| """Plot the actual epoch-average training losses recorded by TensorBoard.""" |
| import csv |
| import json |
| from pathlib import Path |
|
|
| import matplotlib |
| matplotlib.use('Agg') |
| import matplotlib.pyplot as plt |
| from matplotlib.ticker import MultipleLocator, FormatStrFormatter |
| from tensorboard.backend.event_processing.event_accumulator import EventAccumulator |
|
|
| ROOT = Path('/data/realstyle2') |
| OUT = ROOT / 'plots' |
| OUT.mkdir(exist_ok=True) |
| runs = [] |
| for path in sorted((ROOT / 'logs/tensorboard').rglob('events*')): |
| events = EventAccumulator(str(path), size_guidance={'scalars': 0}) |
| events.Reload() |
| if 'loss/epoch_average' in events.Tags().get('scalars', []): |
| runs.append((path, events.Scalars('loss/epoch_average'))) |
| assert len(runs) == 1, 'Expected one completed training run' |
| source, values = runs[0] |
| epochs = [v.step for v in values] |
| losses = [v.value for v in values] |
| assert epochs == list(range(1, 41)) |
| with (OUT / 'mean_loss.csv').open('w', newline='') as f: |
| writer = csv.writer(f) |
| writer.writerow(['epoch', 'mean_loss']) |
| writer.writerows(zip(epochs, losses)) |
| minimum = min(range(len(losses)), key=losses.__getitem__) |
| summary = {'source': str(source), 'scalar': 'loss/epoch_average', |
| 'epochs': len(values), 'first_epoch_loss': losses[0], |
| 'final_epoch_loss': losses[-1], 'minimum_loss_epoch': epochs[minimum], |
| 'minimum_epoch_loss': losses[minimum], 'mean_of_epoch_losses': sum(losses)/len(losses)} |
| (OUT / 'loss_summary.json').write_text(json.dumps(summary, indent=2)) |
|
|
| plt.rcParams.update({'font.family': 'DejaVu Sans', 'font.size': 12, |
| 'axes.spines.top': False, 'axes.spines.right': False, |
| 'axes.labelcolor': '#334155', 'xtick.color': '#64748b', |
| 'ytick.color': '#64748b', 'svg.fonttype': 'none'}) |
| fig, ax = plt.subplots(figsize=(12, 5.9), facecolor='white') |
| fig.subplots_adjust(left=.095, right=.975, top=.77, bottom=.17) |
| fig.text(.095, .92, 'REALSTYLE2 / ANIMA', fontsize=11, weight='bold', color='#64748b') |
| fig.text(.095, .852, 'Mean training loss by epoch', fontsize=24, weight='bold', color='#0f172a') |
| fig.text(.975, .86, '40 epochs 路 4,320 steps', ha='right', fontsize=12, color='#475569') |
| ax.plot(epochs, losses, color='#2563eb', linewidth=2, marker='o', markersize=4.5, |
| markerfacecolor='white', markeredgewidth=1.4, label='Epoch mean loss') |
| ax.scatter([epochs[minimum]], [losses[minimum]], color='#0f766e', s=65, zorder=5) |
| ax.annotate(f'Min: epoch {epochs[minimum]}\n{losses[minimum]:.5f}', |
| xy=(epochs[minimum], losses[minimum]), xytext=(-48, -36), |
| textcoords='offset points', fontsize=11, color='#0f766e', ha='right', |
| arrowprops={'arrowstyle': '-', 'color': '#0f766e', 'lw': 1}) |
| ax.annotate(f'Final: {losses[-1]:.5f}', xy=(epochs[-1], losses[-1]), xytext=(-8, 15), |
| textcoords='offset points', ha='right', fontsize=11, color='#1d4ed8') |
| ax.set(xlim=(.5, 41), ylim=(min(losses)-.0038, max(losses)+.0018), |
| xlabel='Epoch', ylabel='Mean loss') |
| ax.set_xticks([1, 5, 10, 15, 20, 25, 30, 35, 40]) |
| ax.yaxis.set_major_locator(MultipleLocator(.002)) |
| ax.yaxis.set_major_formatter(FormatStrFormatter('%.3f')) |
| ax.grid(axis='y', color='#e2e8f0', linewidth=.8) |
| ax.set_axisbelow(True) |
| ax.spines['left'].set_color('#cbd5e1') |
| ax.spines['bottom'].set_color('#cbd5e1') |
| fig.text(.095, .045, 'Source: TensorBoard loss/epoch_average 路 Raw epoch means, no smoothing 路 Y-axis does not start at zero', |
| fontsize=10, color='#64748b') |
| fig.savefig(OUT / 'mean_loss.png', dpi=200, facecolor='white') |
| fig.savefig(OUT / 'mean_loss.svg', facecolor='white') |
| plt.close(fig) |
| print(json.dumps(summary, indent=2)) |
|
|
