import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense

# Create an 8x3x8 neural network
model = Sequential([
    Dense(8, activation='relu', input_dim=8), # Input layer with 8 neurons
    Dense(3, activation='relu'), # Hidden layer with 3 neurons
    Dense(8, activation='sigmoid') # Output layer with 8 neurons
])

# Display the model summary
model.summary()

# 展示神经网络的形状结构
%pip install pydot graphviz

# Install the graphviz system package
import sys
import os

# Install graphviz based on the operating system
if sys.platform.startswith('darwin'):  # macOS
	!brew install graphviz || echo "Homebrew not available or graphviz already installed"
elif sys.platform.startswith('linux'):
	!apt-get update && apt-get install -y graphviz || echo "apt not available or graphviz already installed"
	# Alternative for other Linux distributions
	!yum install -y graphviz || echo "yum not available or graphviz already installed"

from IPython.display import SVG
from tensorflow.keras.utils import model_to_dot
dot = model_to_dot(model, show_shapes=True)
dot.set('dpi', '60')
print(dot)
SVG(dot.create(prog='dot', format='svg'))

# Compile the model
model.compile(optimizer='adam',
              loss='mse',
              metrics=['accuracy'])

import numpy as np

# 创建训练数据 - 每个输入是一个8维的one-hot向量
X_train = np.array([
    [1, 0, 0, 0, 0, 0, 0, 0],
    [0, 1, 0, 0, 0, 0, 0, 0],
    [0, 0, 1, 0, 0, 0, 0, 0],
    [0, 0, 0, 1, 0, 0, 0, 0],
    [0, 0, 0, 0, 1, 0, 0, 0],
    [0, 0, 0, 0, 0, 1, 0, 0],
    [0, 0, 0, 0, 0, 0, 1, 0],
    [0, 0, 0, 0, 0, 0, 0, 1]
])

# 输出数据与输入相同（恒等函数）
y_train = X_train.copy()

# 训练模型
history = model.fit(X_train, y_train, epochs=5000, verbose=0)

# 评估模型性能
loss, accuracy = model.evaluate(X_train, y_train, verbose=0)
print(f"最终损失：{loss:.4f}， 准确率：{accuracy:.4f}")

# 测试模型
predictions = model.predict(X_train)
print("\n预测结果：")
for i in range(len(X_train)):
    print(f"输入：{X_train[i]} -> 预测输出：{predictions[i]}，期望： {y_train[i]}")
    print("-" * 50)