一、TensorFlow Probability 简介
TensorFlow Probability 是通过概率编程实现深度学习的开源库。
在传统深度学习中,模型参数通常是确定性或固定的。但是,在模型和训练数据不充分或存在噪声时,也许最佳参数的不确定性就会开始出现问题。此时,传统深度学习模型会将噪声视为训练集的一部分,但是概率编程可以处理这种不确定性并将其视为噪声源作为输入。
TensorFlow Probability 建立在 TensorFlow 的基础上,并使用可微分正则概率分布表示神经网络的权重和偏差。
二、概率编程的基本思想
概率编程可以看作是一种建立在概率理论基础上的编程范式。通过概率编程,开发者可以使用概率建模和推断进行灵活计算,能够处理如下问题:
1、在有限的数据集上推断参数
2、解决样本规模很小的问题
3、处理缺失数据集
4、在观测到有限数据时提高模型的不确定性
三、TensorFlow Probability 实现深度学习的方法
1、使用正则化的分布建议网络的权重
import tensorflow as tf import tensorflow_probability as tfp model = tf.keras.Sequential([ tf.keras.layers.Dense( 10, activation=tf.nn.relu, input_shape=(n_features,)), # Apply the L2 regularization with a factor of 1/lambda tfp.layers.DenseVariational( 1, posterior_mean_field, prior_trainable), # input_shape=(10,) ])
2、将权重和偏差使用正则化的分布建议
def build_normal_predictive_distribution(layer): """Create the predictive distribution of a layer output.""" def normal_predictive_distribution(distribution): """Create the predictive distribution of a distribution.""" return tfp.distributions.Normal(distribution.mean(), distribution.stddev()) return tfp.layers.DistributionLambda( make_distribution_fn=normal_predictive_distribution, convert_to_tensor_fn=tfp.distributions.Distribution.sample, name=f'{layer.name}_distribution') model = tf.keras.Sequential([ tfp.layers.DenseVariational( 10, posterior_mean_field, prior_trainable, activation=tf.nn.relu, input_shape=(n_features,), ), build_normal_predictive_distribution(model.layers[-1]), tfp.layers.DistributionLambda( make_distribution_fn=lambda t: tfp.distributions.Normal( t, scale=1), name='y_distribution'), ])
3、使用可微分概率分布进行推断
model.compile( optimizer=tf.optimizers.Adam(lr=learning_rate), loss=neg_log_likelihood, metrics=[neg_log_likelihood, accuracy]) history = model.fit( x_train, y_train, batch_size=batch_size, epochs=n_epochs, validation_data=(x_test, y_test), verbose=0)
四、TensorFlow Probability 实现深度学习的实例
下面的代码实现了在 MNIST 数据集上对手写数字进行分类的概率编程方法——使用 CNN(MLP) + dropout-MLP 模型。
import tensorflow as tf import tensorflow_probability as tfp from sklearn.datasets import fetch_openml from sklearn.model_selection import train_test_split tfd = tfp.distributions # Load data X, y = fetch_openml('mnist_784', version=1, return_X_y=True) X, y = X[:5000, ...].astype('float32'), y[:5000] X /= 255. y = y.astype('int32') x_train, x_test, y_train, y_test = train_test_split(X, y, train_size=0.8, random_state=42) n_samples = 50 n_epochs = 100 batch_size = 128 n_features = x_train.shape[-1] n_classes = 10 def convolutional_input_encoder(x, input_shape): """Pre-process the inputs of the convolutional encoder.""" x = tf.reshape(x, [-1] + list(input_shape) + [1]) return x def convolutional_output_decoder(x): """Re-format and flatten the outputs of the convolutional decoder.""" _, h, w, c = x.shape.as_list() x = tf.reshape(x, [-1, h*w*c]) return x model = tf.keras.Sequential([ tf.keras.layers.Lambda( lambda x: convolutional_input_encoder( x, input_shape=(28, 28)), input_shape=(28, 28)), tf.keras.layers.Conv2D( 32, [5, 5], strides=[1, 1], padding='same', activation=tf.nn.relu), tf.keras.layers.MaxPooling2D( pool_size=[2, 2], strides=[2, 2], padding='same'), tf.keras.layers.Conv2D( 64, [5, 5], strides=[1, 1], padding='same', activation=tf.nn.relu), tf.keras.layers.MaxPooling2D( pool_size=[2, 2], strides=[2, 2], padding='same'), tf.keras.layers.Lambda( lambda x: convolutional_output_decoder(x)), tfp.layers.DenseFlipout( 256, activation=tf.nn.relu), tf.keras.layers.Dropout(0.5), tfp.layers.DenseFlipout( 10), tfp.layers.DistributionLambda( lambda t: tfd.Categorical(logits=t), name='y_dist') ]) def neg_log_likelihood(y_true, y_pred): return -tf.reduce_mean(y_pred.log_prob(tf.squeeze(y_true))) model.compile(optimizer=tf.optimizers.Adam(learning_rate=0.001), loss=neg_log_likelihood) history = model.fit( x_train, y_train, batch_size=batch_size, epochs=n_epochs, verbose=0, validation_data=(x_test, y_test)) _, test_acc = model.evaluate(x_test, y_test, verbose=0) print(f'Test accuracy: {test_acc:.4f}')