import os
import torch
import numpy as np
import matplotlib.pyplot as plt

# 读取 .pt 文件（与脚本同目录）
script_dir = os.path.dirname(os.path.abspath(__file__))
file_path = os.path.join(script_dir, "token_7.pt")
try:
    data = torch.load(file_path, map_location=torch.device('cpu'))[0]
except FileNotFoundError:
    raise FileNotFoundError(f"File {file_path} not found. Please check the path.")
data = data[:, :4000]

# 处理概率数据
if data.dim() == 1:
    probs = data.numpy()
elif data.dim() == 2:
    probs = data.mean(dim=0).numpy()
else:
    raise ValueError(f"Unexpected tensor shape: {data.shape}")

if abs(probs.sum() - 1.0) > 1e-5:
    probs = probs / probs.sum()

labels = np.arange(len(probs))

# 绘制柱状图
plt.rcParams['font.weight'] = 'bold'
plt.figure(figsize=(10, 6))
plt.bar(labels, probs, color='skyblue', edgecolor='black')
plt.xlabel('Text Token Index', fontsize=16, fontweight='bold')
plt.ylabel('Probability', fontsize=16, fontweight='bold')
plt.ylim(0, 0.1)
plt.grid(True, axis='y', linestyle='--', alpha=0.7)
ax = plt.gca()
for label in ax.get_xticklabels() + ax.get_yticklabels():
    label.set_fontweight('bold')
plt.tight_layout()
save_path = os.path.join(script_dir, 'MLLM_P.png')
plt.savefig(save_path, dpi=300, bbox_inches='tight')
print(f"已保存: {save_path}")
plt.show()