文章目录
概要
通过以下实例，你将学会如何查看神经网络结构并打印出训练参数。
流程
- 准备一个简易的二分类数据集，并编写一个单层的神经网络
train_data = np.array([[1, 2, 3, 4, 5],
[7, 7, 2, 4, 10],
[1, 9, 3, 6, 5],
[6, 7, 8, 9, 10]])
train_label = np.array([1, 0, 1, 0]) #标签与样本一一对齐
""" 定义一个单层的神经网络 """
model = tf.keras.models.Sequential([
tf.keras.layers.Dense(1, activation=None)
])- 编译，训练，并保存模型
model.compile(
loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
optimizer='adam'
)
model.fit(train_data,
train_label,
epochs=2750)
tf.saved_model.save(model, "model_dir") #保存到当前目录中，目录名为model_dir- 模型保存形式
- 利用模型的Checkpoint对象查看模型结构和参数
""" 最后的variables是.data-00000-of-00001和 .index文件去掉后缀后的表达形式,
从而统一代表着这两个文件"""
save_path = './model_dir/variables/variables' #
reader = tf.train.load_checkpoint(save_path) # 得到CheckpointReader
""" 打印Checkpoint中存储的所有参数名和参数shape """
for variable_name, variable_shape in reader.get_variable_to_shape_map().items():
print(f'{variable_name} : {variable_shape}')
optimizer/_variables/2/.ATTRIBUTES/VARIABLE_VALUE : [5, 1]
optimizer/_iterations/.ATTRIBUTES/VARIABLE_VALUE : []
_CHECKPOINTABLE_OBJECT_GRAPH : []
keras_api/metrics/0/count/.ATTRIBUTES/VARIABLE_VALUE : []
keras_api/metrics/0/total/.ATTRIBUTES/VARIABLE_VALUE : []
layer_with_weights-0/bias/.ATTRIBUTES/VARIABLE_VALUE : [1]
layer_with_weights-0/kernel/.ATTRIBUTES/VARIABLE_VALUE : [5, 1]
optimizer/_variables/1/.ATTRIBUTES/VARIABLE_VALUE : [5, 1]
optimizer/_learning_rate/.ATTRIBUTES/VARIABLE_VALUE : []
optimizer/_variables/3/.ATTRIBUTES/VARIABLE_VALUE : [1]
optimizer/_variables/4/.ATTRIBUTES/VARIABLE_VALUE : [1]print(reader.get_tensor('layer_with_weights-0/kernel/.ATTRIBUTES/VARIABLE_VALUE'))
print(reader.get_tensor("layer_with_weights-0/bias/.ATTRIBUTES/VARIABLE_VALUE"))
[[-1.7741445 ]
[-0.07314294]
[-0.07213379]
[ 1.1694099 ]
[-0.36803177]]
[1.7487208]- 验证
model = tf.saved_model.load('model_dir')
print(model([[1, 2, 3, 4, 5]]))
output = -1.7741445 - 2*0.07314294 - 3*0.07213379 + 4*1.1694099 - 5*0.36803177+1.7487208
print(output)
tf.Tensor([[2.4493697]], shape=(1, 1), dtype=float32)
2.4493698000000004
