Cleanroom / cleanroom /surrogateAI /second_train.py
bipin-ank's picture
Upload cleanroom dataset
0bd444e
import os
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from timm.models.layers import trunc_normal_
from einops import rearrange
import numpy as np
from typing import List, Dict, Any
import matplotlib.pyplot as plt
import json
import random
import time
from clean_eval import evaluate as rollout_evaluate
# ============================================================================
# 1. PYTORCH DATASET CLASS FOR TIME-STEPPED DATA
# ============================================================================
class TimeStepDataset(Dataset):
"""Dataset for flattened timestep training data"""
def __init__(self, pt_file_path: str):
if not os.path.exists(pt_file_path):
raise FileNotFoundError(f"Dataset file not found at: {pt_file_path}")
self.data = torch.load(pt_file_path, weights_only=False)
print(f"Loaded {len(self.data)} timesteps from {pt_file_path}")
def __len__(self) -> int:
return len(self.data)
def __getitem__(self, idx: int) -> Dict[str, Any]:
sample = self.data[idx]
return {
'coordinates': sample['coordinates'],
'node_type': sample['node_type'],
'current_velocities': sample['current_velocities'],
'target_velocities': sample['target_velocities'],
'meta_info': sample['meta_info']
}
class TrajectoryDataset(Dataset):
"""Dataset for full trajectory evaluation data"""
def __init__(self, pt_file_path: str):
if not os.path.exists(pt_file_path):
raise FileNotFoundError(f"Dataset file not found at: {pt_file_path}")
self.trajectories = torch.load(pt_file_path, weights_only=False)
print(f"Loaded {len(self.trajectories)} trajectories from {pt_file_path}")
def __len__(self) -> int:
return len(self.trajectories)
def __getitem__(self, idx: int) -> Dict[str, Any]:
return self.trajectories[idx]
# ============================================================================
# 2. TRANSOLVER MODEL ARCHITECTURE (Adapted for Time-Dependent Prediction)
# ============================================================================
ACTIVATION = {'gelu': nn.GELU, 'tanh': nn.Tanh, 'sigmoid': nn.Sigmoid, 'relu': nn.ReLU, 'leaky_relu': nn.LeakyReLU(0.1)}
class MLP(nn.Module):
def __init__(self, n_input, n_hidden, n_output, n_layers=1, act='gelu', res=True):
super(MLP, self).__init__()
if act in ACTIVATION.keys():
act = ACTIVATION[act]
else:
raise NotImplementedError
self.linear_pre = nn.Sequential(nn.Linear(n_input, n_hidden), act())
self.linear_post = nn.Linear(n_hidden, n_output)
self.linears = nn.ModuleList([nn.Sequential(nn.Linear(n_hidden, n_hidden), act()) for _ in range(n_layers)])
self.res = res
def forward(self, x):
x = self.linear_pre(x)
for i in range(len(self.linears)):
if self.res:
x = self.linears[i](x) + x
else:
x = self.linears[i](x)
x = self.linear_post(x)
return x
class Physics_Attention_Irregular_Mesh(nn.Module):
def __init__(self, dim, heads=8, dim_head=64, dropout=0., slice_num=64):
super().__init__()
inner_dim = dim_head * heads
self.dim_head = dim_head
self.heads = heads
self.scale = dim_head ** -0.5
self.softmax = nn.Softmax(dim=-1)
self.dropout = nn.Dropout(dropout)
self.temperature = nn.Parameter(torch.ones([1, heads, 1, 1]) * 0.5)
self.in_project_x = nn.Linear(dim, inner_dim)
self.in_project_fx = nn.Linear(dim, inner_dim)
self.in_project_slice = nn.Linear(dim_head, slice_num)
torch.nn.init.orthogonal_(self.in_project_slice.weight)
self.to_q = nn.Linear(dim_head, dim_head, bias=False)
self.to_k = nn.Linear(dim_head, dim_head, bias=False)
self.to_v = nn.Linear(dim_head, dim_head, bias=False)
self.to_out = nn.Sequential(nn.Linear(inner_dim, dim), nn.Dropout(dropout))
def forward(self, x):
B, N, C = x.shape
fx_mid = self.in_project_fx(x).reshape(B, N, self.heads, self.dim_head).permute(0, 2, 1, 3).contiguous()
x_mid = self.in_project_x(x).reshape(B, N, self.heads, self.dim_head).permute(0, 2, 1, 3).contiguous()
slice_weights = self.softmax(self.in_project_slice(x_mid) / self.temperature)
slice_norm = slice_weights.sum(2)
slice_token = torch.einsum("bhnc,bhng->bhgc", fx_mid, slice_weights)
slice_token = slice_token / ((slice_norm + 1e-5)[:, :, :, None].repeat(1, 1, 1, self.dim_head))
q_slice_token = self.to_q(slice_token)
k_slice_token = self.to_k(slice_token)
v_slice_token = self.to_v(slice_token)
dots = torch.matmul(q_slice_token, k_slice_token.transpose(-1, -2)) * self.scale
attn = self.softmax(dots)
attn = self.dropout(attn)
out_slice_token = torch.matmul(attn, v_slice_token)
out_x = torch.einsum("bhgc,bhng->bhnc", out_slice_token, slice_weights)
out_x = rearrange(out_x, 'b h n d -> b n (h d)')
out = self.to_out(out_x)
return out
class Transolver_block(nn.Module):
def __init__(self, num_heads, hidden_dim, dropout, act='gelu', mlp_ratio=4, last_layer=False, out_dim=1, slice_num=32):
super().__init__()
self.last_layer = last_layer
self.ln_1 = nn.LayerNorm(hidden_dim)
self.Attn = Physics_Attention_Irregular_Mesh(hidden_dim, heads=num_heads, dim_head=hidden_dim // num_heads, dropout=dropout, slice_num=slice_num)
self.ln_2 = nn.LayerNorm(hidden_dim)
self.mlp = MLP(hidden_dim, hidden_dim * mlp_ratio, hidden_dim, n_layers=0, res=False, act=act)
if self.last_layer:
self.ln_3 = nn.LayerNorm(hidden_dim)
self.mlp2 = nn.Linear(hidden_dim, out_dim)
def forward(self, fx):
fx2 = self.Attn(self.ln_1(fx)) + fx
fx = self.mlp(self.ln_2(fx2)) + fx2
if self.last_layer:
return self.mlp2(self.ln_3(fx))
return fx
class Model(nn.Module):
def __init__(self, in_dim=13, out_dim=3, n_layers=8, n_hidden=256, dropout=0, n_head=8, act='gelu', mlp_ratio=2, slice_num=32):
super(Model, self).__init__()
self.preprocess = MLP(in_dim, n_hidden * 2, n_hidden, n_layers=0, res=False, act=act)
self.n_hidden = n_hidden
self.blocks = nn.ModuleList([Transolver_block(num_heads=n_head, hidden_dim=n_hidden, dropout=dropout,
act=act, mlp_ratio=mlp_ratio,
out_dim=out_dim,
slice_num=slice_num,
last_layer=(_ == n_layers - 1))
for _ in range(n_layers)])
self.initialize_weights()
self.placeholder = nn.Parameter((1 / n_hidden) * torch.rand(n_hidden, dtype=torch.float))
def initialize_weights(self):
self.apply(self._init_weights)
def _init_weights(self, m):
if isinstance(m, nn.Linear):
trunc_normal_(m.weight, std=0.02)
if isinstance(m, nn.Linear) and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, (nn.LayerNorm, nn.BatchNorm1d)):
nn.init.constant_(m.bias, 0)
nn.init.constant_(m.weight, 1.0)
def forward(self, x):
fx = self.preprocess(x)
fx = fx + self.placeholder[None, None, :]
for block in self.blocks:
fx = block(fx)
return fx
# ============================================================================
# 3. CHECKPOINTING FUNCTIONS
# ============================================================================
def save_checkpoint(model, optimizer, scheduler, epoch, loss, path, norm_stats):
"""Save training checkpoint"""
checkpoint = {
'epoch': epoch,
'model_state_dict': model.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'scheduler_state_dict': scheduler.state_dict() if scheduler else None,
'loss': loss,
'norm_stats': norm_stats # Save normalization stats
}
torch.save(checkpoint, path)
print(f"Checkpoint saved: {path}")
def load_checkpoint(model, optimizer, scheduler, path):
"""Load training checkpoint"""
checkpoint = torch.load(path, map_location='cpu')
model.load_state_dict(checkpoint['model_state_dict'])
optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
if scheduler and checkpoint.get('scheduler_state_dict'):
scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
# Load norm_stats if they exist, otherwise return None
norm_stats = checkpoint.get('norm_stats', None)
if norm_stats is None:
print("Warning: Normalization stats not found in checkpoint.")
return checkpoint['epoch'], checkpoint['loss'], norm_stats
# ============================================================================
# 4. EVALUATION FUNCTIONS
# ============================================================================
def create_wall_mask(node_type):
"""Create mask to ignore wall nodes during loss calculation"""
wall_class_index = 5 # Assuming walls are class 5
node_class_indices = torch.argmax(node_type, dim=-1)
mask = (node_class_indices != wall_class_index).float().unsqueeze(-1)
return mask
def evaluate_rollout(model, trajectory_loader, device, norm_stats):
"""Run rollout evaluation on full trajectories"""
model.eval()
results = []
print(f"Starting rollout evaluation on {len(trajectory_loader)} trajectories...")
with torch.no_grad():
for i, trajectory in enumerate(trajectory_loader):
try:
print(f"\nProcessing trajectory {i+1}/{len(trajectory_loader)}")
# Since trajectory_loader has batch_size=1, we need to extract the single trajectory
single_trajectory = {}
for key, value in trajectory.items():
if key != 'meta_info':
# Remove the batch dimension
single_trajectory[key] = value.squeeze(0)
else:
# meta_info is special - it's a dict where each value is a list/tensor due to batching
# We need to extract the first element from each value in the meta_info dict
single_trajectory[key] = {}
for meta_key, meta_value in value.items():
if isinstance(meta_value, (list, tuple)):
single_trajectory[key][meta_key] = meta_value[0]
elif isinstance(meta_value, dict):
# Handle nested dictionaries (like node_type_counts)
single_trajectory[key][meta_key] = {}
for nested_key, nested_value in meta_value.items():
if isinstance(nested_value, (list, tuple)):
single_trajectory[key][meta_key][nested_key] = nested_value[0]
elif torch.is_tensor(nested_value) and nested_value.numel() > 1:
single_trajectory[key][meta_key][nested_key] = nested_value[0]
else:
single_trajectory[key][meta_key][nested_key] = nested_value
elif torch.is_tensor(meta_value) and meta_value.numel() > 1:
# For tensors with more than one element, take the first element
single_trajectory[key][meta_key] = meta_value[0]
else:
# For scalars or other types, keep as is
single_trajectory[key][meta_key] = meta_value
# Pass the normalization statistics to the evaluation function
_, traj_result = rollout_evaluate(model, single_trajectory, norm_stats)
results.append(traj_result)
print(f"Successfully processed trajectory {i+1}")
except Exception as e:
print(f"Error processing trajectory {i+1}: {str(e)}")
print(f"Trajectory data shapes:")
for key, value in trajectory.items():
if hasattr(value, 'shape'):
print(f" {key}: {value.shape}")
else:
print(f" {key}: {type(value)} - {value}")
raise e
print(f"Completed rollout evaluation on {len(results)} trajectories")
return results
# ============================================================================
# 5. MAIN TRAINING SCRIPT
# ============================================================================
def main():
script_start_time = time.time()
# --- CONFIGURATION ---
preprocessed_dir = r"/home/gd_user1/AnK/project_PINN/cleanroom/rawdataset"
base_filename = "final_data_timestep_data"
output_dir = r"/home/gd_user1/AnK/project_PINN/cleanroom/second_output/output_2"
train_pt_path = os.path.join(preprocessed_dir, f"{base_filename}_train.pt")
val_pt_path = os.path.join(preprocessed_dir, f"{base_filename}_test.pt")
os.makedirs(output_dir, exist_ok=True)
hparams = {
'lr': 0.0005, 'batch_size': 1, 'nb_epochs': 400,
'in_dim': 13, # coordinates(3) + node_type(6) + current_vel(3) + inlet_vel(1)
'out_dim': 3, 'n_hidden': 256, 'n_layers': 6, 'n_head': 8, 'slice_num': 32,
'checkpoint_freq': 10, 'eval_freq': 10
}
# --- SETUP ---
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Using device: {device}")
if torch.cuda.is_available():
print(f"GPU: {torch.cuda.get_device_name(0)}")
print(f"CUDA version: {torch.version.cuda}")
# Set random seeds for reproducibility
torch.manual_seed(42)
random.seed(42)
np.random.seed(42)
# --- DATALOADERS ---
print("Loading datasets...")
train_dataset = TimeStepDataset(train_pt_path)
val_timestep_dataset = TimeStepDataset(val_pt_path) # For single-step validation
test_trajectory_dataset = TrajectoryDataset(val_pt_path) # For rollout evaluation
train_loader = DataLoader(train_dataset, batch_size=hparams['batch_size'], shuffle=True)
val_timestep_loader = DataLoader(val_timestep_dataset, batch_size=hparams['batch_size'], shuffle=False)
test_trajectory_loader = DataLoader(test_trajectory_dataset, batch_size=1, shuffle=False)
# --- COMPUTE NORMALIZATION STATISTICS FROM TRAINING DATA ---
def compute_normalization_stats(train_dataset):
"""Compute normalization statistics from training dataset"""
coords_list = []
vel_list = []
inlet_vel_list = []
print("Computing normalization statistics from training data...")
for sample in train_dataset:
coords_list.append(sample['coordinates'])
vel_list.append(sample['current_velocities'])
# Handle inlet velocity (it's in meta_info and may be tensor or scalar)
inlet_vel = sample['meta_info']['velocity']
if torch.is_tensor(inlet_vel):
inlet_vel_list.append(inlet_vel.item())
else:
inlet_vel_list.append(float(inlet_vel))
# Stack and compute statistics
coords_all = torch.stack(coords_list) # [num_samples, num_nodes, 3]
vel_all = torch.stack(vel_list) # [num_samples, num_nodes, 3]
inlet_vel_all = torch.tensor(inlet_vel_list) # [num_samples]
stats = {
'coords_mean': coords_all.mean(dim=[0, 1]), # Mean across samples and nodes
'coords_std': coords_all.std(dim=[0, 1]),
'vel_mean': vel_all.mean(dim=[0, 1]),
'vel_std': vel_all.std(dim=[0, 1]),
'inlet_vel_mean': inlet_vel_all.mean(),
'inlet_vel_std': inlet_vel_all.std()
}
print(f"Computed stats - Coords mean: {stats['coords_mean']}, std: {stats['coords_std']}")
print(f"Velocities mean: {stats['vel_mean']}, std: {stats['vel_std']}")
print(f"Inlet vel mean: {stats['inlet_vel_mean']:.6f}, std: {stats['inlet_vel_std']:.6f}")
return stats
# Compute normalization statistics
norm_stats = compute_normalization_stats(train_dataset)
coords_mean = norm_stats['coords_mean'].to(device)
coords_std = norm_stats['coords_std'].to(device)
vel_mean = norm_stats['vel_mean'].to(device)
vel_std = norm_stats['vel_std'].to(device)
inlet_vel_mean = norm_stats['inlet_vel_mean'].to(device)
inlet_vel_std = norm_stats['inlet_vel_std'].to(device)
epsilon = 1e-8 # Small constant to prevent division by zero
# --- MODEL, OPTIMIZER, SCHEDULER SETUP ---
model = Model(in_dim=hparams['in_dim'], out_dim=hparams['out_dim'],
n_hidden=hparams['n_hidden'], n_layers=hparams['n_layers'],
n_head=hparams['n_head'], slice_num=hparams['slice_num']).to(device)
optimizer = torch.optim.Adam(model.parameters(), lr=hparams['lr'])
scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=50, gamma=0.9)
criterion = nn.MSELoss()
# Initialize training variables
start_epoch = 0
best_val_loss = float('inf')
# Check for existing checkpoint to resume training
latest_checkpoint_path = os.path.join(output_dir, 'checkpoint_latest.pth')
if os.path.exists(latest_checkpoint_path):
print(f"Found existing checkpoint: {latest_checkpoint_path}")
try:
start_epoch, last_loss, loaded_norm_stats = load_checkpoint(
model, optimizer, scheduler, latest_checkpoint_path
)
if loaded_norm_stats is not None:
norm_stats = loaded_norm_stats
# Update the normalization tensors
coords_mean = norm_stats['coords_mean'].to(device)
coords_std = norm_stats['coords_std'].to(device)
vel_mean = norm_stats['vel_mean'].to(device)
vel_std = norm_stats['vel_std'].to(device)
inlet_vel_mean = norm_stats['inlet_vel_mean'].to(device)
inlet_vel_std = norm_stats['inlet_vel_std'].to(device)
print("Loaded normalization statistics from checkpoint.")
print(f"Resuming training from epoch {start_epoch}")
best_val_loss = last_loss
except Exception as e:
print(f"Error loading checkpoint: {e}")
print("Starting fresh training...")
start_epoch = 0
best_val_loss = float('inf')
print(f"Starting training from epoch {start_epoch} to {hparams['nb_epochs']}")
print("="*60)
train_loss_history = []
val_loss_history = []
lr_history = []
val_epochs = []
# --- TRAINING LOOP ---
for epoch in range(start_epoch, hparams['nb_epochs']):
epoch_start_time = time.time()
model.train()
total_train_loss = 0.0
for batch_idx, batch in enumerate(train_loader):
coordinates = batch['coordinates'].to(device)
node_type = batch['node_type'].to(device)
current_velocities = batch['current_velocities'].to(device)
target_velocities = batch['target_velocities'].to(device)
inlet_velocities = batch['meta_info']['velocity']
if torch.is_tensor(inlet_velocities):
inlet_vel_tensor = inlet_velocities.clone().detach().float().to(device)
else:
inlet_vel_tensor = torch.tensor(inlet_velocities, dtype=torch.float32).to(device)
# --- NORMALIZE FEATURES ---
norm_coords = (coordinates - coords_mean) / (coords_std + epsilon)
norm_current_vel = (current_velocities - vel_mean) / (vel_std + epsilon)
norm_target_vel = (target_velocities - vel_mean) / (vel_std + epsilon)
norm_inlet_vel = (inlet_vel_tensor - inlet_vel_mean) / (inlet_vel_std + epsilon)
batch_size, num_nodes, _ = coordinates.shape
inlet_vel_feature = norm_inlet_vel.view(batch_size, 1, 1).expand(batch_size, num_nodes, 1)
# Input: [normalized coordinates, node_type, normalized velocities, normalized inlet_velocity]
input_features = torch.cat([norm_coords, node_type, norm_current_vel, inlet_vel_feature], dim=-1)
optimizer.zero_grad()
predicted_normalized_velocities = model(input_features)
# Apply wall masking
mask = create_wall_mask(node_type)
loss = criterion(predicted_normalized_velocities * mask, norm_target_vel * mask)
loss.backward()
optimizer.step()
total_train_loss += loss.item()
# Progress update every 50 batches
if batch_idx % 50 == 0:
print(f"Epoch {epoch+1}/{hparams['nb_epochs']} | Batch {batch_idx+1}/{len(train_loader)} | Loss: {loss.item():.6f}")
scheduler.step()
lr_history.append(optimizer.param_groups[0]['lr'])
epoch_time = time.time() - epoch_start_time
avg_train_loss = total_train_loss / len(train_loader)
train_loss_history.append(avg_train_loss)
print(f"Epoch {epoch+1}/{hparams['nb_epochs']} completed | Avg Train Loss: {avg_train_loss:.6f} | Time: {epoch_time:.2f}s")
# --- PERIODIC EVALUATION ---
if (epoch + 1) % hparams['eval_freq'] == 0 or epoch == hparams['nb_epochs'] - 1:
print(f"Epoch {epoch+1}/{hparams['nb_epochs']} | Train Loss: {avg_train_loss:.6f}")
# Rollout evaluation
print(" Running rollout evaluation...")
rollout_results = evaluate_rollout(model, test_trajectory_loader, device, norm_stats)
# Calculate loss from rollout
total_rollout_mse = 0.0
for result in rollout_results:
# Move tensors to the correct device
pred_traj = result['predicted_trajectory'].to(device)
gt_traj = result['ground_truth_trajectory'].to(device)
node_type = result['node_type'].to(device)
# Create a mask to ignore wall nodes
mask = create_wall_mask(node_type).unsqueeze(0) # Add time dim for broadcasting
# Calculate masked MSE
loss = criterion(pred_traj * mask, gt_traj * mask)
total_rollout_mse += loss.item()
avg_rollout_mse = total_rollout_mse / len(rollout_results)
val_loss = avg_rollout_mse
val_loss_history.append(val_loss)
val_epochs.append(epoch + 1)
print(f" Validation Rollout MSE: {val_loss:.6f}")
# Save rollout results
rollout_path = os.path.join(output_dir, f"rollout_results_epoch_{epoch+1}.npy")
np.save(rollout_path, rollout_results, allow_pickle=True) # allow_pickle is good practice for np.save with dicts
print(f" -> Saved rollout results to {rollout_path}")
# Save best model based on rollout loss
if val_loss < best_val_loss:
best_val_loss = val_loss
best_model_path = os.path.join(output_dir, 'best_model.pth')
torch.save(model.state_dict(), best_model_path)
print(f" -> New best model saved with validation loss: {best_val_loss:.6f}")
# --- PERIODIC CHECKPOINTING ---
if (epoch + 1) % hparams['checkpoint_freq'] == 0:
# Use validation loss if available, otherwise use training loss
checkpoint_loss = val_loss if 'val_loss' in locals() else avg_train_loss
checkpoint_path = os.path.join(output_dir, f'checkpoint_epoch_{epoch+1}.pth')
save_checkpoint(model, optimizer, scheduler, epoch + 1, checkpoint_loss, checkpoint_path, norm_stats)
# --- SAVE LATEST CHECKPOINT ---
save_checkpoint(model, optimizer, scheduler, hparams['nb_epochs'], best_val_loss,
os.path.join(output_dir, 'checkpoint_latest.pth'), norm_stats)
print("="*60)
print("Training complete.")
print(f"Best validation loss: {best_val_loss:.6f}")
# --- PLOTTING AND SAVING LOSS HISTORY ---
plt.figure(figsize=(10, 5))
plt.plot(train_loss_history, label='Training Loss')
plt.plot(val_epochs, val_loss_history, 'o-', label='Validation Loss')
plt.title('Training and Validation Loss Over Epochs')
plt.xlabel('Epoch')
plt.ylabel('Loss (MSE)')
plt.legend()
plt.grid(True)
plt.savefig(os.path.join(output_dir, 'training_loss_plot.png'))
print(f"Loss plot saved to {os.path.join(output_dir, 'training_loss_plot.png')}")
loss_data = {
'train_loss_history': train_loss_history,
'val_loss_history': val_loss_history,
'val_epochs': val_epochs,
'lr_history': lr_history,
'best_val_loss': best_val_loss
}
with open(os.path.join(output_dir, 'loss_history.json'), 'w') as f:
json.dump(loss_data, f, indent=4)
print(f"Loss history saved to {os.path.join(output_dir, 'loss_history.json')}")
script_end_time = time.time()
elapsed_time_seconds = script_end_time - script_start_time
print("="*60)
print(f"Total script execution time: {elapsed_time_seconds / 60:.2f} minutes ({elapsed_time_seconds:.2f} seconds)")
if __name__ == '__main__':
main()