Pressure Field work

RegDGCNN_SurfaceFields/Model_Evaluation.sh

@@ -0,0 +1,10 @@
+python run_pipeline.py \
+    --stages evaluate \
+    --exp_name "DrivAerNet_Pressure" \
+    --dataset_path "$HOME/ML_Turbulent/Data_Pressure_Field/Data_Pressure/PressureVTK/N_S_WWS_WM" \
+    --cache_dir "$HOME/ML_Turbulent/Data_Pressure_Field/Cache_data" \
+    --subset_dir "$HOME/ML_Turbulent/DrivAerNet/train_val_test_splits" \
+    --num_points 10000 \
+    --num_eval_samples 5 \
+    --gpus "0"
+

RegDGCNN_SurfaceFields/Model_Preprocess.sh

@@ -0,0 +1,7 @@
+python run_pipeline.py \
+    --stages preprocess \
+    --exp_name "DrivAerNet_Pressure" \
+    --dataset_path "$HOME/ML_Turbulent/Pressure_Field/Data_Pressure/PressureVTK/N_S_WWS_WM" \
+    --cache_dir "$HOME/ML_Turbulent/Pressure_Field/Cache_data" \
+    --subset_dir "$HOME/ML_Turbulent/Pressure_Field/train_val_test_splits" \
+    --num_points 10000

RegDGCNN_SurfaceFields/Model_Test.sh

@@ -0,0 +1,9 @@
+python test_single_vtk.py \
+    --model_checkpoint "experiments/DrivAerNet_Pressure/best_model.pth" \
+    --vtk_file "$HOME/ML_Turbulent/Data_Pressure_Field/Data_Pressure/PressureVTK/N_S_WWS_WM/N_S_WWS_WM_001.vtk" \
+    --output_dir "visualizations" \
+    --num_points 10000 \
+    --k 40 \
+    --emb_dims 1024 \
+    --dropout 0.4
+

RegDGCNN_SurfaceFields/Model_Training.sh

@@ -0,0 +1,22 @@
+python run_pipeline.py \
+    --stages train \
+    --exp_name "DrivAerNet_Pressure" \
+    --dataset_path "$HOME/ML_Turbulent/Data_Pressure_Field/Data_Pressure/PressureVTK/N_S_WWS_WM" \
+    --subset_dir "$HOME/ML_Turbulent/DrivAerNet/train_val_test_splits" \
+    --cache_dir "$HOME/ML_Turbulent/Data_Pressure_Field/Cache_data" \
+    --num_points 10000 \
+    --batch_size 12 \
+    --epochs 150 \
+    --gpus "0, 1"
+
+
+# === Variable  ===
+# --batch_size 
+# It defines how many samples are processed at once per training step
+#
+# --epochs
+# 
+# Total number of training rounds over the whole dataset
+# e.g. 10000 samples, batch_size = 100
+#      10 000 / 100 = 100 steps for one epoch
+#      --epochs 150 have 150 times loop

RegDGCNN_SurfaceFields/My_python_job/Model_Training.sh

@@ -0,0 +1,22 @@
+python run_pipeline.py \
+    --stages train \
+    --exp_name "DrivAerNet_Pressure" \
+    --dataset_path "$HOME/ML_Turbulent/Data_Pressure_Field/Data_Pressure/PressureVTK/N_S_WWS_WM" \
+    --subset_dir "$HOME/ML_Turbulent/DrivAerNet/train_val_test_splits" \
+    --cache_dir "$HOME/ML_Turbulent/Data_Pressure_Field/Cache_data" \
+    --num_points 10000 \
+    --batch_size 12 \
+    --epochs 150 \
+    --gpus "0, 1, 2"
+
+
+# === Variable  ===
+# --batch_size 
+# It defines how many samples are processed at once per training step
+#
+# --epochs
+# 
+# Total number of training rounds over the whole dataset
+# e.g. 10000 samples, batch_size = 100
+#      10 000 / 100 = 100 steps for one epoch
+#      --epochs 150 have 150 times loop

RegDGCNN_SurfaceFields/My_python_job/Usage_data_loader.py

@@ -0,0 +1,3 @@
+"""
+
+"""

RegDGCNN_SurfaceFields/My_python_job/Usage_model_pressure.py

@@ -0,0 +1,11 @@
+import torch
+import torch.nn as nn
+
+"""
+
+"""
+
+
+
+
+

RegDGCNN_SurfaceFields/My_python_job/Usage_train.py

@@ -0,0 +1,148 @@
+# ============ Function Usage ============
+import torch
+import torch.distributed as dist
+import torch.multiprocessing as mp
+
+# Import modules
+from model_pressure import RegDGCNN_pressure
+'''
+
+# DDP: Distributed Data Parallel
+1.----
+    world_size = len(gpu_list.split(','))
+#! 
+    gpu_list                   -> "0, 1, 2"
+    gpu_list.split(',')        -> ['0', '1', '2']
+    len(gpu_list.split(','))   -> 3
+
+2.----
+    exp_dir = os.path.join('experiments', args.exp_name)
+#!
+    Build a path "./experiments/exp_name"
+
+3.----
+    os.makedirs(exp_dir, exist_ok=True)
+#!
+    Create the directory "exp_dir" if it doesn't already exist
+
+4.----
+    mp.spawn(train_and_evaluate, args=(world_size, args), nprocs=world_size, join=True)
+#!
+   train_and_evaluate(rank, world_size, args)
+   # rank :        which GPU this process is using
+   # world_size :  total number of GPUS
+   # args:         your parsed command-line arguments 
+#!
+    mp.spawn(...)
+    ->
+    train_evaluate(rank=0, world_size= , args=args)
+    train_evaluate(rank=1, world_size= , args=args)
+    train_evaluate(rank=2, world_size= , args=args)
+    train_evaluate(rank=3, world_size= , args=args)
+    ...
+
+5.----
+    dist.init_process_group(backend='nccl', init_method='env://', world_size=world_size, rank=rank)
+#!
+    Starts the communication backend for DDP
+    # nccl:        NVDIA backend for multi-GPU
+    # evv://       Uses evvironmen variables(e.g. RANK, WORLD_SIZE, etc)
+
+6.-----
+    torch.cuda.set_device(local_rank)
+#! 
+    Each process uses a single GPU
+    Set the GPU this process will use
+
+7.-----
+    args = vars(args)
+#!
+    Convert it to a regular dictionary
+    e.g. {'epoch': 10, 'lr': 0.0001}
+
+8.-----
+    model = RegDGCNN_pressure(args).to(local_rank)
+#!
+    RegDGCNN_pressure(args): Creates an instance of your custom model class
+    .to(local_rank):         Moves the model to the current corret GPU
+#!
+    The detail please see Usage_model_pressure.pu
+
+10.----
+    model = torch.nn.parallel.DistributedDataParallel(
+        model,
+        device_ids=[local_rank],
+        find_unused_parameters=True, 
+        output_device=local_rank
+    )
+#!
+    Wrap it with DDP so Pytorch handles multi-GPU
+        synchronization and data parallelism
+#!
+    model
+    The model aleady moved to the local GPU
+#!
+    device_ids=[local_rank]
+    Restrict this process to use only one GPU
+#!
+    find_unused_parameters=True
+    It tells pyTorch "Some layers in my model might not be used every time I call forward()
+    -> so please handle that correctly"
+        The detail please see Usage_model_pressure.py forwar() function
+
+#!  
+    output_device=local_rank
+    Ensures outputs go to the same GPU as inputs
+
+11.----
+    criterion = torch.nn.MSELoss()
+#!
+    Loss function:
+    This sets up Mean Squared Error(MSE) as the loss function, commonly used for regression problems
+    It measures the average of the squares of the differences between predicted and actual values:
+        MSE = 1/n * sum((yi - y)^2)      i = 1, ... , n
+
+12.----
+    optimizer = optim.Adam(model.parameters(), lr=args.lr, weight_decay=1e-4)
+#!
+    Uses Adam(Adaptive Moment Estimation), a popular optimizer that adjusts learning rates for each parameter
+#!
+    model.parameter()    
+    ->Tell the optimizer which parameters to update
+#!
+    lr = args.lr
+    ->learning rate 
+#!
+    weight_decay=1e-4
+    ->adds L2 regularization to reduce overfitting
+
+13.----
+    scheduler = ReduceLROnPlateau(optimizer, 'min', patience=10, factor=0.1, verbose=True)
+#!    
+    Automatically reduces the learning rate when the validation loss plateaus(stops improving)
+#!
+    'min'
+    ->Try to minimize val loss
+    -> By default 1e-4
+#!
+    patience=10 
+    ->Wait for 10 epochs without improvement
+#!
+    factor=0.1
+    ->Multiplies LR by 0.1
+#!
+    e.g.
+    Epoch 1 - Validation Loss: 0.470
+    LR : 0.01
+    Epoch 2 - Validation Loss: 0.471
+    LR : 0.01
+    Epoch 3 - Validation Loss: 0.471
+    LR : 0.01
+    Epoch 4 - Validation Loss: 0.470
+    LR : 0.01*0.1
+
+'''
+
+
+
+