ShiyinTan / tensorflow-learning

tensorflow-learning

Install

1. install anaconda
2. conda create -n env_tensorflow pip python=3.7
3. conda activate env_tensorflow
4. pip install --upgrade tensorflow
5. install Microsoft Visual C++ Redistributable for Visual Studio 2019 from https://visualstudio.microsoft.com/zh-hans/downloads/
6. install cuda 10.1 from https://developer.nvidia.com/cuda-toolkit-archive
7. install jupyter notebook conda install nb_conda

error

keras.utils.plot_model():

1. should install graphviz and pydot first
2. pip install pydot
3. downloading graphviz from https://graphviz.gitlab.io/
4. setup graphviz
5. add "C:\programfile\graphviz\bin" to PATH of User
6. add "C:\programfile\graphviz\bin\dot.exe" to PATH of System
7. conda install graphviz
8. conda install pydotplus

keras.datasets.mnist.load():

通常因为国内无法加载数据，那么我们可以先下载数据，然后根据tensorflow提供的源码自己写一个load_data函数

1. download mnist dataset from https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz. 该网址可从keras.datasets.mnist.load()的源码中获取
2. load_data函数，用来读取下载的数据

def load_data(path):
  with np.load(path, allow_pickle=True) as f:
    x_train, y_train = f['x_train'], f['y_train']
    x_test, y_test = f['x_test'], f['y_test']

    return (x_train, y_train), (x_test, y_test)

tensorflow

• logits: in tensorflow, means that unnormalized "probability", often as output of dense layer and input of softmax layer

keras

model

• tf.keras.Model(inputs,outputs,name), 函数式API
• class Mymodel(tf.keras.Model), 子类model, 子类model需要实现__init__构造函数和前向传播call. 其中training可以控制模型在training和testing阶段拥有不同的策略。

class Mymodel(tf.keras.Model):
	def __init__(self):
		super(Mymodel,self).__init__()
		self.dense1 = tf.keras.layers.Dense(4, activation = tf.nn.relu)
		self.dense2 = tf.keras.layers.Dense(5, activation = tf.nn.softmax)
		self.dropout = tf.keras.layers.Dropout(0.5)
	def call(self, inputs, training=False):
		x = self.dense1(inputs)
		if training:
			x = self.dropout(x)
		return self.dense2(x)

注¹.

• model.compile(...): 定义模型的config，例如losses，metrics等。
• model.fit(): train the model
• model.predict(): use the model to do prediction
• model.evaluate():

Footnotes

1. 当子类重写父类构造函数时，如果想要调用父类构造函数时，必须显式调用。 ↩