File size: 10,257 Bytes
cbc5e69 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 | """
Loss logging and visualization utilities for SLTUNET training.
Automatically saves training loss and generates plots.
"""
import os
import csv
from datetime import datetime
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import numpy as np
class LossLogger:
"""Logger for training loss with automatic visualization."""
def __init__(self, output_dir, plot_freq=10):
"""
Args:
output_dir: Directory to save loss logs and plots
plot_freq: Frequency to update plots (in steps)
"""
self.output_dir = output_dir
self.plot_freq = plot_freq
# CSV file for training loss
self.train_loss_file = os.path.join(output_dir, 'train_loss.csv')
self.eval_loss_file = os.path.join(output_dir, 'eval_loss.csv')
# Initialize CSV files
self._init_csv_files()
# Cache for plotting
self.train_losses = []
self.train_steps = []
self.eval_losses = []
self.eval_steps = []
self.eval_bleus = []
def _init_csv_files(self):
"""Initialize CSV files with headers."""
# Training loss CSV
if not os.path.exists(self.train_loss_file):
with open(self.train_loss_file, 'w', newline='') as f:
writer = csv.writer(f)
writer.writerow(['timestamp', 'step', 'epoch', 'loss', 'gnorm', 'pnorm', 'lr'])
# Evaluation loss CSV
if not os.path.exists(self.eval_loss_file):
with open(self.eval_loss_file, 'w', newline='') as f:
writer = csv.writer(f)
writer.writerow(['timestamp', 'step', 'eval_loss', 'bleu4', 'bleu1', 'bleu2', 'bleu3', 'otem2', 'utem4'])
def log_train_step(self, step, epoch, loss, gnorm, pnorm, lr):
"""Log training step information."""
timestamp = datetime.now().strftime('%Y-%m-%d %H:%M:%S')
# Append to CSV
with open(self.train_loss_file, 'a', newline='') as f:
writer = csv.writer(f)
writer.writerow([timestamp, step, epoch, loss, gnorm, pnorm, lr])
# Update cache
self.train_steps.append(step)
self.train_losses.append(loss)
# Generate plot if needed
if step % self.plot_freq == 0:
self.generate_plots()
def log_eval_step(self, step, eval_loss, bleu_score, metrics_dict=None):
"""Log evaluation step information with multiple metrics."""
timestamp = datetime.now().strftime('%Y-%m-%d %H:%M:%S')
# Extract metrics
bleu1 = metrics_dict.get('bleu1', 0.0) if metrics_dict else 0.0
bleu2 = metrics_dict.get('bleu2', 0.0) if metrics_dict else 0.0
bleu3 = metrics_dict.get('bleu3', 0.0) if metrics_dict else 0.0
bleu4 = metrics_dict.get('bleu4', bleu_score) if metrics_dict else bleu_score
otem2 = metrics_dict.get('otem2', 0.0) if metrics_dict else 0.0
utem4 = metrics_dict.get('utem4', 0.0) if metrics_dict else 0.0
# Append to CSV
with open(self.eval_loss_file, 'a', newline='') as f:
writer = csv.writer(f)
writer.writerow([timestamp, step, eval_loss, bleu4, bleu1, bleu2, bleu3, otem2, utem4])
# Update cache
self.eval_steps.append(step)
self.eval_losses.append(eval_loss)
self.eval_bleus.append(bleu4)
# Always generate plot after evaluation
self.generate_plots()
def load_history(self):
"""Load historical data from CSV files."""
# Load training loss
if os.path.exists(self.train_loss_file):
with open(self.train_loss_file, 'r') as f:
reader = csv.DictReader(f)
for row in reader:
self.train_steps.append(int(row['step']))
self.train_losses.append(float(row['loss']))
# Load evaluation loss
if os.path.exists(self.eval_loss_file):
with open(self.eval_loss_file, 'r') as f:
reader = csv.DictReader(f)
for row in reader:
self.eval_steps.append(int(row['step']))
self.eval_losses.append(float(row['eval_loss']))
# Backward compatibility: older CSVs used different BLEU column names
bleu_val = row.get('bleu_score', None)
if bleu_val is None or bleu_val == '':
bleu_val = row.get('bleu4', 0.0)
self.eval_bleus.append(float(bleu_val))
def generate_plots(self):
"""Generate loss and BLEU plots."""
if not self.train_steps:
return
# Create figure with 3 subplots
fig = plt.figure(figsize=(14, 10))
gs = fig.add_gridspec(3, 1, hspace=0.3)
# Plot 1: Training loss
ax1 = fig.add_subplot(gs[0])
if self.train_steps:
ax1.plot(self.train_steps, self.train_losses, 'b-', linewidth=1, alpha=0.6, label='Train Loss')
# Add moving average
if len(self.train_losses) > 20:
window = 20
ma = np.convolve(self.train_losses, np.ones(window)/window, mode='valid')
ma_steps = self.train_steps[window-1:]
ax1.plot(ma_steps, ma, 'r-', linewidth=2, label=f'Moving Avg ({window} steps)')
ax1.set_xlabel('Training Step', fontsize=12, fontweight='bold')
ax1.set_ylabel('Training Loss', fontsize=12, fontweight='bold')
ax1.set_title(f'SLTUNET Training Loss (Current Step: {self.train_steps[-1] if self.train_steps else 0})',
fontsize=14, fontweight='bold')
ax1.grid(True, alpha=0.3, linestyle='--')
ax1.legend(loc='upper right')
# Plot 2: Evaluation loss
ax2 = fig.add_subplot(gs[1])
if self.eval_steps:
ax2.plot(self.eval_steps, self.eval_losses, 'go-', linewidth=2, markersize=6, label='Eval Loss')
# Mark best
if self.eval_losses:
min_loss = min(self.eval_losses)
min_idx = self.eval_losses.index(min_loss)
ax2.plot(self.eval_steps[min_idx], min_loss, 'r*', markersize=15)
ax2.annotate(f'Best: {min_loss:.4f}',
xy=(self.eval_steps[min_idx], min_loss),
xytext=(10, 10), textcoords='offset points',
bbox=dict(boxstyle='round', facecolor='yellow', alpha=0.8),
arrowprops=dict(arrowstyle='->', color='red'))
ax2.set_xlabel('Training Step', fontsize=12, fontweight='bold')
ax2.set_ylabel('Validation Loss', fontsize=12, fontweight='bold')
ax2.set_title('Validation Loss', fontsize=14, fontweight='bold')
ax2.grid(True, alpha=0.3, linestyle='--')
ax2.legend(loc='upper right')
# Plot 3: BLEU score
ax3 = fig.add_subplot(gs[2])
if self.eval_steps:
ax3.plot(self.eval_steps, self.eval_bleus, 'mo-', linewidth=2, markersize=6, label='BLEU Score')
# Mark best
if self.eval_bleus:
max_bleu = max(self.eval_bleus)
max_idx = self.eval_bleus.index(max_bleu)
ax3.plot(self.eval_steps[max_idx], max_bleu, 'r*', markersize=15)
ax3.annotate(f'Best: {max_bleu:.6f}',
xy=(self.eval_steps[max_idx], max_bleu),
xytext=(10, -20), textcoords='offset points',
bbox=dict(boxstyle='round', facecolor='lightgreen', alpha=0.8),
arrowprops=dict(arrowstyle='->', color='green'))
ax3.set_xlabel('Training Step', fontsize=12, fontweight='bold')
ax3.set_ylabel('BLEU Score', fontsize=12, fontweight='bold')
ax3.set_title('BLEU Score (Higher is Better)', fontsize=14, fontweight='bold')
ax3.grid(True, alpha=0.3, linestyle='--')
ax3.legend(loc='lower right')
# Add timestamp
timestamp = datetime.now().strftime('%Y-%m-%d %H:%M:%S')
fig.text(0.99, 0.01, f'Updated: {timestamp}', ha='right', fontsize=9, style='italic')
# Save plot
plot_path = os.path.join(self.output_dir, 'training_curves.png')
plt.savefig(plot_path, dpi=150, bbox_inches='tight')
plt.close()
# Also generate a summary text
self._generate_summary()
def _generate_summary(self):
"""Generate text summary of training."""
summary_path = os.path.join(self.output_dir, 'training_summary.txt')
with open(summary_path, 'w') as f:
f.write("=" * 70 + "\n")
f.write(" SLTUNET Training Summary\n")
f.write(f" Generated: {datetime.now().strftime('%Y-%m-%d %H:%M:%S')}\n")
f.write("=" * 70 + "\n\n")
if self.train_steps:
f.write(f"Training Progress:\n")
f.write(f" Current Step: {self.train_steps[-1]}\n")
f.write(f" Total Steps: {len(self.train_steps)}\n")
f.write(f" Latest Loss: {self.train_losses[-1]:.6f}\n")
if len(self.train_losses) > 10:
recent_avg = np.mean(self.train_losses[-10:])
f.write(f" Recent Avg Loss: {recent_avg:.6f} (last 10 steps)\n")
f.write("\n")
if self.eval_steps:
f.write(f"Evaluation Results:\n")
f.write(f" Total Evaluations: {len(self.eval_steps)}\n")
f.write(f" Best Eval Loss: {min(self.eval_losses):.6f} (step {self.eval_steps[self.eval_losses.index(min(self.eval_losses))]})\n")
f.write(f" Best BLEU Score: {max(self.eval_bleus):.6f} (step {self.eval_steps[self.eval_bleus.index(max(self.eval_bleus))]})\n")
f.write(f" Latest Eval Loss: {self.eval_losses[-1]:.6f}\n")
f.write(f" Latest BLEU: {self.eval_bleus[-1]:.6f}\n")
f.write("\n")
f.write("=" * 70 + "\n")
f.write(f"Loss logs saved to:\n")
f.write(f" - {self.train_loss_file}\n")
f.write(f" - {self.eval_loss_file}\n")
f.write("=" * 70 + "\n")
|