Tensorflow2.0 deep learning simple tutorial of running code

Before written Pytorch, Now work needs are used Tensorflow2.0 Framework , After this period of study , To sum up .Tensorflow2.0 There are two training modes that you can use .

Simple mode

Tensorflow Deep learning simple mode , Its basic idea is to encapsulate all operations , It's easy if you can , So just know the input data format and how it receives data .

Collating data

#  Introduction package 
import tensorflow as tf

#  Do data ,tensorflow A particularly good mechanism is to encapsulate common data packets 
mnist = tf.keras.datasets.mnist

#  there x_train, y_train and x_test, y_test The specific data is ordinary array type 
(x_train, y_train), (x_test, y_test) = mnist.load_data()
x_train, x_test = x_train / 255.0, x_test / 255.0

Model

Tensorflow The grammar of is always encapsulated , Many API It's all packaged , Don't write it yourself , Just find a little trouble

#  The model establishment and pytorch Of `Sequential` As simple as listing , Code execution is sequential execution 
model = tf.keras.models.Sequential([
  tf.keras.layers.Flatten(input_shape=(28, 28)),
  tf.keras.layers.Dense(128, activation='relu'),
  tf.keras.layers.Dropout(0.2),
  tf.keras.layers.Dense(10, activation='softmax')
])

#  Here is how to set the optimizer , Loss function , And index parameters 
model.compile(optimizer='adam',
              loss='sparse_categorical_crossentropy',
              metrics=['accuracy'])

Run

It's much easier to run

#  Training 
model.fit(x_train, y_train, epochs=5)
#  verification 
model.evaluate(x_test,  y_test, verbose=2)

Of course, you can also customize some callback functions , To control the early stop , Model storage , Learning rate optimization and so on .

Expert model

There are many expert modes that we can set ourselves. Let's take a look step by step

Collating data

It should be noted that Tensorflow Unlike Pytorch The same data is <x, y> One to one to ground output , It's a column x And a column y apart .

#  Introduction package 
import tensorflow as tf
import numpy as np
from tensorflow.keras.layers import Dense, Flatten, Conv2D
from tensorflow.keras import Model

#  Make the simplest data 
x = [i for i in range(1, 10000)]
y = [(i+1)*(i+2) for i in x]

#  Turning into Tensorflow Of Tensor The format needs to be converted to numpy Format 
x = np.array(x)
y = np.array(y)

#  This is similar to Pytorch Of Dataset
# batch Namely batchsize
# shuffle Is the degree to which the data order is disrupted 
# map It can carry out secondary processing on the proposed data 

#  The input here should be the same as from_tensor_slices((x, y)) bring into correspondence with , Output with yourself 
def fun(x, y):
    return x, x+2, y

#  there ds Namely Tensorflow Where the data is stored , Be similar to Pytorch Of DataLoader, It can be used for Loop printing 
ds = tf.data.Dataset.from_tensor_slices((x, y)).batch(1000).shuffle(5000).map(fun)

Model

This model is also similar to Pytorch Almost , But you need to rewrite the inheritance method call, It's forward propagation .

class MyModel(Model):
  def __init__(self):
    super(MyModel, self).__init__()
    self.d = Dense(32, activation='relu')
    self.flatten = Flatten()
    self.d1 = Dense(32, activation='relu')
    self.d2 = Dense(10)

  #  Above map Two of them x, Here we also use two parameters to receive 
  def call(self, x, x1):
    x = self.d(x)
    x = self.flatten(x)
    x1 = self.d1(x1)
    return self.d2((x+x1))

#  Create a model instance 
model = MyModel()

#  Then define the optimizer and loss function 
loss_object = tf.keras.losses.MeanSquaredError()
optimizer = tf.keras.optimizers.Adam()

#  Loss storage 
train_loss = tf.keras.metrics.Mean(name='train_loss')

Run

Generally, a training method will be defined , Add... Before the method @tf.function this sentence , Can speed up the training , But you can't print the information of data flow .

@tf.function
def train_step(x, x1, y):
  with tf.GradientTape() as tape:
    # training=True is only needed if there are layers with different
    # behavior during training versus inference (e.g. Dropout).
    predictions = model(x, x1, training=True)
    loss = loss_object(y, predictions)
  gradients = tape.gradient(loss, model.trainable_variables)
  optimizer.apply_gradients(zip(gradients, model.trainable_variables))

  train_loss(loss)

#  Take a few reincarnations 
EPOCHS = 500

for epoch in range(EPOCHS):
  # Reset the metrics at the start of the next epoch
  train_loss.reset_states()
  
  for x, x1, y in ds:
    train_step(x, x1, y)
  print(
    f'Epoch {
      epoch + 1}, '
    f'Loss: {
      train_loss.result()}, '
  )

That's basically it ,Tensorflow Of API Especially rich , Familiar with many functions, you don't have to write them yourself , It's quite convenient , But you need to understand .