TensorFlow Probability의 옵티마이저

TensorFlow.org에서 보기

Google Colab에서 실행

GitHub에서 소스 보기

노트북 다운로드

추상적 인

이 colab에서는 TensorFlow Probability에 구현된 다양한 최적화 프로그램을 사용하는 방법을 보여줍니다.

종속성 및 전제 조건

수입

%matplotlib inline


import contextlib
import functools
import os
import time

import numpy as np
import pandas as pd
import scipy as sp
from six.moves import urllib
from sklearn import preprocessing

import tensorflow.compat.v2 as tf
tf.enable_v2_behavior()

import tensorflow_probability as tfp

BFGS 및 L-BFGS 옵티마이저

Quasi Newton 방법은 널리 사용되는 1차 최적화 알고리즘 클래스입니다. 이러한 방법은 정확한 헤세 행렬에 대한 양의 정부호 근사치를 사용하여 검색 방향을 찾습니다.

Broyden - 플레처 - 골드 파브 - Shanno 알고리즘 ( BFGS는 )이 일반적인 생각의 특정 구현입니다. 그것은 적용되며, 그라데이션 사방 연속 중간 크기의 문제에 대한 선택의 방법 (AN으로 예를 들면 선형 회귀 \(L_2\) 벌금).

L-BFGS는 그 헤 시안 매트릭스 합리적인 비용으로 계산 될 수 없거나 희소 수없는 큰 문제점을 해결하기위한 유용한 BFGS 제한된 메모리 버전이다. 대신에 완전히 조밀 저장하는 \(n \times n\) 헤 시안 행렬의 근사치를, 그들은 단지 길이의 몇 벡터 저장 \(n\) 암묵적으로 이러한 근사치를 나타냅니다.

도우미 기능

CACHE_DIR = os.path.join(os.sep, 'tmp', 'datasets')


def make_val_and_grad_fn(value_fn):
  @functools.wraps(value_fn)
  def val_and_grad(x):
    return tfp.math.value_and_gradient(value_fn, x)
  return val_and_grad


@contextlib.contextmanager
def timed_execution():
  t0 = time.time()
  yield
  dt = time.time() - t0
  print('Evaluation took: %f seconds' % dt)


def np_value(tensor):
  """Get numpy value out of possibly nested tuple of tensors."""
  if isinstance(tensor, tuple):
    return type(tensor)(*(np_value(t) for t in tensor))
  else:
    return tensor.numpy()

def run(optimizer):
  """Run an optimizer and measure it's evaluation time."""
  optimizer()  # Warmup.
  with timed_execution():
    result = optimizer()
  return np_value(result)

단순 이차 함수에 대한 L-BFGS

# Fix numpy seed for reproducibility
np.random.seed(12345)

# The objective must be supplied as a function that takes a single
# (Tensor) argument and returns a tuple. The first component of the
# tuple is the value of the objective at the supplied point and the
# second value is the gradient at the supplied point. The value must
# be a scalar and the gradient must have the same shape as the
# supplied argument.

# The `make_val_and_grad_fn` decorator helps transforming a function
# returning the objective value into one that returns both the gradient
# and the value. It also works for both eager and graph mode.

dim = 10
minimum = np.ones([dim])
scales = np.exp(np.random.randn(dim))

@make_val_and_grad_fn
def quadratic(x):
  return tf.reduce_sum(scales * (x - minimum) ** 2, axis=-1)

# The minimization routine also requires you to supply an initial
# starting point for the search. For this example we choose a random
# starting point.
start = np.random.randn(dim)

# Finally an optional argument called tolerance let's you choose the
# stopping point of the search. The tolerance specifies the maximum
# (supremum) norm of the gradient vector at which the algorithm terminates.
# If you don't have a specific need for higher or lower accuracy, leaving
# this parameter unspecified (and hence using the default value of 1e-8)
# should be good enough.
tolerance = 1e-10

@tf.function
def quadratic_with_lbfgs():
  return tfp.optimizer.lbfgs_minimize(
    quadratic,
    initial_position=tf.constant(start),
    tolerance=tolerance)

results = run(quadratic_with_lbfgs)

# The optimization results contain multiple pieces of information. The most
# important fields are: 'converged' and 'position'.
# Converged is a boolean scalar tensor. As the name implies, it indicates
# whether the norm of the gradient at the final point was within tolerance.
# Position is the location of the minimum found. It is important to check
# that converged is True before using the value of the position.

print('L-BFGS Results')
print('Converged:', results.converged)
print('Location of the minimum:', results.position)
print('Number of iterations:', results.num_iterations)

Evaluation took: 0.014586 seconds
L-BFGS Results
Converged: True
Location of the minimum: [1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
Number of iterations: 10

BFGS와 동일한 문제

@tf.function
def quadratic_with_bfgs():
  return tfp.optimizer.bfgs_minimize(
    quadratic,
    initial_position=tf.constant(start),
    tolerance=tolerance)

results = run(quadratic_with_bfgs)

print('BFGS Results')
print('Converged:', results.converged)
print('Location of the minimum:', results.position)
print('Number of iterations:', results.num_iterations)

Evaluation took: 0.010468 seconds
BFGS Results
Converged: True
Location of the minimum: [1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
Number of iterations: 10

L1 패널티가 있는 선형 회귀: 전립선암 데이터

트레버 해 스티, 로버트 팁쉬 라니와 제롬 프리드먼에 의해 통계 학습, 데이터 마이닝, 추론 및 예측의 요소 : 책에서 예.

이것은 L1 패널티의 최적화 문제입니다.

데이터세트 가져오기

def cache_or_download_file(cache_dir, url_base, filename):
  """Read a cached file or download it."""
  filepath = os.path.join(cache_dir, filename)
  if tf.io.gfile.exists(filepath):
    return filepath
  if not tf.io.gfile.exists(cache_dir):
    tf.io.gfile.makedirs(cache_dir)
  url = url_base + filename
  print("Downloading {url} to {filepath}.".format(url=url, filepath=filepath))
  urllib.request.urlretrieve(url, filepath)
  return filepath

def get_prostate_dataset(cache_dir=CACHE_DIR):
  """Download the prostate dataset and read as Pandas dataframe."""
  url_base = 'http://web.stanford.edu/~hastie/ElemStatLearn/datasets/'
  return pd.read_csv(
      cache_or_download_file(cache_dir, url_base, 'prostate.data'),
      delim_whitespace=True, index_col=0)

prostate_df = get_prostate_dataset()

Downloading http://web.stanford.edu/~hastie/ElemStatLearn/datasets/prostate.data to /tmp/datasets/prostate.data.

문제 정의

np.random.seed(12345)

feature_names = ['lcavol', 'lweight',   'age',  'lbph', 'svi', 'lcp',   
                 'gleason', 'pgg45']

# Normalize features
scalar = preprocessing.StandardScaler()
prostate_df[feature_names] = pd.DataFrame(
    scalar.fit_transform(
        prostate_df[feature_names].astype('float64')))

# select training set
prostate_df_train = prostate_df[prostate_df.train == 'T'] 

# Select features and labels 
features = prostate_df_train[feature_names]
labels =  prostate_df_train[['lpsa']]

# Create tensors
feat = tf.constant(features.values, dtype=tf.float64)
lab = tf.constant(labels.values, dtype=tf.float64)

dtype = feat.dtype

regularization = 0 # regularization parameter
dim = 8 # number of features

# We pick a random starting point for the search
start = np.random.randn(dim + 1)

def regression_loss(params):
  """Compute loss for linear regression model with L1 penalty

  Args:
    params: A real tensor of shape [dim + 1]. The zeroth component
      is the intercept term and the rest of the components are the
      beta coefficients.

  Returns:
    The mean square error loss including L1 penalty.
  """
  params = tf.squeeze(params)
  intercept, beta  = params[0], params[1:]
  pred = tf.matmul(feat, tf.expand_dims(beta, axis=-1)) + intercept
  mse_loss = tf.reduce_sum(
      tf.cast(
        tf.losses.mean_squared_error(y_true=lab, y_pred=pred), tf.float64))
  l1_penalty = regularization * tf.reduce_sum(tf.abs(beta))
  total_loss = mse_loss + l1_penalty
  return total_loss

L-BFGS로 풀기

L-BFGS를 사용하여 맞춤. L1 패널티가 파생 불연속성을 도입하더라도 실제로 L-BFGS는 여전히 잘 작동합니다.

@tf.function
def l1_regression_with_lbfgs():
  return tfp.optimizer.lbfgs_minimize(
    make_val_and_grad_fn(regression_loss),
    initial_position=tf.constant(start),
    tolerance=1e-8)

results = run(l1_regression_with_lbfgs)
minimum = results.position
fitted_intercept = minimum[0]
fitted_beta = minimum[1:]

print('L-BFGS Results')
print('Converged:', results.converged)
print('Intercept: Fitted ({})'.format(fitted_intercept))
print('Beta:      Fitted {}'.format(fitted_beta))

Evaluation took: 0.017987 seconds
L-BFGS Results
Converged: True
Intercept: Fitted (2.3879985744556484)
Beta:      Fitted [ 0.68626215  0.28193532 -0.17030254  0.10799274  0.33634988 -0.24888523
  0.11992237  0.08689026]

넬더 미드로 해결하기

Nelder 미드 방법은 가장 인기 파생 무료 최소화 방법 중 하나입니다. 이 최적화 프로그램은 기울기 정보를 사용하지 않으며 대상 함수의 미분 가능성을 가정하지 않습니다. 따라서 L1 패널티가 있는 최적화 문제와 같이 평활하지 않은 목적 함수에 적합합니다.

의 최적화 문제를 들어 \(n\)-dimensions는 일련의 유지\(n+1\) 비 퇴화 한 단면에 걸쳐 후보 솔루션을. 각 정점의 함수 값을 사용하여 일련의 움직임(반사, 확장, 수축 및 수축)을 기반으로 심플렉스를 연속적으로 수정합니다.

# Nelder mead expects an initial_vertex of shape [n + 1, 1].
initial_vertex = tf.expand_dims(tf.constant(start, dtype=dtype), axis=-1)

@tf.function
def l1_regression_with_nelder_mead():
  return tfp.optimizer.nelder_mead_minimize(
      regression_loss,
      initial_vertex=initial_vertex,
      func_tolerance=1e-10,
      position_tolerance=1e-10)

results = run(l1_regression_with_nelder_mead)
minimum = results.position.reshape([-1])
fitted_intercept = minimum[0]
fitted_beta = minimum[1:]

print('Nelder Mead Results')
print('Converged:', results.converged)
print('Intercept: Fitted ({})'.format(fitted_intercept))
print('Beta:      Fitted {}'.format(fitted_beta))

Evaluation took: 0.325643 seconds
Nelder Mead Results
Converged: True
Intercept: Fitted (2.387998456121595)
Beta:      Fitted [ 0.68626266  0.28193456 -0.17030291  0.10799375  0.33635132 -0.24888703
  0.11992244  0.08689023]

L2 패널티를 사용한 로지스틱 회귀

이 예에서는 분류를 위한 합성 데이터 세트를 만들고 L-BFGS 옵티마이저를 사용하여 매개변수에 맞춥니다.

np.random.seed(12345)

dim = 5  # The number of features
n_obs = 10000  # The number of observations

betas = np.random.randn(dim)  # The true beta
intercept = np.random.randn()  # The true intercept

features = np.random.randn(n_obs, dim)  # The feature matrix
probs = sp.special.expit(
    np.matmul(features, np.expand_dims(betas, -1)) + intercept)

labels = sp.stats.bernoulli.rvs(probs)  # The true labels

regularization = 0.8
feat = tf.constant(features)
lab = tf.constant(labels, dtype=feat.dtype)

@make_val_and_grad_fn
def negative_log_likelihood(params):
  """Negative log likelihood for logistic model with L2 penalty

  Args:
    params: A real tensor of shape [dim + 1]. The zeroth component
      is the intercept term and the rest of the components are the
      beta coefficients.

  Returns:
    The negative log likelihood plus the penalty term. 
  """
  intercept, beta  = params[0], params[1:]
  logit = tf.matmul(feat, tf.expand_dims(beta, -1)) + intercept
  log_likelihood = tf.reduce_sum(tf.nn.sigmoid_cross_entropy_with_logits(
      labels=lab, logits=logit))
  l2_penalty = regularization * tf.reduce_sum(beta ** 2)
  total_loss = log_likelihood + l2_penalty
  return total_loss

start = np.random.randn(dim + 1)

@tf.function
def l2_regression_with_lbfgs():
  return tfp.optimizer.lbfgs_minimize(
      negative_log_likelihood,
      initial_position=tf.constant(start),
      tolerance=1e-8)

results = run(l2_regression_with_lbfgs)
minimum = results.position
fitted_intercept = minimum[0]
fitted_beta = minimum[1:]

print('Converged:', results.converged)
print('Intercept: Fitted ({}), Actual ({})'.format(fitted_intercept, intercept))
print('Beta:\n\tFitted {},\n\tActual {}'.format(fitted_beta, betas))

Evaluation took: 0.056751 seconds
Converged: True
Intercept: Fitted (1.4111415084244365), Actual (1.3934058329729904)
Beta:
    Fitted [-0.18016612  0.53121578 -0.56420632 -0.5336374   2.00499675],
    Actual [-0.20470766  0.47894334 -0.51943872 -0.5557303   1.96578057]

일괄 지원

BFGS와 L-BFGS는 모두 일괄 계산을 지원합니다. 예를 들어 다양한 시작점에서 단일 함수를 최적화하기 위한 것입니다. 또는 단일 지점에서 여러 매개변수 함수.

단일 기능, 다중 시작점

Himmelblau의 기능은 표준 최적화 테스트 케이스입니다. 함수는 다음과 같이 제공됩니다.

\[f(x, y) = (x^2 + y - 11)^2 + (x + y^2 - 7)^2\]

이 함수에는 다음 위치에 있는 4개의 최소값이 있습니다.

• (3, 2),
• (-2.805118, 3.131312),
• (-3.779310, -3.283186),
• (3.584428, -1.848126).

이 모든 최소값은 적절한 시작점에서 도달할 수 있습니다.

# The function to minimize must take as input a tensor of shape [..., n]. In
# this n=2 is the size of the domain of the input and [...] are batching
# dimensions. The return value must be of shape [...], i.e. a batch of scalars
# with the objective value of the function evaluated at each input point.

@make_val_and_grad_fn
def himmelblau(coord):
  x, y = coord[..., 0], coord[..., 1]
  return (x * x + y - 11) ** 2 + (x + y * y - 7) ** 2

starts = tf.constant([[1, 1],
                      [-2, 2],
                      [-1, -1],
                      [1, -2]], dtype='float64')

# The stopping_condition allows to further specify when should the search stop.
# The default, tfp.optimizer.converged_all, will proceed until all points have
# either converged or failed. There is also a tfp.optimizer.converged_any to
# stop as soon as the first point converges, or all have failed.

@tf.function
def batch_multiple_starts():
  return tfp.optimizer.lbfgs_minimize(
      himmelblau, initial_position=starts,
      stopping_condition=tfp.optimizer.converged_all,
      tolerance=1e-8)

results = run(batch_multiple_starts)
print('Converged:', results.converged)
print('Minima:', results.position)

Evaluation took: 0.019095 seconds
Converged: [ True  True  True  True]
Minima: [[ 3.          2.        ]
 [-2.80511809  3.13131252]
 [-3.77931025 -3.28318599]
 [ 3.58442834 -1.84812653]]

여러 기능

데모 목적으로 이 예에서 우리는 동시에 많은 수의 고차원 무작위 생성 2차 보울을 최적화합니다.

np.random.seed(12345)

dim = 100
batches = 500
minimum = np.random.randn(batches, dim)
scales = np.exp(np.random.randn(batches, dim))

@make_val_and_grad_fn
def quadratic(x):
  return tf.reduce_sum(input_tensor=scales * (x - minimum)**2, axis=-1)

# Make all starting points (1, 1, ..., 1). Note not all starting points need
# to be the same.
start = tf.ones((batches, dim), dtype='float64')

@tf.function
def batch_multiple_functions():
  return tfp.optimizer.lbfgs_minimize(
      quadratic, initial_position=start,
      stopping_condition=tfp.optimizer.converged_all,
      max_iterations=100,
      tolerance=1e-8)

results = run(batch_multiple_functions)
print('All converged:', np.all(results.converged))
print('Largest error:', np.max(results.position - minimum))

Evaluation took: 1.994132 seconds
All converged: True
Largest error: 4.4131473142527966e-08