188 lines
7.4 KiB
Python
188 lines
7.4 KiB
Python
import sys
|
|
sys.path.append('./src')
|
|
from omegaconf import OmegaConf ### for yaml config parsing
|
|
import torch
|
|
import numpy as np
|
|
import models
|
|
from datasets import *
|
|
import torch.optim as optim
|
|
from tqdm import tqdm
|
|
from torchinfo import summary
|
|
from pathlib import Path
|
|
|
|
from models import get_model_class
|
|
from datasets import singlePhotonDataset
|
|
|
|
### random seed for reproducibility
|
|
torch.manual_seed(0)
|
|
torch.cuda.manual_seed(0)
|
|
np.random.seed(0)
|
|
torch.backends.cudnn.deterministic = True
|
|
torch.backends.cudnn.benchmark = False
|
|
|
|
conf = OmegaConf.load("Configs/train_1photon.yaml")
|
|
|
|
def prepare_output_folder(conf):
|
|
from datetime import datetime
|
|
date = datetime.now().strftime("%y%m%d") ## YYMMDD format
|
|
# find the next index for experiment name
|
|
exp_index = 0
|
|
while True:
|
|
exp_name = f'{date}_1ph_{conf.data.energy}keV_v{conf.model.version}_{exp_index:02d}'
|
|
if not Path(f'Results/{exp_name}').exists():
|
|
break
|
|
exp_index += 1
|
|
Path(f'Results/{exp_name}').mkdir(parents=True, exist_ok=True)
|
|
Path(f'Results/{exp_name}/Models').mkdir(parents=True, exist_ok=True)
|
|
Path(f'Results/{exp_name}/Plots').mkdir(parents=True, exist_ok=True)
|
|
OmegaConf.save(conf, f'Results/{exp_name}/config.yaml')
|
|
return exp_name
|
|
|
|
def get_loss_function(conf):
|
|
if conf.loss.type == "weightedL1":
|
|
alpha = conf.loss.alpha
|
|
beta = conf.loss.beta
|
|
def weighted_loss(pred, target):
|
|
# weighted L1 loss for x,y position
|
|
pred = pred[:, :2]
|
|
target = target[:, :2]
|
|
|
|
direction_weight = 1.0 + alpha * torch.abs(target) # (B, 2)
|
|
r = torch.norm(target, dim=1, keepdim=True)
|
|
radial_weight = 1.0 + beta * r # (B, 1) →
|
|
weights = radial_weight * direction_weight
|
|
|
|
loss = weights * torch.abs(pred - target)
|
|
return loss.mean()
|
|
return weighted_loss
|
|
|
|
def train(model, trainLoader, optimizer, loss_fn):
|
|
model.train()
|
|
batchLoss = 0
|
|
total_samples = 0
|
|
rms_x, rms_y = 0, 0
|
|
for batch_idx, (sample, label) in enumerate(trainLoader):
|
|
sample, label = sample.cuda(), label.cuda()
|
|
x, y, z, e = label[:,0], label[:,1], label[:,2], label[:,3]
|
|
optimizer.zero_grad()
|
|
output = model(sample)
|
|
loss = loss_fn(output, torch.stack((x, y, z), axis=1))
|
|
loss.backward()
|
|
optimizer.step()
|
|
batchLoss += loss.item() * sample.shape[0]
|
|
total_samples += sample.shape[0]
|
|
rms_x += torch.sum((output[:,0] - x)**2).item()
|
|
rms_y += torch.sum((output[:,1] - y)**2).item()
|
|
avgLoss = batchLoss / total_samples
|
|
rms_x = np.sqrt(rms_x / total_samples)
|
|
rms_y = np.sqrt(rms_y / total_samples)
|
|
|
|
datasetName = trainLoader.dataset.datasetName
|
|
print(f"[{datasetName}]\t Average Loss: {avgLoss:.6f} (sigma = {np.sqrt(avgLoss):.6f}) \t RMS X: {rms_x:.6f} \t RMS Y: {rms_y:.6f}")
|
|
return avgLoss
|
|
|
|
def evaluate(model, testLoader, loss_fn):
|
|
model.eval()
|
|
batchLoss = 0
|
|
rms_x, rms_y = 0, 0
|
|
with torch.no_grad():
|
|
for batch_idx, (sample, label) in enumerate(testLoader):
|
|
sample, label = sample.cuda(), label.cuda()
|
|
x, y, z, e = label[:,0], label[:,1], label[:,2], label[:,3]
|
|
output = model(sample)
|
|
loss = loss_fn(output, torch.stack((x, y, z), axis=1))
|
|
batchLoss += loss.item() * sample.shape[0]
|
|
rms_x += torch.sum((output[:,0] - x)**2).item()
|
|
rms_y += torch.sum((output[:,1] - y)**2).item()
|
|
avgLoss = batchLoss / len(testLoader.dataset)
|
|
rms_x = np.sqrt(rms_x / len(testLoader.dataset))
|
|
rms_y = np.sqrt(rms_y / len(testLoader.dataset))
|
|
|
|
datasetName = testLoader.dataset.datasetName
|
|
print(f"[{datasetName}]\t Average Loss: {avgLoss:.6f} (sigma = {np.sqrt(avgLoss):.6f}) \t RMS X: {rms_x:.6f} \t RMS Y: {rms_y:.6f}")
|
|
return avgLoss
|
|
|
|
def get_dataloaders(conf):
|
|
"""construct all dataLoaders"""
|
|
datasets = {}
|
|
loaders = {}
|
|
splits = ['train', 'val', 'test']
|
|
keys = ['train_files', 'val_files', 'test_files']
|
|
batch_keys = ['batch_size_train', 'batch_size_val', 'batch_size_test']
|
|
file_range_keys = ['train_file_range', 'val_file_range', 'test_file_range']
|
|
|
|
for split, f_key, b_key, fr_key in zip(splits, keys, batch_keys, file_range_keys):
|
|
file_paths = [f"{conf.data.sample_folder}/{conf.data.energy}keV_Moench040_150V_{i}.npz" for i in range(conf.data[fr_key][0], conf.data[fr_key][1] + 1)]
|
|
|
|
dataset = singlePhotonDataset(
|
|
file_paths,
|
|
sampleRatio=1.0,
|
|
datasetName=split.capitalize(),
|
|
noiseKeV=conf.data.noise_keV,
|
|
numberOfAugOps=conf.data.num_aug_ops,
|
|
normalize=conf.data.normalize,
|
|
noiseThreshold=conf.data.noise_threshold * conf.data.noise_keV
|
|
)
|
|
datasets[split] = dataset
|
|
|
|
loaders[split] = torch.utils.data.DataLoader(
|
|
dataset,
|
|
batch_size=conf.data[b_key],
|
|
shuffle=(split == 'train'),
|
|
num_workers=conf.data.num_workers,
|
|
pin_memory=True
|
|
)
|
|
return loaders['train'], loaders['val'], loaders['test']
|
|
|
|
def plot_loss_curve(train_losses, val_losses, test_loss, exp_name, conf):
|
|
import matplotlib.pyplot as plt
|
|
plt.figure(figsize=(8,6))
|
|
plt.plot(train_losses, label='Train Loss', color='blue')
|
|
plt.plot(val_losses, label='Validation Loss', color='orange')
|
|
if test_loss > 0:
|
|
plt.axhline(y=test_loss, color='green', linestyle='--', label='Test Loss')
|
|
plt.yscale('log')
|
|
plt.xlabel('Epoch')
|
|
plt.ylabel('MSE Loss')
|
|
plt.legend()
|
|
plt.grid()
|
|
plotName = f'loss_curve_singlePhoton_{conf.model.version}'
|
|
if conf.data.normalize:
|
|
plotName += '_normalized'
|
|
plt.savefig(f'Results/{exp_name}/Plots/{plotName}.png')
|
|
|
|
def get_model_name(epoch, conf):
|
|
modelName = f'singlePhoton{conf.model.version}_{conf.data.energy}keV_Noise{conf.data.noise_keV}keV_E{epoch}_aug{conf.data.num_aug_ops}'
|
|
if conf.data.normalize:
|
|
modelName += '_normalized'
|
|
return modelName
|
|
|
|
if __name__ == "__main__":
|
|
|
|
exp_name = prepare_output_folder(conf)
|
|
model = models.get_model_class(conf.model.version)().cuda()
|
|
# summary(model, input_size=(128, 1, 3, 3))
|
|
|
|
loss_fn = get_loss_function(conf)
|
|
optimizer = torch.optim.Adam(model.parameters(), lr=conf.training.learning_rate)
|
|
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, 'min', factor=conf.training.scheduler_factor, patience = conf.training.scheduler_patience)
|
|
|
|
trainLoader, valLoader, testLoader = get_dataloaders(conf)
|
|
|
|
train_losses = []
|
|
val_losses = []
|
|
|
|
for epoch in tqdm(range(1, conf.training.epochs + 1)):
|
|
t_loss = train(model, trainLoader, optimizer, loss_fn)
|
|
train_losses.append(t_loss)
|
|
v_loss = evaluate(model, valLoader, loss_fn)
|
|
val_losses.append(v_loss)
|
|
scheduler.step(val_losses[-1])
|
|
print(f"Learning Rate: {optimizer.param_groups[0]['lr']}")
|
|
if epoch in conf.training.checkpoint_epochs or epoch == conf.training.epochs:
|
|
modelName = get_model_name(epoch, conf)
|
|
torch.save(model.state_dict(), f'Results/{exp_name}/Models/{modelName}.pth')
|
|
print(f"Saved model checkpoint: {modelName}.pth")
|
|
plot_loss_curve(train_losses, val_losses, test_loss=-1, exp_name=exp_name, conf=conf)
|
|
test_loss = evaluate(model, testLoader, loss_fn)
|
|
plot_loss_curve(train_losses, val_losses, test_loss=test_loss, exp_name=exp_name, conf=conf) |