Neural Network Architecture Loss Raincloud

import numpy as np
import matplotlib.pyplot as plt
import scipy.stats as stats

np.random.seed(123)

# Final training loss
cnn = np.random.lognormal(-2.5, 0.4, 100)
transformer = np.random.lognormal(-3.2, 0.35, 100)
lstm = np.random.lognormal(-2.8, 0.45, 100)
mlp = np.random.lognormal(-2.0, 0.5, 100)

cnn = np.clip(cnn, 0.01, 0.3)
transformer = np.clip(transformer, 0.005, 0.15)
lstm = np.clip(lstm, 0.01, 0.25)
mlp = np.clip(mlp, 0.02, 0.5)

F_stat, p_value = stats.f_oneway(cnn, transformer, lstm, mlp)

BG_COLOR = "#0d1117"
COLOR_SCALE = ["#F5B027", "#6CF527", "#27D3F5", "#F5276C"]

fig, ax = plt.subplots(figsize=(10, 6), facecolor=BG_COLOR)
ax.set_facecolor(BG_COLOR)

y_data = [cnn, transformer, lstm, mlp]
positions = [0, 1, 2, 3]
labels = ["CNN", "Transformer", "LSTM", "MLP"]

for h in [0.05, 0.1, 0.15, 0.2]:
    ax.axhline(h, color='#333333', ls=(0, (5, 5)), alpha=0.6, zorder=0)

violins = ax.violinplot(y_data, positions=positions, widths=0.5, 
                         bw_method="silverman", showmeans=False, 
                         showmedians=False, showextrema=False)
for pc in violins["bodies"]:
    pc.set_facecolor("none")
    pc.set_edgecolor("white")
    pc.set_linewidth(1.8)

bp = ax.boxplot(y_data, positions=positions, showfliers=False, showcaps=False,
                medianprops=dict(linewidth=3, color='#F5D327'),
                whiskerprops=dict(linewidth=2, color='#666666'),
                boxprops=dict(linewidth=2, color='#666666'))

for i, (y, color) in enumerate(zip(y_data, COLOR_SCALE)):
    x_jitter = np.array([i] * len(y)) + stats.t(df=6, scale=0.04).rvs(len(y))
    ax.scatter(x_jitter, y, s=50, color=color, alpha=0.5, zorder=2)

means = [y.mean() for y in y_data]
for i, (mean, color) in enumerate(zip(means, COLOR_SCALE)):
    ax.scatter(i, mean, s=180, color='#C82909', zorder=5, edgecolors='white', linewidths=2)
    ax.plot([i, i + 0.28], [mean, mean], ls="dashdot", color="white", zorder=3, lw=1.5)
    ax.text(i + 0.3, mean, f"μ={mean:.3f}", fontsize=10, va="center", color='white',
            bbox=dict(facecolor=BG_COLOR, edgecolor=color, boxstyle="round,pad=0.15", lw=2))

# Best model bracket
ax.annotate('', xy=(1, transformer.mean()-0.01), xytext=(1, 0.005),
            arrowprops=dict(arrowstyle='->', color='#6CF527', lw=2))
ax.text(1, 0.002, "Best", ha='center', color='#6CF527', fontsize=10, fontweight='bold')

ax.spines["right"].set_visible(False)
ax.spines["top"].set_visible(False)
ax.spines["left"].set_color('#444444')
ax.spines["bottom"].set_color('#444444')
ax.tick_params(colors='#888888', length=0)

xlabels = [f"{l}\n(n={len(y_data[i])})" for i, l in enumerate(labels)]
ax.set_xticks(positions)
ax.set_xticklabels(xlabels, size=12, color='white')
ax.set_ylabel("Final Loss (Cross-Entropy)", size=14, color='white', fontweight='bold')

ax.set_title("Neural Network Architecture Comparison", fontsize=14, color="white", fontweight="bold", pad=20)


ax.set_ylim(-0.01, 0.35)
plt.tight_layout()
plt.show()