TensorFlow2.0使用keras訓(xùn)練模型的實(shí)現(xiàn)
更新時(shí)間：2021年02月20日 09:38:27 作者：Doit_
這篇文章主要介紹了TensorFlow2.0使用keras訓(xùn)練模型的實(shí)現(xiàn)，文中通過示例代碼介紹的非常詳細(xì)，對大家的學(xué)習(xí)或者工作具有一定的參考學(xué)習(xí)價(jià)值，需要的朋友們下面隨著小編來一起學(xué)習(xí)學(xué)習(xí)吧
1.一般的模型構(gòu)造、訓(xùn)練、測試流程
# 模型構(gòu)造
inputs = keras.Input(shape=(784,), name='mnist_input')
h1 = layers.Dense(64, activation='relu')(inputs)
h1 = layers.Dense(64, activation='relu')(h1)
outputs = layers.Dense(10, activation='softmax')(h1)
model = keras.Model(inputs, outputs)
# keras.utils.plot_model(model, 'net001.png', show_shapes=True)
model.compile(optimizer=keras.optimizers.RMSprop(),
loss=keras.losses.SparseCategoricalCrossentropy(),
metrics=[keras.metrics.SparseCategoricalAccuracy()])
# 載入數(shù)據(jù)
(x_train, y_train), (x_test, y_test) = keras.datasets.mnist.load_data()
x_train = x_train.reshape(60000, 784).astype('float32') /255
x_test = x_test.reshape(10000, 784).astype('float32') /255
x_val = x_train[-10000:]
y_val = y_train[-10000:]
x_train = x_train[:-10000]
y_train = y_train[:-10000]
# 訓(xùn)練模型
history = model.fit(x_train, y_train, batch_size=64, epochs=3,
validation_data=(x_val, y_val))
print('history:')
print(history.history)
result = model.evaluate(x_test, y_test, batch_size=128)
print('evaluate:')
print(result)
pred = model.predict(x_test[:2])
print('predict:')
print(pred)
2.自定義損失和指標(biāo)
自定義指標(biāo)只需繼承Metric類， 并重寫一下函數(shù)
_init_(self)，初始化。
update_state(self，y_true，y_pred，sample_weight = None)，它使用目標(biāo)y_true和模型預(yù)測y_pred來更新狀態(tài)變量。
result(self)，它使用狀態(tài)變量來計(jì)算最終結(jié)果。
reset_states(self)，重新初始化度量的狀態(tài)。
# 這是一個(gè)簡單的示例，顯示如何實(shí)現(xiàn)CatgoricalTruePositives指標(biāo)，該指標(biāo)計(jì)算正確分類為屬于給定類的樣本數(shù)量
class CatgoricalTruePostives(keras.metrics.Metric):
def __init__(self, name='binary_true_postives', **kwargs):
super(CatgoricalTruePostives, self).__init__(name=name, **kwargs)
self.true_postives = self.add_weight(name='tp', initializer='zeros')
def update_state(self, y_true, y_pred, sample_weight=None):
y_pred = tf.argmax(y_pred)
y_true = tf.equal(tf.cast(y_pred, tf.int32), tf.cast(y_true, tf.int32))
y_true = tf.cast(y_true, tf.float32)
if sample_weight is not None:
sample_weight = tf.cast(sample_weight, tf.float32)
y_true = tf.multiply(sample_weight, y_true)
return self.true_postives.assign_add(tf.reduce_sum(y_true))
def result(self):
return tf.identity(self.true_postives)
def reset_states(self):
self.true_postives.assign(0.)
model.compile(optimizer=keras.optimizers.RMSprop(1e-3),
loss=keras.losses.SparseCategoricalCrossentropy(),
metrics=[CatgoricalTruePostives()])
model.fit(x_train, y_train,
batch_size=64, epochs=3)
# 以定義網(wǎng)絡(luò)層的方式添加網(wǎng)絡(luò)loss
class ActivityRegularizationLayer(layers.Layer):
def call(self, inputs):
self.add_loss(tf.reduce_sum(inputs) * 0.1)
return inputs
inputs = keras.Input(shape=(784,), name='mnist_input')
h1 = layers.Dense(64, activation='relu')(inputs)
h1 = ActivityRegularizationLayer()(h1)
h1 = layers.Dense(64, activation='relu')(h1)
outputs = layers.Dense(10, activation='softmax')(h1)
model = keras.Model(inputs, outputs)
# keras.utils.plot_model(model, 'net001.png', show_shapes=True)
model.compile(optimizer=keras.optimizers.RMSprop(),
loss=keras.losses.SparseCategoricalCrossentropy(),
metrics=[keras.metrics.SparseCategoricalAccuracy()])
model.fit(x_train, y_train, batch_size=32, epochs=1)
# 也可以以定義網(wǎng)絡(luò)層的方式添加要統(tǒng)計(jì)的metric
class MetricLoggingLayer(layers.Layer):
def call(self, inputs):
self.add_metric(keras.backend.std(inputs),
name='std_of_activation',
aggregation='mean')
return inputs
inputs = keras.Input(shape=(784,), name='mnist_input')
h1 = layers.Dense(64, activation='relu')(inputs)
h1 = MetricLoggingLayer()(h1)
h1 = layers.Dense(64, activation='relu')(h1)
outputs = layers.Dense(10, activation='softmax')(h1)
model = keras.Model(inputs, outputs)
# keras.utils.plot_model(model, 'net001.png', show_shapes=True)
model.compile(optimizer=keras.optimizers.RMSprop(),
loss=keras.losses.SparseCategoricalCrossentropy(),
metrics=[keras.metrics.SparseCategoricalAccuracy()])
model.fit(x_train, y_train, batch_size=32, epochs=1)
# 也可以直接在model上面加
# 也可以以定義網(wǎng)絡(luò)層的方式添加要統(tǒng)計(jì)的metric
class MetricLoggingLayer(layers.Layer):
def call(self, inputs):
self.add_metric(keras.backend.std(inputs),
name='std_of_activation',
aggregation='mean')
return inputs
inputs = keras.Input(shape=(784,), name='mnist_input')
h1 = layers.Dense(64, activation='relu')(inputs)
h2 = layers.Dense(64, activation='relu')(h1)
outputs = layers.Dense(10, activation='softmax')(h2)
model = keras.Model(inputs, outputs)
model.add_metric(keras.backend.std(inputs),
name='std_of_activation',
aggregation='mean')
model.add_loss(tf.reduce_sum(h1)*0.1)
# keras.utils.plot_model(model, 'net001.png', show_shapes=True)
model.compile(optimizer=keras.optimizers.RMSprop(),
loss=keras.losses.SparseCategoricalCrossentropy(),
metrics=[keras.metrics.SparseCategoricalAccuracy()])
model.fit(x_train, y_train, batch_size=32, epochs=1)
處理使用validation_data傳入測試數(shù)據(jù)，還可以使用validation_split劃分驗(yàn)證數(shù)據(jù)
ps:validation_split只能在用numpy數(shù)據(jù)訓(xùn)練的情況下使用
model.fit(x_train, y_train, batch_size=32, epochs=1, validation_split=0.2)
3.使用tf.data構(gòu)造數(shù)據(jù)
def get_compiled_model():
inputs = keras.Input(shape=(784,), name='mnist_input')
h1 = layers.Dense(64, activation='relu')(inputs)
h2 = layers.Dense(64, activation='relu')(h1)
outputs = layers.Dense(10, activation='softmax')(h2)
model = keras.Model(inputs, outputs)
model.compile(optimizer=keras.optimizers.RMSprop(),
loss=keras.losses.SparseCategoricalCrossentropy(),
metrics=[keras.metrics.SparseCategoricalAccuracy()])
return model
model = get_compiled_model()
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(64)
val_dataset = tf.data.Dataset.from_tensor_slices((x_val, y_val))
val_dataset = val_dataset.batch(64)
# model.fit(train_dataset, epochs=3)
# steps_per_epoch 每個(gè)epoch只訓(xùn)練幾步
# validation_steps 每次驗(yàn)證，驗(yàn)證幾步
model.fit(train_dataset, epochs=3, steps_per_epoch=100,
validation_data=val_dataset, validation_steps=3)
4.樣本權(quán)重和類權(quán)重
“樣本權(quán)重”數(shù)組是一個(gè)數(shù)字?jǐn)?shù)組，用于指定批處理中每個(gè)樣本在計(jì)算總損失時(shí)應(yīng)具有多少權(quán)重。 它通常用于不平衡的分類問題（這個(gè)想法是為了給予很少見的類更多的權(quán)重）。 當(dāng)使用的權(quán)重是1和0時(shí)，該數(shù)組可以用作損失函數(shù)的掩碼（完全丟棄某些樣本對總損失的貢獻(xiàn)）。
“類權(quán)重”dict是同一概念的更具體的實(shí)例：它將類索引映射到應(yīng)該用于屬于該類的樣本的樣本權(quán)重。 例如，如果類“0”比數(shù)據(jù)中的類“1”少兩倍，則可以使用class_weight = {0：1.，1：0.5}。
# 增加第5類的權(quán)重
import numpy as np
# 樣本權(quán)重
model = get_compiled_model()
class_weight = {i:1.0 for i in range(10)}
class_weight[5] = 2.0
print(class_weight)
model.fit(x_train, y_train,
class_weight=class_weight,
batch_size=64,
epochs=4)
# 類權(quán)重
model = get_compiled_model()
sample_weight = np.ones(shape=(len(y_train),))
sample_weight[y_train == 5] = 2.0
model.fit(x_train, y_train,
sample_weight=sample_weight,
batch_size=64,
epochs=4)
# tf.data數(shù)據(jù)
model = get_compiled_model()
sample_weight = np.ones(shape=(len(y_train),))
sample_weight[y_train == 5] = 2.0
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train,
sample_weight))
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(64)
val_dataset = tf.data.Dataset.from_tensor_slices((x_val, y_val))
val_dataset = val_dataset.batch(64)
model.fit(train_dataset, epochs=3, )
5.多輸入多輸出模型
image_input = keras.Input(shape=(32, 32, 3), name='img_input')
timeseries_input = keras.Input(shape=(None, 10), name='ts_input')
x1 = layers.Conv2D(3, 3)(image_input)
x1 = layers.GlobalMaxPooling2D()(x1)
x2 = layers.Conv1D(3, 3)(timeseries_input)
x2 = layers.GlobalMaxPooling1D()(x2)
x = layers.concatenate([x1, x2])
score_output = layers.Dense(1, name='score_output')(x)
class_output = layers.Dense(5, activation='softmax', name='class_output')(x)
model = keras.Model(inputs=[image_input, timeseries_input],
outputs=[score_output, class_output])
keras.utils.plot_model(model, 'multi_input_output_model.png'
, show_shapes=True)
# 可以為模型指定不同的loss和metrics
model.compile(
optimizer=keras.optimizers.RMSprop(1e-3),
loss=[keras.losses.MeanSquaredError(),
keras.losses.CategoricalCrossentropy()])
# 還可以指定loss的權(quán)重
model.compile(
optimizer=keras.optimizers.RMSprop(1e-3),
loss={'score_output': keras.losses.MeanSquaredError(),
'class_output': keras.losses.CategoricalCrossentropy()},
metrics={'score_output': [keras.metrics.MeanAbsolutePercentageError(),
keras.metrics.MeanAbsoluteError()],
'class_output': [keras.metrics.CategoricalAccuracy()]},
loss_weight={'score_output': 2., 'class_output': 1.})
# 可以把不需要傳播的loss置0
model.compile(
optimizer=keras.optimizers.RMSprop(1e-3),
loss=[None, keras.losses.CategoricalCrossentropy()])
# Or dict loss version
model.compile(
optimizer=keras.optimizers.RMSprop(1e-3),
loss={'class_output': keras.losses.CategoricalCrossentropy()})
6.使用回 調(diào)
Keras中的回調(diào)是在訓(xùn)練期間（在epoch開始時(shí)，batch結(jié)束時(shí)，epoch結(jié)束時(shí)等）在不同點(diǎn)調(diào)用的對象，可用于實(shí)現(xiàn)以下行為：
- 在培訓(xùn)期間的不同時(shí)間點(diǎn)進(jìn)行驗(yàn)證（超出內(nèi)置的每個(gè)時(shí)期驗(yàn)證）
- 定期檢查模型或超過某個(gè)精度閾值
- 在訓(xùn)練似乎平穩(wěn)時(shí)改變模型的學(xué)習(xí)率
- 在訓(xùn)練似乎平穩(wěn)時(shí)對頂層進(jìn)行微調(diào)
- 在培訓(xùn)結(jié)束或超出某個(gè)性能閾值時(shí)發(fā)送電子郵件或即時(shí)消息通知等等。
可使用的內(nèi)置回調(diào)有
- ModelCheckpoint：定期保存模型。
- EarlyStopping：當(dāng)訓(xùn)練不再改進(jìn)驗(yàn)證指標(biāo)時(shí)停止培訓(xùn)。
- TensorBoard：定期編寫可在TensorBoard中顯示的模型日志（更多細(xì)節(jié)見“可視化”）。
- CSVLogger：將丟失和指標(biāo)數(shù)據(jù)流式傳輸?shù)紺SV文件。
- 等等
6.1回調(diào)使用
model = get_compiled_model() callbacks = [ keras.callbacks.EarlyStopping( # Stop training when `val_loss` is no longer improving monitor='val_loss', # "no longer improving" being defined as "no better than 1e-2 less" min_delta=1e-2, # "no longer improving" being further defined as "for at least 2 epochs" patience=2, verbose=1) ] model.fit(x_train, y_train, epochs=20, batch_size=64, callbacks=callbacks, validation_split=0.2)
# checkpoint模型回調(diào)
model = get_compiled_model()
check_callback = keras.callbacks.ModelCheckpoint(
filepath='mymodel_{epoch}.h5',
save_best_only=True,
monitor='val_loss',
verbose=1
)
model.fit(x_train, y_train,
epochs=3,
batch_size=64,
callbacks=[check_callback],
validation_split=0.2)
# 動(dòng)態(tài)調(diào)整學(xué)習(xí)率 initial_learning_rate = 0.1 lr_schedule = keras.optimizers.schedules.ExponentialDecay( initial_learning_rate, decay_steps=10000, decay_rate=0.96, staircase=True ) optimizer = keras.optimizers.RMSprop(learning_rate=lr_schedule)
# 使用tensorboard tensorboard_cbk = keras.callbacks.TensorBoard(log_dir='./full_path_to_your_logs') model.fit(x_train, y_train, epochs=5, batch_size=64, callbacks=[tensorboard_cbk], validation_split=0.2)
6.2創(chuàng)建自己的回調(diào)方法
class LossHistory(keras.callbacks.Callback):
def on_train_begin(self, logs):
self.losses = []
def on_epoch_end(self, batch, logs):
self.losses.append(logs.get('loss'))
print('\nloss:',self.losses[-1])
model = get_compiled_model()
callbacks = [
LossHistory()
]
model.fit(x_train, y_train,
epochs=3,
batch_size=64,
callbacks=callbacks,
validation_split=0.2)
7.自己構(gòu)造訓(xùn)練和驗(yàn)證循環(huán)
# Get the model.
inputs = keras.Input(shape=(784,), name='digits')
x = layers.Dense(64, activation='relu', name='dense_1')(inputs)
x = layers.Dense(64, activation='relu', name='dense_2')(x)
outputs = layers.Dense(10, activation='softmax', name='predictions')(x)
model = keras.Model(inputs=inputs, outputs=outputs)
# Instantiate an optimizer.
optimizer = keras.optimizers.SGD(learning_rate=1e-3)
# Instantiate a loss function.
loss_fn = keras.losses.SparseCategoricalCrossentropy()
# Prepare the training dataset.
batch_size = 64
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(batch_size)
# 自己構(gòu)造循環(huán)
for epoch in range(3):
print('epoch: ', epoch)
for step, (x_batch_train, y_batch_train) in enumerate(train_dataset):
# 開一個(gè)gradient tape, 計(jì)算梯度
with tf.GradientTape() as tape:
logits = model(x_batch_train)
loss_value = loss_fn(y_batch_train, logits)
grads = tape.gradient(loss_value, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
if step % 200 == 0:
print('Training loss (for one batch) at step %s: %s' % (step, float(loss_value)))
print('Seen so far: %s samples' % ((step + 1) * 64))
# 訓(xùn)練并驗(yàn)證
# Get model
inputs = keras.Input(shape=(784,), name='digits')
x = layers.Dense(64, activation='relu', name='dense_1')(inputs)
x = layers.Dense(64, activation='relu', name='dense_2')(x)
outputs = layers.Dense(10, activation='softmax', name='predictions')(x)
model = keras.Model(inputs=inputs, outputs=outputs)
# Instantiate an optimizer to train the model.
optimizer = keras.optimizers.SGD(learning_rate=1e-3)
# Instantiate a loss function.
loss_fn = keras.losses.SparseCategoricalCrossentropy()
# Prepare the metrics.
train_acc_metric = keras.metrics.SparseCategoricalAccuracy()
val_acc_metric = keras.metrics.SparseCategoricalAccuracy()
# Prepare the training dataset.
batch_size = 64
train_dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
train_dataset = train_dataset.shuffle(buffer_size=1024).batch(batch_size)
# Prepare the validation dataset.
val_dataset = tf.data.Dataset.from_tensor_slices((x_val, y_val))
val_dataset = val_dataset.batch(64)
# Iterate over epochs.
for epoch in range(3):
print('Start of epoch %d' % (epoch,))
# Iterate over the batches of the dataset.
for step, (x_batch_train, y_batch_train) in enumerate(train_dataset):
with tf.GradientTape() as tape:
logits = model(x_batch_train)
loss_value = loss_fn(y_batch_train, logits)
grads = tape.gradient(loss_value, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
# Update training metric.
train_acc_metric(y_batch_train, logits)
# Log every 200 batches.
if step % 200 == 0:
print('Training loss (for one batch) at step %s: %s' % (step, float(loss_value)))
print('Seen so far: %s samples' % ((step + 1) * 64))
# Display metrics at the end of each epoch.
train_acc = train_acc_metric.result()
print('Training acc over epoch: %s' % (float(train_acc),))
# Reset training metrics at the end of each epoch
train_acc_metric.reset_states()
# Run a validation loop at the end of each epoch.
for x_batch_val, y_batch_val in val_dataset:
val_logits = model(x_batch_val)
# Update val metrics
val_acc_metric(y_batch_val, val_logits)
val_acc = val_acc_metric.result()
val_acc_metric.reset_states()
print('Validation acc: %s' % (float(val_acc),))
## 添加自己構(gòu)造的loss, 每次只能看到最新一次訓(xùn)練增加的loss class ActivityRegularizationLayer(layers.Layer): def call(self, inputs): self.add_loss(1e-2 * tf.reduce_sum(inputs)) return inputs inputs = keras.Input(shape=(784,), name='digits') x = layers.Dense(64, activation='relu', name='dense_1')(inputs) # Insert activity regularization as a layer x = ActivityRegularizationLayer()(x) x = layers.Dense(64, activation='relu', name='dense_2')(x) outputs = layers.Dense(10, activation='softmax', name='predictions')(x) model = keras.Model(inputs=inputs, outputs=outputs) logits = model(x_train[:64]) print(model.losses) logits = model(x_train[:64]) logits = model(x_train[64: 128]) logits = model(x_train[128: 192]) print(model.losses)
# 將loss添加進(jìn)求導(dǎo)中
optimizer = keras.optimizers.SGD(learning_rate=1e-3)
for epoch in range(3):
print('Start of epoch %d' % (epoch,))
for step, (x_batch_train, y_batch_train) in enumerate(train_dataset):
with tf.GradientTape() as tape:
logits = model(x_batch_train)
loss_value = loss_fn(y_batch_train, logits)
# Add extra losses created during this forward pass:
loss_value += sum(model.losses)
grads = tape.gradient(loss_value, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
# Log every 200 batches.
if step % 200 == 0:
print('Training loss (for one batch) at step %s: %s' % (step, float(loss_value)))
print('Seen so far: %s samples' % ((step + 1) * 64))
到此這篇關(guān)于TensorFlow2.0使用keras訓(xùn)練模型的實(shí)現(xiàn)的文章就介紹到這了,更多相關(guān)TensorFlow2.0 keras訓(xùn)練模型內(nèi)容請搜索腳本之家以前的文章或繼續(xù)瀏覽下面的相關(guān)文章希望大家以后多多支持腳本之家！
相關(guān)文章
python 基于空間相似度的K-means軌跡聚類的實(shí)現(xiàn)
這篇文章主要介紹了python 基于空間相似度的K-means軌跡聚類的實(shí)現(xiàn)，文中通過示例代碼介紹的非常詳細(xì)，對大家的學(xué)習(xí)或者工作具有一定的參考學(xué)習(xí)價(jià)值，需要的朋友們下面隨著小編來一起學(xué)習(xí)學(xué)習(xí)吧2021-03-03
大家都知道Django 是一個(gè)由 Python 編寫的一個(gè)開放源代碼的 Web 應(yīng)用框架。接下來通過本文給大家介紹Django的基本運(yùn)用之Django垃圾分類詳解，感興趣的朋友一起看看吧2021-09-09
python常用函數(shù)random()函數(shù)詳解
這篇文章主要介紹了python常用函數(shù)random()函數(shù),本文通過實(shí)例代碼給大家介紹的非常詳細(xì)，對大家的學(xué)習(xí)或工作具有一定的參考借鑒價(jià)值，需要的朋友可以參考下2023-02-02
這篇文章主要介紹了python學(xué)習(xí)實(shí)操案例，這一篇小編給大家?guī)淼氖橇斜恚赃@里是和列表有關(guān)的案例，需要的小伙伴可以參考一下，希望對你有所幫助<BR>2022-02-02
這篇文章主要介紹了Python中的is和==比較兩個(gè)對象的兩種方法的相關(guān)資料,希望通過本文能幫助到大家，需要的朋友可以參考下2017-09-09
python入門字符串拼接\截取\轉(zhuǎn)數(shù)字理解學(xué)習(xí)
本篇內(nèi)容我們主要講有關(guān)Python字符串的用法，包括字符串的拼接、字符串怎么轉(zhuǎn)數(shù)字、字符串的格式化、字符串函數(shù)等內(nèi)容,有需要的朋友可以借鑒參考下2021-09-09
這篇文章將介紹圖像量化處理，即將圖像像素點(diǎn)對應(yīng)亮度的連續(xù)變化區(qū)間轉(zhuǎn)換為單個(gè)特定值的過程。文中的示例代碼講解詳細(xì)，需要的可以參考一下2022-02-02
pandas數(shù)據(jù)清洗(缺失值和重復(fù)值的處理)
這篇文章主要介紹了pandas數(shù)據(jù)清洗(缺失值和重復(fù)值的處理)，pandas對大數(shù)據(jù)有很多便捷的清洗用法，尤其針對缺失值和重復(fù)值，詳細(xì)介紹感興趣的小伙伴可以參考下面文章內(nèi)容2022-08-08
Python數(shù)據(jù)可視化之用Matplotlib繪制常用圖形
Matplotlib能夠繪制折線圖、散點(diǎn)圖、柱狀圖、直方圖、餅圖. 我們需要知道不同的統(tǒng)計(jì)圖的意義,以此來決定選擇哪種統(tǒng)計(jì)圖來呈現(xiàn)我們的數(shù)據(jù),今天就帶大家詳細(xì)了解如何繪制這些常用圖形,需要的朋友可以參考下2021-06-06
Python自動(dòng)化測試Eclipse+Pydev 搭建開發(fā)環(huán)境
本文主要介紹python自動(dòng)化測試環(huán)境搭建,這里對Eclipse+Pydev 搭建開發(fā)環(huán)境做了圖文詳解，有需要的小伙伴可以參考下2016-08-08