要求动手从0实现 softmax 回归（只借助Tensor和Numpy相关的库）在Fashion-MNIST数据 集 上进行训练和测试，并从loss、训练集以及测试集上的准确率等多个角度对结果进行分析；给出代码

为了在Fashion-MNIST数据集上使用手动实现的softmax回归，你需要完成以下步骤：

1. **导入所需库**：

   import numpy as np
   from sklearn.datasets import fetch_openml
   from sklearn.model_selection import train_test_split
   from sklearn.metrics import accuracy_score

2. **加载数据**：

   data = fetch_openml('Fashion_MNIST', version=1)
   X = data.data.reshape(-1, 784) / 255.0  # 归一化像素值
   y = data.target

3. **分割数据集**：

   X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

4. **定义权重初始化**和**softmax 函数**：

   def init_weights(n_features, n_classes):
       return np.random.randn(n_classes, n_features)

   def softmax(x):
       e_x = np.exp(x - np.max(x))
       return e_x / e_x.sum(axis=1, keepdims=True)

5. **实现前向传播和损失函数**：

   def forward_pass(weights, inputs, targets):
       scores = np.dot(inputs, weights.T)
       predictions = softmax(scores)
       loss = -np.mean(np.log(predictions[range(len(targets)), targets]))
       return predictions, loss

   def update_weights(loss, learning_rate, weights, inputs, targets):
       dW = np.zeros_like(weights)
       for i in range(len(inputs)):
           dW += (predictions[i] - targets[i]) * inputs[i].reshape(1, -1)
       weights -= learning_rate * dW / len(inputs)
       return weights

6. **训练和测试模型**：

   num_epochs = 10
   learning_rate = 0.01
   weights = init_weights(X.shape[1], 10)  # 假设 Fashion-MNIST 有 10 类

   for epoch in range(num_epochs):
       predictions, loss = forward_pass(weights, X_train, y_train)
       weights = update_weights(loss, learning_rate, weights, X_train, y_train)

   # 测试模型
   test_predictions, _ = forward_pass(weights, X_test, y_test)
   test_accuracy = accuracy_score(y_test, np.argmax(test_predictions, axis=1))

7. **分析结果**：
- 记录每次迭代的训练损失，观察是否收敛。
- 输出训练集和测试集的准确率，比较两者，看模型泛化能力。