使用混合密度網路逼近非函數映射
作者： lukewood
建立日期 2023/07/15
上次修改日期 2023/07/15
說明： 使用混合密度網路逼近非一對一映射。
逼近非函數
神經網路是通用的函數逼近器。關鍵字：函數！雖然是強大的函數逼近器，但神經網路無法逼近非函數。關於函數，要記住一個重要的限制 - 它們有一個輸入、一個輸出！當訓練集對於單個 X 有多個 Y 值時，神經網路會受到很大的影響。
在本指南中，我將向您展示如何逼近非函數類別，該類別包含從 x -> y 的映射，使得對於給定的 x 可能存在多個 y。我們將使用一種稱為「混合密度網路」的神經網路類別。
我將使用新的 多後端 Keras Core 專案 來建立我的混合密度網路。Keras 團隊在這個專案上做得很好 - 能夠在一行程式碼中交換框架真是太棒了。
一些壞消息：我在本指南中使用 TensorFlow probability... 所以它實際上無法與其他後端一起使用。
無論如何，讓我們從安裝相依性和整理匯入開始
!pip install -q --upgrade tensorflow-probability keras-core
import numpy as np
import matplotlib.pyplot as plt
import math
import random
from keras_core import callbacks
import keras_core
import tensorflow as tf
from keras_core import layers
from keras_core import optimizers
from tensorflow_probability import distributions as tfd
Using TensorFlow backend
接下來，讓我們產生一個有雜訊的螺旋線，我們將嘗試逼近它。我在下面定義了一些函數來做到這一點
def normalize(x):
return (x - np.min(x)) / (np.max(x) - np.min(x))
def create_noisy_spiral(n, jitter_std=0.2, revolutions=2):
angle = np.random.uniform(0, 2 * np.pi * revolutions, [n])
r = angle
x = r * np.cos(angle)
y = r * np.sin(angle)
result = np.stack([x, y], axis=1)
result = result + np.random.normal(scale=jitter_std, size=[n, 2])
result = 5 * normalize(result)
return result
接下來，讓我們多次調用此函數以建構一個樣本資料集
xy = create_noisy_spiral(10000)
x, y = xy[:, 0:1], xy[:, 1:]
plt.scatter(x, y)
plt.show()
如您所見，對於給定的 X，Y 有多個可能的值。普通的神經網路將簡單地學習這些點相對於幾何空間的平均值。
我們可以透過一個簡單的線性模型快速展示這一點
N_HIDDEN = 128
model = keras_core.Sequential(
[
layers.Dense(N_HIDDEN, activation="relu"),
layers.Dense(N_HIDDEN, activation="relu"),
layers.Dense(1),
]
)
讓我們使用均方誤差以及 adam 最佳化器。這些往往是合理的原型選擇
model.compile(optimizer="adam", loss="mse")
我們可以很容易地擬合這個模型
model.fit(
x,
y,
epochs=300,
batch_size=128,
validation_split=0.15,
callbacks=[callbacks.EarlyStopping(monitor="val_loss", patience=10)],
)
Epoch 1/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 3s 8ms/step - loss: 2.6971 - val_loss: 1.6366
Epoch 2/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.5672 - val_loss: 1.2341
Epoch 3/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.1751 - val_loss: 1.0113
Epoch 4/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0322 - val_loss: 1.0108
Epoch 5/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0625 - val_loss: 1.0212
Epoch 6/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0290 - val_loss: 1.0022
Epoch 7/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0469 - val_loss: 1.0033
Epoch 8/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0247 - val_loss: 1.0011
Epoch 9/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0313 - val_loss: 0.9997
Epoch 10/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0252 - val_loss: 0.9995
Epoch 11/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0369 - val_loss: 1.0015
Epoch 12/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0203 - val_loss: 0.9958
Epoch 13/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0305 - val_loss: 0.9960
Epoch 14/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0283 - val_loss: 1.0081
Epoch 15/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0331 - val_loss: 0.9943
Epoch 16/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: 1.0244 - val_loss: 1.0021
Epoch 17/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0496 - val_loss: 1.0077
Epoch 18/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0367 - val_loss: 0.9940
Epoch 19/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0201 - val_loss: 0.9927
Epoch 20/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0501 - val_loss: 1.0133
Epoch 21/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0098 - val_loss: 0.9980
Epoch 22/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0195 - val_loss: 0.9907
Epoch 23/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0196 - val_loss: 0.9899
Epoch 24/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0170 - val_loss: 1.0033
Epoch 25/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0169 - val_loss: 0.9963
Epoch 26/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0141 - val_loss: 0.9895
Epoch 27/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0367 - val_loss: 0.9916
Epoch 28/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0301 - val_loss: 0.9991
Epoch 29/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0097 - val_loss: 1.0004
Epoch 30/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0415 - val_loss: 1.0062
Epoch 31/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0186 - val_loss: 0.9888
Epoch 32/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0230 - val_loss: 0.9910
Epoch 33/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0217 - val_loss: 0.9910
Epoch 34/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0180 - val_loss: 0.9945
Epoch 35/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0329 - val_loss: 0.9963
Epoch 36/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0190 - val_loss: 0.9912
Epoch 37/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0341 - val_loss: 0.9894
Epoch 38/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0100 - val_loss: 0.9920
Epoch 39/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0097 - val_loss: 0.9899
Epoch 40/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0216 - val_loss: 0.9948
Epoch 41/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0115 - val_loss: 0.9923
<keras_core.src.callbacks.history.History at 0x12e0b4dd0>
讓我們看看結果
y_pred = model.predict(x)
313/313 ━━━━━━━━━━━━━━━━━━━━ 0s 851us/step
正如預期的那樣，該模型學習了給定 x 的所有 y 中點的幾何平均值。
plt.scatter(x, y)
plt.scatter(x, y_pred)
plt.show()
混合密度網路
混合密度網路可以緩解這個問題。混合密度是一種複雜密度的類別，可以用更簡單的密度來表示。它們實際上是許多機率分佈的總和。混合密度網路學習根據給定的訓練集參數化混合密度分佈。
作為一名從業者，您需要知道的是，混合密度網路解決了給定 X 的 Y 有多個值的問題。我希望為您的工具箱添加一個工具 - 但我不會在本指南中正式解釋混合密度網路的推導。最重要的是要知道混合密度網路學習參數化混合密度分佈。這是透過計算相對於提供的 y_i 標籤以及對應 x_i 的預測分佈的特殊損失來完成的。此損失函數透過計算 y_i 從預測的混合分佈中抽取的機率來運作。
讓我們實作一個混合密度網路。以下根據舊的 Keras 程式庫 Keras 混合密度網路層 定義了大量的輔助函數。
我已經調整程式碼以與 Keras core 一起使用。
讓我們開始編寫混合密度網路！首先，我們需要一個特殊的激活函數：ELU 加上一個很小的 epsilon。這有助於防止 ELU 輸出 0，這會導致混合密度網路損失評估中的 NaN。
def elu_plus_one_plus_epsilon(x):
return keras_core.activations.elu(x) + 1 + keras_core.backend.epsilon()
接下來，讓我們實際定義一個 MixtureDensity 層，該層輸出從學習的混合分佈中取樣所需的所有值
class MixtureDensityOutput(layers.Layer):
def __init__(self, output_dimension, num_mixtures, **kwargs):
super().__init__(**kwargs)
self.output_dim = output_dimension
self.num_mix = num_mixtures
self.mdn_mus = layers.Dense(
self.num_mix * self.output_dim, name="mdn_mus"
) # mix*output vals, no activation
self.mdn_sigmas = layers.Dense(
self.num_mix * self.output_dim,
activation=elu_plus_one_plus_epsilon,
name="mdn_sigmas",
) # mix*output vals exp activation
self.mdn_pi = layers.Dense(self.num_mix, name="mdn_pi") # mix vals, logits
def build(self, input_shape):
self.mdn_mus.build(input_shape)
self.mdn_sigmas.build(input_shape)
self.mdn_pi.build(input_shape)
super().build(input_shape)
@property
def trainable_weights(self):
return (
self.mdn_mus.trainable_weights
+ self.mdn_sigmas.trainable_weights
+ self.mdn_pi.trainable_weights
)
@property
def non_trainable_weights(self):
return (
self.mdn_mus.non_trainable_weights
+ self.mdn_sigmas.non_trainable_weights
+ self.mdn_pi.non_trainable_weights
)
def call(self, x, mask=None):
return layers.concatenate(
[self.mdn_mus(x), self.mdn_sigmas(x), self.mdn_pi(x)], name="mdn_outputs"
)
讓我們使用我們的新層建構一個混合密度網路
OUTPUT_DIMS = 1
N_MIXES = 20
mdn_network = keras_core.Sequential(
[
layers.Dense(N_HIDDEN, activation="relu"),
layers.Dense(N_HIDDEN, activation="relu"),
MixtureDensityOutput(OUTPUT_DIMS, N_MIXES),
]
)
接下來，讓我們根據真實值和我們的預期輸出，實作一個自訂損失函數來訓練混合密度網路層
def get_mixture_loss_func(output_dim, num_mixes):
def mdn_loss_func(y_true, y_pred):
# Reshape inputs in case this is used in a TimeDistributed layer
y_pred = tf.reshape(
y_pred,
[-1, (2 * num_mixes * output_dim) + num_mixes],
name="reshape_ypreds",
)
y_true = tf.reshape(y_true, [-1, output_dim], name="reshape_ytrue")
# Split the inputs into parameters
out_mu, out_sigma, out_pi = tf.split(
y_pred,
num_or_size_splits=[
num_mixes * output_dim,
num_mixes * output_dim,
num_mixes,
],
axis=-1,
name="mdn_coef_split",
)
# Construct the mixture models
cat = tfd.Categorical(logits=out_pi)
component_splits = [output_dim] * num_mixes
mus = tf.split(out_mu, num_or_size_splits=component_splits, axis=1)
sigs = tf.split(out_sigma, num_or_size_splits=component_splits, axis=1)
coll = [
tfd.MultivariateNormalDiag(loc=loc, scale_diag=scale)
for loc, scale in zip(mus, sigs)
]
mixture = tfd.Mixture(cat=cat, components=coll)
loss = mixture.log_prob(y_true)
loss = tf.negative(loss)
loss = tf.reduce_mean(loss)
return loss
return mdn_loss_func
mdn_network.compile(loss=get_mixture_loss_func(OUTPUT_DIMS, N_MIXES), optimizer="adam")
最後，我們可以像任何其他 Keras 模型一樣呼叫 model.fit()。
mdn_network.fit(
x,
y,
epochs=300,
batch_size=128,
validation_split=0.15,
callbacks=[
callbacks.EarlyStopping(monitor="loss", patience=10, restore_best_weights=True),
callbacks.ReduceLROnPlateau(monitor="loss", patience=5),
],
)
Epoch 1/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 20s 89ms/step - loss: 2.5088 - val_loss: 1.6384 - learning_rate: 0.0010
Epoch 2/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.5954 - val_loss: 1.4872 - learning_rate: 0.0010
Epoch 3/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.4818 - val_loss: 1.4026 - learning_rate: 0.0010
Epoch 4/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.3818 - val_loss: 1.3327 - learning_rate: 0.0010
Epoch 5/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.3478 - val_loss: 1.3034 - learning_rate: 0.0010
Epoch 6/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: 1.3045 - val_loss: 1.2684 - learning_rate: 0.0010
Epoch 7/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: 1.2836 - val_loss: 1.2381 - learning_rate: 0.0010
Epoch 8/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: 1.2582 - val_loss: 1.2047 - learning_rate: 0.0010
Epoch 9/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.2212 - val_loss: 1.1915 - learning_rate: 0.0010
Epoch 10/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.1907 - val_loss: 1.1903 - learning_rate: 0.0010
Epoch 11/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.1456 - val_loss: 1.0221 - learning_rate: 0.0010
Epoch 12/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 1.0075 - val_loss: 0.9356 - learning_rate: 0.0010
Epoch 13/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.9413 - val_loss: 0.8409 - learning_rate: 0.0010
Epoch 14/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: 0.8646 - val_loss: 0.8717 - learning_rate: 0.0010
Epoch 15/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: 0.8053 - val_loss: 0.8080 - learning_rate: 0.0010
Epoch 16/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.7568 - val_loss: 0.6381 - learning_rate: 0.0010
Epoch 17/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.6638 - val_loss: 0.6175 - learning_rate: 0.0010
Epoch 18/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.5893 - val_loss: 0.5387 - learning_rate: 0.0010
Epoch 19/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.5835 - val_loss: 0.5449 - learning_rate: 0.0010
Epoch 20/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 4ms/step - loss: 0.5137 - val_loss: 0.4536 - learning_rate: 0.0010
Epoch 21/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: 0.4808 - val_loss: 0.4779 - learning_rate: 0.0010
Epoch 22/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.4592 - val_loss: 0.4359 - learning_rate: 0.0010
Epoch 23/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.4303 - val_loss: 0.4768 - learning_rate: 0.0010
Epoch 24/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.4505 - val_loss: 0.4084 - learning_rate: 0.0010
Epoch 25/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.4033 - val_loss: 0.3484 - learning_rate: 0.0010
Epoch 26/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.3696 - val_loss: 0.4844 - learning_rate: 0.0010
Epoch 27/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.3868 - val_loss: 0.3406 - learning_rate: 0.0010
Epoch 28/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.3214 - val_loss: 0.2739 - learning_rate: 0.0010
Epoch 29/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.3154 - val_loss: 0.3286 - learning_rate: 0.0010
Epoch 30/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.2930 - val_loss: 0.2263 - learning_rate: 0.0010
Epoch 31/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.2946 - val_loss: 0.2927 - learning_rate: 0.0010
Epoch 32/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.2739 - val_loss: 0.2026 - learning_rate: 0.0010
Epoch 33/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.2454 - val_loss: 0.2451 - learning_rate: 0.0010
Epoch 34/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.2146 - val_loss: 0.1722 - learning_rate: 0.0010
Epoch 35/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.2041 - val_loss: 0.2774 - learning_rate: 0.0010
Epoch 36/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.2020 - val_loss: 0.1257 - learning_rate: 0.0010
Epoch 37/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.1614 - val_loss: 0.1128 - learning_rate: 0.0010
Epoch 38/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.1676 - val_loss: 0.1908 - learning_rate: 0.0010
Epoch 39/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: 0.1511 - val_loss: 0.1045 - learning_rate: 0.0010
Epoch 40/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: 0.1061 - val_loss: 0.1321 - learning_rate: 0.0010
Epoch 41/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.1170 - val_loss: 0.0879 - learning_rate: 0.0010
Epoch 42/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.1045 - val_loss: 0.0307 - learning_rate: 0.0010
Epoch 43/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.1066 - val_loss: 0.0637 - learning_rate: 0.0010
Epoch 44/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0960 - val_loss: 0.0304 - learning_rate: 0.0010
Epoch 45/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0747 - val_loss: 0.0211 - learning_rate: 0.0010
Epoch 46/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0733 - val_loss: -0.0155 - learning_rate: 0.0010
Epoch 47/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0339 - val_loss: 0.0079 - learning_rate: 0.0010
Epoch 48/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0597 - val_loss: 0.0223 - learning_rate: 0.0010
Epoch 49/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0370 - val_loss: 0.0549 - learning_rate: 0.0010
Epoch 50/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0343 - val_loss: 0.0031 - learning_rate: 0.0010
Epoch 51/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0132 - val_loss: -0.0304 - learning_rate: 0.0010
Epoch 52/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0326 - val_loss: 0.0584 - learning_rate: 0.0010
Epoch 53/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0512 - val_loss: -0.0166 - learning_rate: 0.0010
Epoch 54/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0210 - val_loss: -0.0433 - learning_rate: 0.0010
Epoch 55/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0261 - val_loss: 0.0317 - learning_rate: 0.0010
Epoch 56/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0185 - val_loss: -0.0210 - learning_rate: 0.0010
Epoch 57/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0021 - val_loss: -0.0218 - learning_rate: 0.0010
Epoch 58/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0100 - val_loss: -0.0488 - learning_rate: 0.0010
Epoch 59/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0126 - val_loss: -0.0504 - learning_rate: 0.0010
Epoch 60/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0278 - val_loss: -0.0622 - learning_rate: 0.0010
Epoch 61/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0180 - val_loss: -0.0756 - learning_rate: 0.0010
Epoch 62/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: -0.0198 - val_loss: -0.0427 - learning_rate: 0.0010
Epoch 63/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0129 - val_loss: -0.0483 - learning_rate: 0.0010
Epoch 64/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0221 - val_loss: -0.0379 - learning_rate: 0.0010
Epoch 65/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: -0.0177 - val_loss: -0.0626 - learning_rate: 0.0010
Epoch 66/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0045 - val_loss: -0.0148 - learning_rate: 0.0010
Epoch 67/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0045 - val_loss: -0.0570 - learning_rate: 0.0010
Epoch 68/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0304 - val_loss: -0.0062 - learning_rate: 0.0010
Epoch 69/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: 0.0053 - val_loss: -0.0553 - learning_rate: 0.0010
Epoch 70/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0364 - val_loss: -0.1112 - learning_rate: 0.0010
Epoch 71/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0017 - val_loss: -0.0865 - learning_rate: 0.0010
Epoch 72/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0082 - val_loss: -0.1180 - learning_rate: 0.0010
Epoch 73/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0501 - val_loss: -0.1028 - learning_rate: 0.0010
Epoch 74/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0452 - val_loss: -0.0381 - learning_rate: 0.0010
Epoch 75/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0397 - val_loss: -0.0517 - learning_rate: 0.0010
Epoch 76/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0317 - val_loss: -0.1144 - learning_rate: 0.0010
Epoch 77/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0400 - val_loss: -0.1283 - learning_rate: 0.0010
Epoch 78/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0756 - val_loss: -0.0749 - learning_rate: 0.0010
Epoch 79/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0459 - val_loss: -0.1229 - learning_rate: 0.0010
Epoch 80/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0485 - val_loss: -0.0896 - learning_rate: 0.0010
Epoch 81/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: -0.0351 - val_loss: -0.1037 - learning_rate: 0.0010
Epoch 82/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0617 - val_loss: -0.0949 - learning_rate: 0.0010
Epoch 83/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0614 - val_loss: -0.1044 - learning_rate: 0.0010
Epoch 84/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0650 - val_loss: -0.1128 - learning_rate: 0.0010
Epoch 85/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0710 - val_loss: -0.1236 - learning_rate: 0.0010
Epoch 86/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0504 - val_loss: -0.0149 - learning_rate: 0.0010
Epoch 87/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0561 - val_loss: -0.1095 - learning_rate: 0.0010
Epoch 88/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0527 - val_loss: -0.0929 - learning_rate: 0.0010
Epoch 89/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0704 - val_loss: -0.1062 - learning_rate: 0.0010
Epoch 90/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.0386 - val_loss: -0.1433 - learning_rate: 0.0010
Epoch 91/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1129 - val_loss: -0.1698 - learning_rate: 1.0000e-04
Epoch 92/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1210 - val_loss: -0.1696 - learning_rate: 1.0000e-04
Epoch 93/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1315 - val_loss: -0.1663 - learning_rate: 1.0000e-04
Epoch 94/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1207 - val_loss: -0.1696 - learning_rate: 1.0000e-04
Epoch 95/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1208 - val_loss: -0.1606 - learning_rate: 1.0000e-04
Epoch 96/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1157 - val_loss: -0.1728 - learning_rate: 1.0000e-04
Epoch 97/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1367 - val_loss: -0.1691 - learning_rate: 1.0000e-04
Epoch 98/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1237 - val_loss: -0.1740 - learning_rate: 1.0000e-04
Epoch 99/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1271 - val_loss: -0.1593 - learning_rate: 1.0000e-04
Epoch 100/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1358 - val_loss: -0.1738 - learning_rate: 1.0000e-04
Epoch 101/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1260 - val_loss: -0.1669 - learning_rate: 1.0000e-04
Epoch 102/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1184 - val_loss: -0.1660 - learning_rate: 1.0000e-04
Epoch 103/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1221 - val_loss: -0.1740 - learning_rate: 1.0000e-04
Epoch 104/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1207 - val_loss: -0.1498 - learning_rate: 1.0000e-04
Epoch 105/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1210 - val_loss: -0.1695 - learning_rate: 1.0000e-04
Epoch 106/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1264 - val_loss: -0.1477 - learning_rate: 1.0000e-04
Epoch 107/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1217 - val_loss: -0.1717 - learning_rate: 1.0000e-04
Epoch 108/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1182 - val_loss: -0.1748 - learning_rate: 1.0000e-05
Epoch 109/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1394 - val_loss: -0.1757 - learning_rate: 1.0000e-05
Epoch 110/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1363 - val_loss: -0.1762 - learning_rate: 1.0000e-05
Epoch 111/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1292 - val_loss: -0.1765 - learning_rate: 1.0000e-05
Epoch 112/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1330 - val_loss: -0.1737 - learning_rate: 1.0000e-05
Epoch 113/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1341 - val_loss: -0.1769 - learning_rate: 1.0000e-05
Epoch 114/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1318 - val_loss: -0.1771 - learning_rate: 1.0000e-05
Epoch 115/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1285 - val_loss: -0.1756 - learning_rate: 1.0000e-05
Epoch 116/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1211 - val_loss: -0.1764 - learning_rate: 1.0000e-05
Epoch 117/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1434 - val_loss: -0.1755 - learning_rate: 1.0000e-05
Epoch 118/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 1s 2ms/step - loss: -0.1375 - val_loss: -0.1757 - learning_rate: 1.0000e-05
Epoch 119/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1407 - val_loss: -0.1740 - learning_rate: 1.0000e-05
Epoch 120/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1406 - val_loss: -0.1754 - learning_rate: 1.0000e-06
Epoch 121/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1258 - val_loss: -0.1761 - learning_rate: 1.0000e-06
Epoch 122/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1384 - val_loss: -0.1762 - learning_rate: 1.0000e-06
Epoch 123/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1522 - val_loss: -0.1764 - learning_rate: 1.0000e-06
Epoch 124/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1310 - val_loss: -0.1763 - learning_rate: 1.0000e-06
Epoch 125/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1434 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 126/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1329 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 127/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1392 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 128/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1300 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 129/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: -0.1347 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 130/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: -0.1200 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 131/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: -0.1415 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 132/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1270 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 133/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1329 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 134/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1265 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 135/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1329 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 136/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1429 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 137/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1394 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 138/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1315 - val_loss: -0.1763 - learning_rate: 1.0000e-07
Epoch 139/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1253 - val_loss: -0.1763 - learning_rate: 1.0000e-08
Epoch 140/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1346 - val_loss: -0.1763 - learning_rate: 1.0000e-08
Epoch 141/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1418 - val_loss: -0.1763 - learning_rate: 1.0000e-08
Epoch 142/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 2ms/step - loss: -0.1279 - val_loss: -0.1763 - learning_rate: 1.0000e-08
Epoch 143/300
67/67 ━━━━━━━━━━━━━━━━━━━━ 0s 3ms/step - loss: -0.1224 - val_loss: -0.1763 - learning_rate: 1.0000e-08
<keras_core.src.callbacks.history.History at 0x148c20890>
讓我們做一些預測！
y_pred_mixture = mdn_network.predict(x)
print(y_pred_mixture.shape)
313/313 ━━━━━━━━━━━━━━━━━━━━ 0s 811us/step
(10000, 60)
MDN 不會輸出單個值；而是輸出值以參數化混合分佈。為了視覺化這些輸出，讓我們從分佈中取樣。
請注意，取樣是一個有損的過程。如果您想保留所有資訊作為更大的潛在表示的一部分（即用於下游處理），我建議您直接保留分佈參數。
def split_mixture_params(params, output_dim, num_mixes):
mus = params[: num_mixes * output_dim]
sigs = params[num_mixes * output_dim : 2 * num_mixes * output_dim]
pi_logits = params[-num_mixes:]
return mus, sigs, pi_logits
def softmax(w, t=1.0):
e = np.array(w) / t # adjust temperature
e -= e.max() # subtract max to protect from exploding exp values.
e = np.exp(e)
dist = e / np.sum(e)
return dist
def sample_from_categorical(dist):
r = np.random.rand(1) # uniform random number in [0,1]
accumulate = 0
for i in range(0, dist.size):
accumulate += dist[i]
if accumulate >= r:
return i
tf.logging.info("Error sampling categorical model.")
return -1
def sample_from_output(params, output_dim, num_mixes, temp=1.0, sigma_temp=1.0):
mus, sigs, pi_logits = split_mixture_params(params, output_dim, num_mixes)
pis = softmax(pi_logits, t=temp)
m = sample_from_categorical(pis)
# Alternative way to sample from categorical:
# m = np.random.choice(range(len(pis)), p=pis)
mus_vector = mus[m * output_dim : (m + 1) * output_dim]
sig_vector = sigs[m * output_dim : (m + 1) * output_dim]
scale_matrix = np.identity(output_dim) * sig_vector # scale matrix from diag
cov_matrix = np.matmul(scale_matrix, scale_matrix.T) # cov is scale squared.
cov_matrix = cov_matrix * sigma_temp # adjust for sigma temperature
sample = np.random.multivariate_normal(mus_vector, cov_matrix, 1)
return sample
接下來讓我們使用我們的取樣函數
# Sample from the predicted distributions
y_samples = np.apply_along_axis(
sample_from_output, 1, y_pred_mixture, 1, N_MIXES, temp=1.0
)
最後，我們可以視覺化我們的網路輸出
plt.scatter(x, y, alpha=0.05, color="blue", label="Ground Truth")
plt.scatter(
x,
y_samples[:, :, 0],
color="green",
alpha=0.05,
label="Mixture Density Network prediction",
)
plt.show()
太棒了。很樂意看到它
結論
神經網路是通用的函數逼近器 - 但它們只能逼近函數。混合密度網路可以使用一些巧妙的機率技巧來逼近任意 x->y 映射。
如需更多 tensorflow_probability 的範例，請從這裡開始。
再一張漂亮的圖供您參考
fig, axs = plt.subplots(1, 3)
fig.set_figheight(3)
fig.set_figwidth(12)
axs[0].set_title("Ground Truth")
axs[0].scatter(x, y, alpha=0.05, color="blue")
xlim = axs[0].get_xlim()
ylim = axs[0].get_ylim()
axs[1].set_title("Normal Model prediction")
axs[1].scatter(x, y_pred, alpha=0.05, color="red")
axs[1].set_xlim(xlim)
axs[1].set_ylim(ylim)
axs[2].scatter(
x,
y_samples[:, :, 0],
color="green",
alpha=0.05,
label="Mixture Density Network prediction",
)
axs[2].set_title("Mixture Density Network prediction")
axs[2].set_xlim(xlim)
axs[2].set_ylim(ylim)
plt.show()
