在TensorFlow 1.15的自定义训练循环中将张量转换为numpy数组

import tensorflow as tf
import numpy as np

# 定义模型
model = ...

# 定义优化器
optimizer = ...

# 定义损失函数
loss_fn = ...

# 定义评估指标
metrics = ...

# 定义训练数据集
train_dataset = ...

# 定义训练步数
num_epochs = ...

# 自定义训练循环
for epoch in range(num_epochs):
    for batch, (x, y) in enumerate(train_dataset):
        with tf.GradientTape() as tape:
            # 前向传播
            logits = model(x, training=True)
            # 计算损失
            loss_value = loss_fn(y, logits)
        
        # 计算梯度
        grads = tape.gradient(loss_value, model.trainable_variables)
        # 更新模型参数
        optimizer.apply_gradients(zip(grads, model.trainable_variables))
        
        # 更新评估指标
        metrics.update_state(y, logits)
        
        # 打印训练进度
        if batch % 100 == 0:
            print('Epoch {} Batch {} Loss {:.4f}'.format(epoch, batch, loss_value))
    
    # 打印每个epoch的评估指标
    print('Epoch {} Loss {:.4f} Accuracy {:.4f}'.format(epoch, loss_value, metrics.result()))
    # 重置评估指标
    metrics.reset_states()

numpy_array = tensor.numpy()