Files
DSAI/image-inpainting/src/train.py

546 lines
22 KiB
Python

"""
Author: Your Name
HTL-Grieskirchen 5. Jahrgang, Schuljahr 2025/26
train.py
"""
import datasets
from architecture import MyModel
from utils import plot, evaluate_model
import torch
import torch.nn as nn
import numpy as np
import os
import json
from torchvision import models
import torch.nn.functional as F
from torch.utils.data import DataLoader
from torch.utils.data import Subset
import wandb
def load_runtime_config(config_path, current_params):
"""Load runtime configuration from JSON file and update parameters"""
try:
if os.path.exists(config_path):
with open(config_path, 'r') as f:
new_config = json.load(f)
# Update modifiable parameters
updated = False
modifiable_keys = ['n_updates', 'plot_at', 'early_stopping_patience',
'print_stats_at', 'print_train_stats_at', 'validate_at',
'learningrate', 'weight_decay']
for key in modifiable_keys:
if key in new_config and new_config[key] != current_params.get(key):
old_val = current_params.get(key)
current_params[key] = new_config[key]
print(f"\n[CONFIG UPDATE] {key}: {old_val} -> {new_config[key]}")
updated = True
# Check for command flags
commands = new_config.get('commands', {})
current_params['commands'] = commands
if updated:
print("[CONFIG UPDATE] Runtime configuration updated successfully!\n")
except Exception as e:
print(f"Warning: Could not load runtime config: {e}")
return current_params
def clear_command_flag(config_path, command_name):
"""Clear a specific command flag after execution"""
try:
if os.path.exists(config_path):
with open(config_path, 'r') as f:
config = json.load(f)
if 'commands' in config and command_name in config['commands']:
config['commands'][command_name] = False
with open(config_path, 'w') as f:
json.dump(config, f, indent=2)
except Exception as e:
print(f"Warning: Could not clear command flag: {e}")
class RMSELoss(nn.Module):
"""RMSE loss for direct optimization of evaluation metric"""
def __init__(self):
super().__init__()
self.mse = nn.MSELoss()
def forward(self, pred, target):
mse = self.mse(pred, target)
# Larger epsilon for numerical stability
rmse = torch.sqrt(mse + 1e-6)
return rmse
class PerceptualLoss(nn.Module):
"""Perceptual loss using VGG16 features for better texture and detail preservation"""
def __init__(self, device):
super().__init__()
# Load pre-trained VGG16 and use specific layers
vgg = models.vgg16(pretrained=True).features.to(device).eval()
# Freeze VGG parameters
for param in vgg.parameters():
param.requires_grad = False
# Use early and middle layers for perceptual loss
self.slice1 = nn.Sequential(*list(vgg.children())[:4]) # relu1_2
self.slice2 = nn.Sequential(*list(vgg.children())[4:9]) # relu2_2
self.slice3 = nn.Sequential(*list(vgg.children())[9:16]) # relu3_3
# Normalization for VGG (ImageNet stats)
self.register_buffer('mean', torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1))
self.register_buffer('std', torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1))
def normalize(self, x):
"""Normalize images for VGG with clamping for stability"""
# Clamp input to valid range
x = torch.clamp(x, 0.0, 1.0)
return (x - self.mean) / (self.std + 1e-8)
def forward(self, pred, target):
# Clamp inputs to prevent extreme values
pred = torch.clamp(pred, 0.0, 1.0)
target = torch.clamp(target, 0.0, 1.0)
# Normalize inputs
pred = self.normalize(pred)
target = self.normalize(target)
# Extract features from multiple layers
pred_f1 = self.slice1(pred)
pred_f2 = self.slice2(pred_f1)
pred_f3 = self.slice3(pred_f2)
target_f1 = self.slice1(target)
target_f2 = self.slice2(target_f1)
target_f3 = self.slice3(target_f2)
# Compute losses at multiple scales
loss = F.l1_loss(pred_f1, target_f1) + \
F.l1_loss(pred_f2, target_f2) + \
F.l1_loss(pred_f3, target_f3)
return loss
class CombinedLoss(nn.Module):
"""Combined loss optimized for RMSE evaluation with optional perceptual component"""
def __init__(self, device, use_perceptual=True, perceptual_weight=0.05):
super().__init__()
self.use_perceptual = use_perceptual
if use_perceptual:
self.perceptual_loss = PerceptualLoss(device)
# Use MSE instead of RMSE for training (more stable gradients)
self.mse_loss = nn.MSELoss()
self.rmse_loss = RMSELoss() # For logging only
self.perceptual_weight = perceptual_weight
self.mse_weight = 1.0 - perceptual_weight
def forward(self, pred, target):
# Clamp predictions to valid range
pred = torch.clamp(pred, 0.0, 1.0)
target = torch.clamp(target, 0.0, 1.0)
# Check for NaN in inputs
if not torch.isfinite(pred).all() or not torch.isfinite(target).all():
print("Warning: NaN detected in loss inputs")
return (torch.tensor(float('nan'), device=pred.device),) * 4
# Primary loss: MSE (equivalent to RMSE but more stable)
mse = self.mse_loss(pred, target)
rmse = self.rmse_loss(pred, target) # For logging
if self.use_perceptual:
# Optional small perceptual component for texture quality
perceptual = self.perceptual_loss(pred, target)
# Check perceptual loss validity
if not torch.isfinite(perceptual):
perceptual = torch.tensor(0.0, device=pred.device)
total_loss = self.mse_weight * mse + self.perceptual_weight * perceptual
else:
# Pure MSE optimization
perceptual = torch.tensor(0.0, device=pred.device)
total_loss = mse
# Validate loss is not NaN or Inf
if not torch.isfinite(total_loss):
# Return MSE only as fallback
total_loss = mse
if not torch.isfinite(total_loss):
print("Warning: MSE is NaN")
return (torch.tensor(float('nan'), device=pred.device),) * 4
return total_loss, perceptual, mse, rmse
def train(seed, testset_ratio, validset_ratio, data_path, results_path, early_stopping_patience, device, learningrate,
weight_decay, n_updates, use_wandb, print_train_stats_at, print_stats_at, plot_at, validate_at, batchsize,
network_config: dict, testset_path=None, save_path=None, plot_path_predictions=None):
np.random.seed(seed=seed)
torch.manual_seed(seed=seed)
if device is None:
device = torch.device(
"cuda:0" if torch.cuda.is_available() else "mps" if torch.backends.mps.is_available() else "cpu")
if isinstance(device, str):
device = torch.device(device)
# Enable mixed precision training for memory efficiency
use_amp = torch.cuda.is_available()
if use_amp:
scaler = torch.amp.GradScaler('cuda', init_scale=2048.0, growth_interval=100)
else:
scaler = None
if use_wandb:
wandb.login()
wandb.init(project="image_inpainting", config={
"learning_rate": learningrate,
"weight_decay": weight_decay,
"n_updates": n_updates,
"batch_size": batchsize,
"validation_ratio": validset_ratio,
"testset_ratio": testset_ratio,
"early_stopping_patience": early_stopping_patience,
})
# Prepare a path to plot to
plotpath = os.path.join(results_path, "plots")
os.makedirs(plotpath, exist_ok=True)
image_dataset = datasets.ImageDataset(datafolder=data_path)
n_total = len(image_dataset)
n_test = int(n_total * testset_ratio)
n_valid = int(n_total * validset_ratio)
n_train = n_total - n_test - n_valid
indices = np.random.permutation(n_total)
dataset_train = Subset(image_dataset, indices=indices[0:n_train])
dataset_valid = Subset(image_dataset, indices=indices[n_train:n_train + n_valid])
dataset_test = Subset(image_dataset, indices=indices[n_train + n_valid:n_total])
assert len(image_dataset) == len(dataset_train) + len(dataset_test) + len(dataset_valid)
del image_dataset
dataloader_train = DataLoader(dataset=dataset_train, batch_size=batchsize,
num_workers=0, shuffle=True)
dataloader_valid = DataLoader(dataset=dataset_valid, batch_size=1,
num_workers=0, shuffle=False)
dataloader_test = DataLoader(dataset=dataset_test, batch_size=1,
num_workers=0, shuffle=False)
# initializing the model
network = MyModel(**network_config)
network.to(device)
network.train()
# defining the loss - Optimized for RMSE evaluation
# Set use_perceptual=False for pure MSE training, or keep True with 5% weight for texture quality
# TEMPORARILY DISABLED due to NaN issues - re-enable once training is stable
combined_loss = CombinedLoss(device, use_perceptual=False, perceptual_weight=0.0).to(device)
mse_loss = torch.nn.MSELoss() # Keep for evaluation
# defining the optimizer with AdamW for better weight decay handling
optimizer = torch.optim.AdamW(network.parameters(), lr=learningrate, weight_decay=weight_decay, betas=(0.9, 0.999))
# Learning rate warmup
warmup_steps = min(1000, n_updates // 10)
# Cosine annealing with warm restarts for long training
scheduler_main = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(
optimizer, T_0=n_updates//4, T_mult=1, eta_min=learningrate/100
)
# Warmup scheduler
def get_lr_scale(step):
if step < warmup_steps:
return step / warmup_steps
return 1.0
if use_wandb:
wandb.watch(network, mse_loss, log="all", log_freq=10)
i = 0
counter = 0
best_validation_loss = np.inf
loss_list = []
saved_model_path = os.path.join(results_path, "best_model.pt")
# Save runtime configuration to JSON file for dynamic updates
config_json_path = os.path.join(results_path, "runtime_config.json")
runtime_params = {
'learningrate': learningrate,
'weight_decay': weight_decay,
'n_updates': n_updates,
'plot_at': plot_at,
'early_stopping_patience': early_stopping_patience,
'print_stats_at': print_stats_at,
'print_train_stats_at': print_train_stats_at,
'validate_at': validate_at,
'commands': {
'save_checkpoint': False,
'run_test_validation': False,
'generate_predictions': False
}
}
with open(config_json_path, 'w') as f:
json.dump(runtime_params, f, indent=2)
print(f"Started training on device {device}")
print(f"Runtime config saved to: {config_json_path}")
print(f"You can modify this file during training to change parameters dynamically!")
print(f"Set command flags to true to trigger actions (save_checkpoint, run_test_validation, generate_predictions)\n")
while i < n_updates:
for input, target in dataloader_train:
input, target = input.to(device), target.to(device)
# Check for runtime config updates every 5 steps
if i % 5 == 0 and i > 0:
runtime_params = load_runtime_config(config_json_path, runtime_params)
n_updates = runtime_params['n_updates']
plot_at = runtime_params['plot_at']
early_stopping_patience = runtime_params['early_stopping_patience']
print_stats_at = runtime_params['print_stats_at']
print_train_stats_at = runtime_params['print_train_stats_at']
validate_at = runtime_params['validate_at']
# Update optimizer parameters if changed
if 'learningrate' in runtime_params:
new_lr = runtime_params['learningrate']
current_lr = optimizer.param_groups[0]['lr']
if abs(new_lr - current_lr) > 1e-10: # Float comparison with tolerance
for param_group in optimizer.param_groups:
param_group['lr'] = new_lr
if 'weight_decay' in runtime_params:
new_wd = runtime_params['weight_decay']
current_wd = optimizer.param_groups[0]['weight_decay']
if abs(new_wd - current_wd) > 1e-10: # Float comparison with tolerance
for param_group in optimizer.param_groups:
param_group['weight_decay'] = new_wd
# Execute runtime commands
commands = runtime_params.get('commands', {})
# Command: Save checkpoint
if commands.get('save_checkpoint', False):
checkpoint_path = os.path.join(results_path, f"checkpoint_step_{i}.pt")
torch.save(network.state_dict(), checkpoint_path)
print(f"\n[COMMAND] Checkpoint saved to: {checkpoint_path}\n")
clear_command_flag(config_json_path, 'save_checkpoint')
# Command: Generate predictions
if commands.get('generate_predictions', False) and testset_path is not None:
print(f"\n[COMMAND] Generating predictions at step {i}...")
try:
from utils import create_predictions
pred_save_path = save_path or os.path.join(results_path, "runtime_predictions", f"step_{i}")
pred_plot_path = plot_path_predictions or os.path.join(results_path, "runtime_predictions", "plots", f"step_{i}")
os.makedirs(pred_plot_path, exist_ok=True)
# Save current state temporarily
temp_state_path = os.path.join(results_path, f"temp_state_step_{i}.pt")
torch.save(network.state_dict(), temp_state_path)
# Generate predictions
create_predictions(network_config, temp_state_path, testset_path, None,
pred_save_path, pred_plot_path, plot_at=20, rmse_value=None)
print(f"[COMMAND] Predictions saved to: {pred_save_path}")
print(f"[COMMAND] Plots saved to: {pred_plot_path}\n")
# Clean up temp file
if os.path.exists(temp_state_path):
os.remove(temp_state_path)
except Exception as e:
print(f"[COMMAND] Error generating predictions: {e}\n")
network.train()
clear_command_flag(config_json_path, 'generate_predictions')
# Command: Run test validation
if commands.get('run_test_validation', False):
print(f"\n[COMMAND] Running test set validation at step {i}...")
network.eval()
test_loss, test_rmse = evaluate_model(network, dataloader_test, mse_loss, device)
print(f"[COMMAND] Test Loss: {test_loss:.6f}, Test RMSE: {test_rmse:.6f}\n")
network.train()
clear_command_flag(config_json_path, 'run_test_validation')
if (i + 1) % print_train_stats_at == 0:
print(f'Update Step {i + 1} of {n_updates}: Current loss: {loss_list[-1]}')
optimizer.zero_grad()
# Mixed precision training for memory efficiency
if use_amp:
with torch.amp.autocast('cuda'):
output = network(input)
total_loss, perceptual, mse, rmse = combined_loss(output, target)
# Check for NaN before backward
if not torch.isfinite(total_loss):
continue
scaler.scale(total_loss).backward()
# Unscale and check gradients
scaler.unscale_(optimizer)
# Check for NaN in gradients
has_nan = False
for name, param in network.named_parameters():
if param.grad is not None:
if not torch.isfinite(param.grad).all():
print(f"NaN gradient detected in {name}")
has_nan = True
break
if has_nan:
print(f"Skipping step {i+1}: NaN gradients detected")
optimizer.zero_grad()
scaler.update()
# Reset scaler if NaN persists
if (i + 1) % 10 == 0:
scaler = torch.amp.GradScaler('cuda', init_scale=2048.0, growth_interval=100)
continue
# More aggressive gradient clipping for stability
grad_norm = torch.nn.utils.clip_grad_norm_(network.parameters(), max_norm=1.0)
# Skip update if gradient norm is too large
if grad_norm > 100.0:
print(f"Skipping step {i+1}: Gradient norm too large: {grad_norm:.2f}")
optimizer.zero_grad()
scaler.update()
continue
scaler.step(optimizer)
scaler.update()
else:
output = network(input)
total_loss, perceptual, mse, rmse = combined_loss(output, target)
# Check for NaN before backward
if not torch.isfinite(total_loss):
print(f"Skipping step {i+1}: NaN or Inf loss detected")
continue
total_loss.backward()
# Check for NaN in gradients
has_nan = False
for name, param in network.named_parameters():
if param.grad is not None and not torch.isfinite(param.grad).all():
print(f"NaN gradient detected in {name}")
has_nan = True
break
if has_nan:
print(f"Skipping step {i+1}: NaN gradients detected")
optimizer.zero_grad()
continue
# More aggressive gradient clipping
grad_norm = torch.nn.utils.clip_grad_norm_(network.parameters(), max_norm=1.0)
if grad_norm > 100.0:
print(f"Skipping step {i+1}: Gradient norm too large: {grad_norm:.2f}")
optimizer.zero_grad()
continue
optimizer.step()
# Apply learning rate scheduling with warmup
lr_scale = get_lr_scale(i)
for param_group in optimizer.param_groups:
param_group['lr'] = learningrate * lr_scale
if i >= warmup_steps:
scheduler_main.step()
loss_list.append(total_loss.item())
# writing the stats to wandb
if use_wandb and (i+1) % print_stats_at == 0:
wandb.log({
"training/loss_total": total_loss.item(),
"training/loss_mse": mse.item(),
"training/loss_rmse": rmse.item(),
"training/loss_perceptual": perceptual.item() if isinstance(perceptual, torch.Tensor) else perceptual,
"training/learning_rate": optimizer.param_groups[0]['lr']
}, step=i)
# plotting
if (i + 1) % plot_at == 0:
print(f"Plotting images, current update {i + 1}")
# Convert to float32 for matplotlib compatibility
plot(input.float().cpu().numpy(),
target.detach().float().cpu().numpy(),
output.detach().float().cpu().numpy(),
plotpath, i)
# evaluating model every validate_at sample
if (i + 1) % validate_at == 0:
print(f"Evaluation of the model:")
val_loss, val_rmse = evaluate_model(network, dataloader_valid, mse_loss, device)
print(f"val_loss: {val_loss}")
print(f"val_RMSE: {val_rmse}")
if use_wandb:
wandb.log({"validation/loss": val_loss,
"validation/RMSE": val_rmse}, step=i)
# wandb histogram
# Save best model for early stopping
if val_loss < best_validation_loss:
best_validation_loss = val_loss
torch.save(network.state_dict(), saved_model_path)
print(f"Saved new best model with val_loss: {best_validation_loss}")
counter = 0
else:
counter += 1
if counter >= early_stopping_patience:
print("Stopped training because of early stopping")
i = n_updates
break
i += 1
if i >= n_updates:
print("Finished training because maximum number of updates reached")
break
print("Evaluating the self-defined testset")
network.load_state_dict(torch.load(saved_model_path))
testset_loss, testset_rmse = evaluate_model(network=network, dataloader=dataloader_test, loss_fn=mse_loss,
device=device)
print(f'testset_loss of model: {testset_loss}, RMSE = {testset_rmse}')
if use_wandb:
wandb.summary["testset/loss"] = testset_loss
wandb.summary["testset/RMSE"] = testset_rmse
wandb.finish()
return testset_rmse