+      --model $model_name \
+      --data Traffic \
+      --features M \
+      --seq_len 96 \
+      --label_len 24 \
+      --pred_len $pred_len \
+      --factor 3 \
+      --enc_in 7 \
+      --dec_in 7 \
+      --c_out 7 \
+      --des 'Exp' \
+      --itr 1 \
+      --d_model $d_model \
+      --d_ff $d_ff \
+      --batch_size $batch_size \
+      --learning_rate $learning_rate \
+      --llm_layers $llm_layers \
+      --train_epochs $train_epochs \
+      --model_comment "$comment" \
+      --llm_model $llm_model \
+      --llm_dim $llm_dim \
+      --num_params $num_params \
+      --period_of_interest "$period_of_interest" \
+      --seed $seed
+
+    echo "ETTh1 with pred_len $pred_len and seed $seed completed, saved to $comment"
+  done
+done