在 subclassed_model.py 中,通过对 tf.keras.Model 进行子类化,设计了两个自定义模型。

 import tensorflow as tf
tf.enable_eager_execution() # parameters
UNITS = 8 class Encoder(tf.keras.Model):
def __init__(self):
super(Encoder, self).__init__()
self.fc1 = tf.keras.layers.Dense(units=UNITS * 2, activation='relu')
self.fc2 = tf.keras.layers.Dense(units=UNITS, activation='relu') def call(self, inputs):
r = self.fc1(inputs)
return self.fc2(r) class Decoder(tf.keras.Model):
def __init__(self):
super(Decoder, self).__init__()
self.fc = tf.keras.layers.Dense(units=1) def call(self, inputs):
return self.fc(inputs)

在 save_subclassed_model.py 中,创建了 5000 组训练数据集,实例化 Encoder()、Decoder() 模型,优化器采用 tf.train.AdamOptimizer(),以均方误差作为 Loss 函数。训练过程中,每 5 个 epoch 保存一次模型。

 from subclassed_model import *

 import numpy as np
import matplotlib.pyplot as plt
import os import tensorflow as tf
tf.enable_eager_execution() # create training data
X = np.linspace(-1, 1, 5000)
np.random.shuffle(X)
y = X ** 3 + 1 + np.random.normal(0, 0.05, (5000,)) # plot data
plt.scatter(X, y)
plt.show() # training dataset
BATCH_SIZE = 16
BUFFER_SIZE = 128 training_dataset = tf.data.Dataset.from_tensor_slices((X, y)).batch(BATCH_SIZE).shuffle(BUFFER_SIZE) # initialize subclassed models
encoder = Encoder()
decoder = Decoder() optimizer = tf.train.AdamOptimizer() # loss function
def loss_function(real, pred):
return tf.losses.mean_squared_error(labels=real, predictions=pred) # training
EPOCHS = 15 # checkpoint
checkpoint_dir = './training_checkpoints'
checkpoint_prefix = os.path.join(checkpoint_dir, 'ckpt')
checkpoint = tf.train.Checkpoint(optimizer=optimizer,
encoder=encoder,
decoder=decoder)
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir) for epoch in range(EPOCHS):
epoch_loss = 0 for (batch, (x, y)) in enumerate(training_dataset):
x = tf.cast(x, tf.float32)
y = tf.cast(y, tf.float32)
x = tf.expand_dims(x, axis=1) # tf.Tensor([...], shape=(16, 1), dtype=float32)
y = tf.expand_dims(y, axis=1) # tf.Tensor([...], shape=(16, 1), dtype=float32) with tf.GradientTape() as tape:
y_ = encoder(x)
prediction = decoder(y_)
batch_loss = loss_function(real=y, pred=prediction) grads = tape.gradient(batch_loss, encoder.variables + decoder.variables)
optimizer.apply_gradients(zip(grads, encoder.variables + decoder.variables),
global_step=tf.train.get_or_create_global_step()) epoch_loss += batch_loss if (batch + 1) % 100 == 0:
print('Epoch {} Batch {} Loss {:.4f}'.format(epoch + 1,
batch + 1,
batch_loss.numpy())) print('Epoch {} Loss {:.4f}'.format(epoch + 1, epoch_loss / len(X))) if (epoch + 1) % 5 == 0:
checkpoint.save(file_prefix=checkpoint_prefix)

运行 save_subclassed_model.py。

2019-06-27 12:57:14.253635: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
2019-06-27 12:57:15.660142: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1432] Found device 0 with properties:
name: GeForce GTX 1060 major: 6 minor: 1 memoryClockRate(GHz): 1.6705
pciBusID: 0000:01:00.0
totalMemory: 6.00GiB freeMemory: 4.97GiB
2019-06-27 12:57:15.660397: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1511] Adding visible gpu devices: 0
2019-06-27 12:57:16.488227: I tensorflow/core/common_runtime/gpu/gpu_device.cc:982] Device interconnect StreamExecutor with strength 1 edge matrix:
2019-06-27 12:57:16.488385: I tensorflow/core/common_runtime/gpu/gpu_device.cc:988] 0
2019-06-27 12:57:16.488476: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1001] 0: N
2019-06-27 12:57:16.488772: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1115] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 4722 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1060, pci bus id: 0000:01:00.0, compute capability: 6.1)
Epoch 1 Batch 100 Loss 0.1120
Epoch 1 Batch 200 Loss 0.0179
Epoch 1 Batch 300 Loss 0.0347
Epoch 1 Loss 0.0111
Epoch 2 Batch 100 Loss 0.0144
Epoch 2 Batch 200 Loss 0.0097
Epoch 2 Batch 300 Loss 0.0141
Epoch 2 Loss 0.0012
Epoch 3 Batch 100 Loss 0.0060
Epoch 3 Batch 200 Loss 0.0037
Epoch 3 Batch 300 Loss 0.0054
Epoch 3 Loss 0.0007
Epoch 4 Batch 100 Loss 0.0088
Epoch 4 Batch 200 Loss 0.0038
Epoch 4 Batch 300 Loss 0.0093
Epoch 4 Loss 0.0004
Epoch 5 Batch 100 Loss 0.0039
Epoch 5 Batch 200 Loss 0.0044
Epoch 5 Batch 300 Loss 0.0031
Epoch 5 Loss 0.0003
Epoch 6 Batch 100 Loss 0.0025
Epoch 6 Batch 200 Loss 0.0038
Epoch 6 Batch 300 Loss 0.0027
Epoch 6 Loss 0.0002
Epoch 7 Batch 100 Loss 0.0026
Epoch 7 Batch 200 Loss 0.0032
Epoch 7 Batch 300 Loss 0.0041
Epoch 7 Loss 0.0002
Epoch 8 Batch 100 Loss 0.0022
Epoch 8 Batch 200 Loss 0.0031
Epoch 8 Batch 300 Loss 0.0026
Epoch 8 Loss 0.0002
Epoch 9 Batch 100 Loss 0.0040
Epoch 9 Batch 200 Loss 0.0014
Epoch 9 Batch 300 Loss 0.0040
Epoch 9 Loss 0.0002
Epoch 10 Batch 100 Loss 0.0023
Epoch 10 Batch 200 Loss 0.0030
Epoch 10 Batch 300 Loss 0.0038
Epoch 10 Loss 0.0002
Epoch 11 Batch 100 Loss 0.0028
Epoch 11 Batch 200 Loss 0.0020
Epoch 11 Batch 300 Loss 0.0025
Epoch 11 Loss 0.0002
Epoch 12 Batch 100 Loss 0.0027
Epoch 12 Batch 200 Loss 0.0045
Epoch 12 Batch 300 Loss 0.0021
Epoch 12 Loss 0.0002
Epoch 13 Batch 100 Loss 0.0016
Epoch 13 Batch 200 Loss 0.0033
Epoch 13 Batch 300 Loss 0.0024
Epoch 13 Loss 0.0002
Epoch 14 Batch 100 Loss 0.0034
Epoch 14 Batch 200 Loss 0.0028
Epoch 14 Batch 300 Loss 0.0033
Epoch 14 Loss 0.0002
Epoch 15 Batch 100 Loss 0.0019
Epoch 15 Batch 200 Loss 0.0030
Epoch 15 Batch 300 Loss 0.0037
Epoch 15 Loss 0.0002 Process finished with exit code 0

查看 checkpoint_dir 目录下的文件。


在 load_subclassed_model.py 中,创建了 200 组测试数据,加载了 the latest checkpoint 中保存的模型参数,对模型进行了测试。

 from subclassed_model import *

 import numpy as np
import matplotlib.pyplot as plt import tensorflow as tf
tf.enable_eager_execution() # load model
encoder = Encoder()
decoder = Decoder()
optimizer = tf.train.AdamOptimizer() checkpoint_dir = './training_checkpoints' checkpoint = tf.train.Checkpoint(optimizer=optimizer,
encoder=encoder,
decoder=decoder)
checkpoint.restore(tf.train.latest_checkpoint(checkpoint_dir)) # build model
BATCH_SIZE = 16 encoder.build(input_shape=tf.TensorShape((BATCH_SIZE, 1)))
decoder.build(input_shape=tf.TensorShape((BATCH_SIZE, UNITS))) encoder.summary()
decoder.summary() # create validation data
X_test = np.linspace(-1, 1, 200) # validation dataset
val_dataset = tf.data.Dataset.from_tensor_slices(X_test).batch(1) # predict and plot
results = []
for (batch, x) in enumerate(val_dataset):
x = tf.cast(x, tf.float32)
x = tf.expand_dims(x, axis=1)
y_ = encoder(x)
prediction = decoder(y_)
# print(prediction.numpy()[0][0])
results.append(prediction.numpy()[0][0]) # plot results
plt.scatter(X_test, results)
plt.show()

运行 load_subclassed_model.py。

2019-06-27 13:27:40.712260: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2
2019-06-27 13:27:42.105938: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1432] Found device 0 with properties:
name: GeForce GTX 1060 major: 6 minor: 1 memoryClockRate(GHz): 1.6705
pciBusID: 0000:01:00.0
totalMemory: 6.00GiB freeMemory: 4.97GiB
2019-06-27 13:27:42.106200: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1511] Adding visible gpu devices: 0
2019-06-27 13:27:42.921364: I tensorflow/core/common_runtime/gpu/gpu_device.cc:982] Device interconnect StreamExecutor with strength 1 edge matrix:
2019-06-27 13:27:42.921510: I tensorflow/core/common_runtime/gpu/gpu_device.cc:988] 0
2019-06-27 13:27:42.921594: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1001] 0: N
2019-06-27 13:27:42.921777: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1115] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 4722 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1060, pci bus id: 0000:01:00.0, compute capability: 6.1)
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense (Dense) multiple 32
_________________________________________________________________
dense_1 (Dense) multiple 136
=================================================================
Total params: 168
Trainable params: 168
Non-trainable params: 0
_________________________________________________________________
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_2 (Dense) multiple 9
=================================================================
Total params: 9
Trainable params: 9
Non-trainable params: 0
_________________________________________________________________ Process finished with exit code 0

版权声明:本文为博主原创文章,欢迎转载,转载请注明作者及原文出处!

[Tensorflow] 使用 tf.train.Checkpoint() 保存 / 加载 keras subclassed model的更多相关文章

  1. [Tensorflow] 使用 model.save_weights() 保存 / 加载 Keras Subclassed Model

    在 parameters.py 中,定义了各类参数. # training data directory TRAINING_DATA_DIR = './data/' # checkpoint dire ...

  2. docker 保存 加载(导入 导出镜像

    tensorflow 的docker镜像很大,pull一次由于墙经常失败.其实docker 可以将镜像导出再导入. 保存加载(tensorflow)镜像 1) 查看镜像 docker images 如 ...

  3. 优化tableView加载cell与model的过程

    优化tableView加载cell与model的过程 效果图 说明 1. 用多态的特性来优化tableView加载cell与model的过程 2. swift写起来果然要比Objective-C简洁了 ...

  4. TensorFlow:tf.train.Saver()模型保存与恢复

    1.保存 将训练好的模型参数保存起来,以便以后进行验证或测试.tf里面提供模型保存的是tf.train.Saver()模块. 模型保存,先要创建一个Saver对象:如 saver=tf.train.S ...

  5. tensorflow的tf.train.Saver()模型保存与恢复

    将训练好的模型参数保存起来,以便以后进行验证或测试.tf里面提供模型保存的是tf.train.Saver()模块. 模型保存,先要创建一个Saver对象:如 saver=tf.train.Saver( ...

  6. tensorflow中 tf.train.slice_input_producer 和 tf.train.batch 函数(转)

    tensorflow数据读取机制 tensorflow中为了充分利用GPU,减少GPU等待数据的空闲时间,使用了两个线程分别执行数据读入和数据计算. 具体来说就是使用一个线程源源不断的将硬盘中的图片数 ...

  7. tensorflow中 tf.train.slice_input_producer 和 tf.train.batch 函数

    tensorflow数据读取机制 tensorflow中为了充分利用GPU,减少GPU等待数据的空闲时间,使用了两个线程分别执行数据读入和数据计算. 具体来说就是使用一个线程源源不断的将硬盘中的图片数 ...

  8. 【转载】 tensorflow中 tf.train.slice_input_producer 和 tf.train.batch 函数

    原文地址: https://blog.csdn.net/dcrmg/article/details/79776876 ----------------------------------------- ...

  9. tensorflow之tf.train.exponential_decay()指数衰减法

    exponential_decay(learning_rate,  global_steps, decay_steps, decay_rate, staircase=False, name=None) ...

随机推荐

  1. 属性 Attribute

    一.创建属性 [AttributeUsage(AttributeTargets.Class | AttributeTargets.Constructor, AllowMultiple = true, ...

  2. Fibonacci数性质

    Fibonacci数性质 0.\(F_{n-1}+F_{n-2}=F_{n} ,特殊的 F_{0}=1,F_{1}=1\) 上述式子为定义式 1.\(F_{0}+F_{1}+...+F_{n}=F_{ ...

  3. 15、Spring Boot 2.x 集成 Swagger UI

    1.15.Spring Boot 2.x 集成 Swagger UI 完整源码: Spring-Boot-Demos 1.15.1 pom文件添加swagger包 <swagger2.versi ...

  4. django 内置标签

    1.autoescape 自动转义开关 官网:https://docs.djangoproject.com/en/2.2/ref/templates/builtins/ 作用:将 html 内容解析成 ...

  5. Codeforces Round #571 (Div. 2)

    A. Vus the Cossack and a Contest 签. #include <bits/stdc++.h> using namespace std; int main() { ...

  6. SSRF漏洞攻击利用从浅到深

    梳理一下ssrf 不详细 简单记录 0x01 SSRF成因和基本利用0x02 内网打未授权redis0x03 关于ssrf打授权的redis0x04 写redis shell和密钥的一点问题0x05 ...

  7. 计算机中buffer和cache的理解

    Linux中Buffer和Cache的区别 Cache 和 Buffer的区别 作者:知乎用户链接:https://www.zhihu.com/question/26190832/answer/323 ...

  8. New in Python 3.8.0

    Python 3.8.0 发布时间: Oct. 14, 2019 这是一个Python3.8.0的稳定发行版. Python3.8.0是最新的Python编程语言发行版,ta包含了许多新的特征和优化. ...

  9. 黑马vue---56-58、vue组件创建的三种方式

    黑马vue---56-58.vue组件创建的三种方式 一.总结 一句话总结: 不论是哪种方式创建出来的组件,组件的 template 属性指向的模板内容,必须有且只能有唯一的一个根元素 1.使用 Vu ...

  10. Python 自学笔记(五)

    1.布尔值 1-1.概念 定义计算机中的逻辑判断,只有两种结果,True和False. if,while后面的判断条件就是布尔值,只有条件为True的时候才执行. 1-2.数值比较 1-3.数值运算 ...