premade_models.ipynb

Kernel: Python 3

Copyright 2020 The TensorFlow Authors.

#@title Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# https://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

Modelos pré-fabricados do TF Lattice

Ver em TensorFlow.org

Executar no Google Colab

Ver fonte no GitHub

Baixar notebook

Visão geral

Os modelos pré-fabricados são maneiras rápidas e fáceis de criar instâncias do tf.keras.model para casos de uso típicos. Este guia descreve as etapas necessárias para criar e treinar/testar um modelo pré-fabricado do TFL.

Configuração

Instale o pacote do TF Lattice:

#@test {"skip": true}
!pip install tensorflow-lattice pydot

Importe os pacotes necessários:

import tensorflow as tf

import copy
import logging
import numpy as np
import pandas as pd
import sys
import tensorflow_lattice as tfl
logging.disable(sys.maxsize)

Configure os valores padrão usados para treinamento neste guia:

LEARNING_RATE = 0.01
BATCH_SIZE = 128
NUM_EPOCHS = 500
PREFITTING_NUM_EPOCHS = 10

Baixe o dataset UCI Statlog (Heart):

heart_csv_file = tf.keras.utils.get_file(
    'heart.csv',
    'http://storage.googleapis.com/download.tensorflow.org/data/heart.csv')
heart_df = pd.read_csv(heart_csv_file)
thal_vocab_list = ['normal', 'fixed', 'reversible']
heart_df['thal'] = heart_df['thal'].map(
    {v: i for i, v in enumerate(thal_vocab_list)})
heart_df = heart_df.astype(float)

heart_train_size = int(len(heart_df) * 0.8)
heart_train_dict = dict(heart_df[:heart_train_size])
heart_test_dict = dict(heart_df[heart_train_size:])

# This ordering of input features should match the feature configs. If no
# feature config relies explicitly on the data (i.e. all are 'quantiles'),
# then you can construct the feature_names list by simply iterating over each
# feature config and extracting it's name.
feature_names = [
    'age', 'sex', 'cp', 'chol', 'fbs', 'trestbps', 'thalach', 'restecg',
    'exang', 'oldpeak', 'slope', 'ca', 'thal'
]

# Since we have some features that manually construct their input keypoints,
# we need an index mapping of the feature names.
feature_name_indices = {name: index for index, name in enumerate(feature_names)}

label_name = 'target'
heart_train_xs = [
    heart_train_dict[feature_name] for feature_name in feature_names
]
heart_test_xs = [heart_test_dict[feature_name] for feature_name in feature_names]
heart_train_ys = heart_train_dict[label_name]
heart_test_ys = heart_test_dict[label_name]

Configurações de características

A calibração de características e as configurações por característica são definidas usando tfl.configs.FeatureConfig. As configurações de características incluem restrições de monotonicidade, regularização por recurso (ver tfl.configs.RegularizerConfig) e tamanhos de lattice para modelos lattice.

É necessário especificar totalmente a configuração para qualquer característica que queremos que o modelo reconheça. Caso contrário, o modelo não terá como saber se essa característica existe.

Defina as configurações das características

Agora que os quantis podem ser computados, defina uma configuração para cada característica que o modelo deve aceitar como entrada.

# Features:
# - age
# - sex
# - cp        chest pain type (4 values)
# - trestbps  resting blood pressure
# - chol      serum cholestoral in mg/dl
# - fbs       fasting blood sugar > 120 mg/dl
# - restecg   resting electrocardiographic results (values 0,1,2)
# - thalach   maximum heart rate achieved
# - exang     exercise induced angina
# - oldpeak   ST depression induced by exercise relative to rest
# - slope     the slope of the peak exercise ST segment
# - ca        number of major vessels (0-3) colored by flourosopy
# - thal      normal; fixed defect; reversable defect
#
# Feature configs are used to specify how each feature is calibrated and used.
heart_feature_configs = [
    tfl.configs.FeatureConfig(
        name='age',
        lattice_size=3,
        monotonicity='increasing',
        # We must set the keypoints manually.
        pwl_calibration_num_keypoints=5,
        pwl_calibration_input_keypoints='quantiles',
        pwl_calibration_clip_max=100,
        # Per feature regularization.
        regularizer_configs=[
            tfl.configs.RegularizerConfig(name='calib_wrinkle', l2=0.1),
        ],
    ),
    tfl.configs.FeatureConfig(
        name='sex',
        num_buckets=2,
    ),
    tfl.configs.FeatureConfig(
        name='cp',
        monotonicity='increasing',
        # Keypoints that are uniformly spaced.
        pwl_calibration_num_keypoints=4,
        pwl_calibration_input_keypoints=np.linspace(
            np.min(heart_train_xs[feature_name_indices['cp']]),
            np.max(heart_train_xs[feature_name_indices['cp']]),
            num=4),
    ),
    tfl.configs.FeatureConfig(
        name='chol',
        monotonicity='increasing',
        # Explicit input keypoints initialization.
        pwl_calibration_input_keypoints=[126.0, 210.0, 247.0, 286.0, 564.0],
        # Calibration can be forced to span the full output range by clamping.
        pwl_calibration_clamp_min=True,
        pwl_calibration_clamp_max=True,
        # Per feature regularization.
        regularizer_configs=[
            tfl.configs.RegularizerConfig(name='calib_hessian', l2=1e-4),
        ],
    ),
    tfl.configs.FeatureConfig(
        name='fbs',
        # Partial monotonicity: output(0) <= output(1)
        monotonicity=[(0, 1)],
        num_buckets=2,
    ),
    tfl.configs.FeatureConfig(
        name='trestbps',
        monotonicity='decreasing',
        pwl_calibration_num_keypoints=5,
        pwl_calibration_input_keypoints='quantiles',
    ),
    tfl.configs.FeatureConfig(
        name='thalach',
        monotonicity='decreasing',
        pwl_calibration_num_keypoints=5,
        pwl_calibration_input_keypoints='quantiles',
    ),
    tfl.configs.FeatureConfig(
        name='restecg',
        # Partial monotonicity: output(0) <= output(1), output(0) <= output(2)
        monotonicity=[(0, 1), (0, 2)],
        num_buckets=3,
    ),
    tfl.configs.FeatureConfig(
        name='exang',
        # Partial monotonicity: output(0) <= output(1)
        monotonicity=[(0, 1)],
        num_buckets=2,
    ),
    tfl.configs.FeatureConfig(
        name='oldpeak',
        monotonicity='increasing',
        pwl_calibration_num_keypoints=5,
        pwl_calibration_input_keypoints='quantiles',
    ),
    tfl.configs.FeatureConfig(
        name='slope',
        # Partial monotonicity: output(0) <= output(1), output(1) <= output(2)
        monotonicity=[(0, 1), (1, 2)],
        num_buckets=3,
    ),
    tfl.configs.FeatureConfig(
        name='ca',
        monotonicity='increasing',
        pwl_calibration_num_keypoints=4,
        pwl_calibration_input_keypoints='quantiles',
    ),
    tfl.configs.FeatureConfig(
        name='thal',
        # Partial monotonicity:
        # output(normal) <= output(fixed)
        # output(normal) <= output(reversible)
        monotonicity=[('normal', 'fixed'), ('normal', 'reversible')],
        num_buckets=3,
        # We must specify the vocabulary list in order to later set the
        # monotonicities since we used names and not indices.
        vocabulary_list=thal_vocab_list,
    ),
]

Defina monotonicidades e keypoints

Em seguida, defina corretamente as monotonicidades para as características em que usamos um vocabulário personalizado (como "thal" acima).

tfl.premade_lib.set_categorical_monotonicities(heart_feature_configs)

Por fim, conclua as configurações das características ao calcular e configurar keypoints.

feature_keypoints = tfl.premade_lib.compute_feature_keypoints(
    feature_configs=heart_feature_configs, features=heart_train_dict)
tfl.premade_lib.set_feature_keypoints(
    feature_configs=heart_feature_configs,
    feature_keypoints=feature_keypoints,
    add_missing_feature_configs=False)

Modelo linear calibrado

Para criar um modelo pré-fabricado do TFL, primeiro construa uma configuração de modelo a partir de tfl.configs. Um modelo linear calibrado é criado usando a tfl.configs.CalibratedLinearConfig. Ela aplica a calibração categórica e por partes nas características de entrada e, em seguida, uma combinação linear e uma calibração linear por partes de saída opcional. Ao usar a calibração de saída ou especificar limites de saída, a camada linear aplicará a média ponderada nas saídas calibradas.

Este exemplo cria um modelo linear calibrado nas primeiras 5 características.

# Model config defines the model structure for the premade model.
linear_model_config = tfl.configs.CalibratedLinearConfig(
    feature_configs=heart_feature_configs[:5],
    use_bias=True,
    output_calibration=True,
    output_calibration_num_keypoints=10,
    # We initialize the output to [-2.0, 2.0] since we'll be using logits.
    output_initialization=np.linspace(-2.0, 2.0, num=10),
    regularizer_configs=[
        # Regularizer for the output calibrator.
        tfl.configs.RegularizerConfig(name='output_calib_hessian', l2=1e-4),
    ])
# A CalibratedLinear premade model constructed from the given model config.
linear_model = tfl.premade.CalibratedLinear(linear_model_config)
# Let's plot our model.
tf.keras.utils.plot_model(linear_model, show_layer_names=False, rankdir='LR')

Agora, como com qualquer outro tf.keras.Model, é preciso compilar e ajustar o modelo aos dados.

linear_model.compile(
    loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.AUC(from_logits=True)],
    optimizer=tf.keras.optimizers.Adam(LEARNING_RATE))
linear_model.fit(
    heart_train_xs[:5],
    heart_train_ys,
    epochs=NUM_EPOCHS,
    batch_size=BATCH_SIZE,
    verbose=False)

Após o treinamento do modelo, é possível avaliá-lo no dataset de teste.

print('Test Set Evaluation...')
print(linear_model.evaluate(heart_test_xs[:5], heart_test_ys))

Modelo lattice calibrado

Um modelo lattice calibrado é criado usando tfl.configs.CalibratedLatticeConfig. Esse tipo de modelo aplica calibração linear por partes e categórica nas características de entrada e, em seguida, um modelo lattice e uma calibração linear por partes de saída opcional.

Este exemplo cria um modelo lattice calibrado nas primeiras 5 características.

# This is a calibrated lattice model: inputs are calibrated, then combined
# non-linearly using a lattice layer.
lattice_model_config = tfl.configs.CalibratedLatticeConfig(
    feature_configs=heart_feature_configs[:5],
    # We initialize the output to [-2.0, 2.0] since we'll be using logits.
    output_initialization=[-2.0, 2.0],
    regularizer_configs=[
        # Torsion regularizer applied to the lattice to make it more linear.
        tfl.configs.RegularizerConfig(name='torsion', l2=1e-2),
        # Globally defined calibration regularizer is applied to all features.
        tfl.configs.RegularizerConfig(name='calib_hessian', l2=1e-2),
    ])
# A CalibratedLattice premade model constructed from the given model config.
lattice_model = tfl.premade.CalibratedLattice(lattice_model_config)
# Let's plot our model.
tf.keras.utils.plot_model(lattice_model, show_layer_names=False, rankdir='LR')

Como antes, compile, ajuste e avalie o modelo.

lattice_model.compile(
    loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.AUC(from_logits=True)],
    optimizer=tf.keras.optimizers.Adam(LEARNING_RATE))
lattice_model.fit(
    heart_train_xs[:5],
    heart_train_ys,
    epochs=NUM_EPOCHS,
    batch_size=BATCH_SIZE,
    verbose=False)
print('Test Set Evaluation...')
print(lattice_model.evaluate(heart_test_xs[:5], heart_test_ys))

Modelo ensemble lattice calibrado

Quando o número de características for grande, você pode usar um modelo ensemble, que cria vários lattices menores para subconjuntos de características e calcula a média da saída deles, em vez de só criar um lattice grande. Os modelos lattice ensemble são criados usando tfl.configs.CalibratedLatticeEnsembleConfig. Um modelo ensemble lattice calibrado aplica a calibração linear por partes e categórica na característica de entrada e, em seguida, um ensemble de modelos lattice e uma calibração linear por partes de saída opcional.

Inicialização explícita de ensemble lattice

Se você já sabe com que subconjuntos de características quer alimentar os lattices, é possível definir os lattices de maneira explícita usando nomes de características. Este exemplo cria um modelo ensemble lattice calibrado com 5 lattices e 3 características por lattice.

# This is a calibrated lattice ensemble model: inputs are calibrated, then
# combined non-linearly and averaged using multiple lattice layers.
explicit_ensemble_model_config = tfl.configs.CalibratedLatticeEnsembleConfig(
    feature_configs=heart_feature_configs,
    lattices=[['trestbps', 'chol', 'ca'], ['fbs', 'restecg', 'thal'],
              ['fbs', 'cp', 'oldpeak'], ['exang', 'slope', 'thalach'],
              ['restecg', 'age', 'sex']],
    num_lattices=5,
    lattice_rank=3,
    # We initialize the output to [-2.0, 2.0] since we'll be using logits.
    output_initialization=[-2.0, 2.0])
# A CalibratedLatticeEnsemble premade model constructed from the given
# model config.
explicit_ensemble_model = tfl.premade.CalibratedLatticeEnsemble(
    explicit_ensemble_model_config)
# Let's plot our model.
tf.keras.utils.plot_model(
    explicit_ensemble_model, show_layer_names=False, rankdir='LR')

Como antes, compile, ajuste e avalie o modelo.

explicit_ensemble_model.compile(
    loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.AUC(from_logits=True)],
    optimizer=tf.keras.optimizers.Adam(LEARNING_RATE))
explicit_ensemble_model.fit(
    heart_train_xs,
    heart_train_ys,
    epochs=NUM_EPOCHS,
    batch_size=BATCH_SIZE,
    verbose=False)
print('Test Set Evaluation...')
print(explicit_ensemble_model.evaluate(heart_test_xs, heart_test_ys))

Ensemble lattice aleatório

Se você não sabe com que subconjuntos de características quer alimentar os lattices, outra opção é usar subconjuntos aleatórios para cada lattice. Este exemplo cria um modelo ensemble lattice calibrado com 5 lattices e 3 características por lattice.

# This is a calibrated lattice ensemble model: inputs are calibrated, then
# combined non-linearly and averaged using multiple lattice layers.
random_ensemble_model_config = tfl.configs.CalibratedLatticeEnsembleConfig(
    feature_configs=heart_feature_configs,
    lattices='random',
    num_lattices=5,
    lattice_rank=3,
    # We initialize the output to [-2.0, 2.0] since we'll be using logits.
    output_initialization=[-2.0, 2.0],
    random_seed=42)
# Now we must set the random lattice structure and construct the model.
tfl.premade_lib.set_random_lattice_ensemble(random_ensemble_model_config)
# A CalibratedLatticeEnsemble premade model constructed from the given
# model config.
random_ensemble_model = tfl.premade.CalibratedLatticeEnsemble(
    random_ensemble_model_config)
# Let's plot our model.
tf.keras.utils.plot_model(
    random_ensemble_model, show_layer_names=False, rankdir='LR')

Como antes, compile, ajuste e avalie o modelo.

random_ensemble_model.compile(
    loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.AUC(from_logits=True)],
    optimizer=tf.keras.optimizers.Adam(LEARNING_RATE))
random_ensemble_model.fit(
    heart_train_xs,
    heart_train_ys,
    epochs=NUM_EPOCHS,
    batch_size=BATCH_SIZE,
    verbose=False)
print('Test Set Evaluation...')
print(random_ensemble_model.evaluate(heart_test_xs, heart_test_ys))

Ensemble lattice aleatório de camada RTL

Ao usar um ensemble lattice aleatório, você pode especificar para o modelo usar uma única camada tfl.layers.RTL. Ressaltamos que tfl.layers.RTL só é compatível com as restrições de monotonicidade, precisa apresentar o mesmo tamanho de lattice para todas as características e não pode ter regularização por característica. Observe que o uso de uma camada tfl.layers.RTL permite escalar para ensembles muito maiores em comparação com o uso de instâncias tfl.layers.Lattice.

Este exemplo cria um modelo ensemble lattice calibrado com 5 lattices e 3 características por lattice.

# Make sure our feature configs have the same lattice size, no per-feature
# regularization, and only monotonicity constraints.
rtl_layer_feature_configs = copy.deepcopy(heart_feature_configs)
for feature_config in rtl_layer_feature_configs:
  feature_config.lattice_size = 2
  feature_config.unimodality = 'none'
  feature_config.reflects_trust_in = None
  feature_config.dominates = None
  feature_config.regularizer_configs = None
# This is a calibrated lattice ensemble model: inputs are calibrated, then
# combined non-linearly and averaged using multiple lattice layers.
rtl_layer_ensemble_model_config = tfl.configs.CalibratedLatticeEnsembleConfig(
    feature_configs=rtl_layer_feature_configs,
    lattices='rtl_layer',
    num_lattices=5,
    lattice_rank=3,
    # We initialize the output to [-2.0, 2.0] since we'll be using logits.
    output_initialization=[-2.0, 2.0],
    random_seed=42)
# A CalibratedLatticeEnsemble premade model constructed from the given
# model config. Note that we do not have to specify the lattices by calling
# a helper function (like before with random) because the RTL Layer will take
# care of that for us.
rtl_layer_ensemble_model = tfl.premade.CalibratedLatticeEnsemble(
    rtl_layer_ensemble_model_config)
# Let's plot our model.
tf.keras.utils.plot_model(
    rtl_layer_ensemble_model, show_layer_names=False, rankdir='LR')

Como antes, compile, ajuste e avalie o modelo.

rtl_layer_ensemble_model.compile(
    loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.AUC(from_logits=True)],
    optimizer=tf.keras.optimizers.Adam(LEARNING_RATE))
rtl_layer_ensemble_model.fit(
    heart_train_xs,
    heart_train_ys,
    epochs=NUM_EPOCHS,
    batch_size=BATCH_SIZE,
    verbose=False)
print('Test Set Evaluation...')
print(rtl_layer_ensemble_model.evaluate(heart_test_xs, heart_test_ys))

Ensemble lattice do Crystals

Os modelos pré-fabricados também fornecem um algoritmo de arranjo de características heurístico chamado Crystals. Para usar o algoritmo Crystals, primeiro é preciso treinar um modelo de prefitting que estima as interações de características por partes. Em seguida, arranjamos o ensemble final para que as características com mais interações não lineares fiquem nos mesmos lattices.

A biblioteca Premade oferece funções helper para construir a configuração de modelo de prefitting e extrair a estrutura de crystals. Observe que o modelo de prefitting não precisa ser totalmente treinado, então algumas épocas deve ser o suficiente.

Este exemplo cria um modelo ensemble lattice calibrado com 5 lattices e 3 características por lattice.

# This is a calibrated lattice ensemble model: inputs are calibrated, then
# combines non-linearly and averaged using multiple lattice layers.
crystals_ensemble_model_config = tfl.configs.CalibratedLatticeEnsembleConfig(
    feature_configs=heart_feature_configs,
    lattices='crystals',
    num_lattices=5,
    lattice_rank=3,
    # We initialize the output to [-2.0, 2.0] since we'll be using logits.
    output_initialization=[-2.0, 2.0],
    random_seed=42)
# Now that we have our model config, we can construct a prefitting model config.
prefitting_model_config = tfl.premade_lib.construct_prefitting_model_config(
    crystals_ensemble_model_config)
# A CalibratedLatticeEnsemble premade model constructed from the given
# prefitting model config.
prefitting_model = tfl.premade.CalibratedLatticeEnsemble(
    prefitting_model_config)
# We can compile and train our prefitting model as we like.
prefitting_model.compile(
    loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
    optimizer=tf.keras.optimizers.Adam(LEARNING_RATE))
prefitting_model.fit(
    heart_train_xs,
    heart_train_ys,
    epochs=PREFITTING_NUM_EPOCHS,
    batch_size=BATCH_SIZE,
    verbose=False)
# Now that we have our trained prefitting model, we can extract the crystals.
tfl.premade_lib.set_crystals_lattice_ensemble(crystals_ensemble_model_config,
                                              prefitting_model_config,
                                              prefitting_model)
# A CalibratedLatticeEnsemble premade model constructed from the given
# model config.
crystals_ensemble_model = tfl.premade.CalibratedLatticeEnsemble(
    crystals_ensemble_model_config)
# Let's plot our model.
tf.keras.utils.plot_model(
    crystals_ensemble_model, show_layer_names=False, rankdir='LR')

Como antes, compile, ajuste e avalie o modelo.

crystals_ensemble_model.compile(
    loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
    metrics=[tf.keras.metrics.AUC(from_logits=True)],
    optimizer=tf.keras.optimizers.Adam(LEARNING_RATE))
crystals_ensemble_model.fit(
    heart_train_xs,
    heart_train_ys,
    epochs=NUM_EPOCHS,
    batch_size=BATCH_SIZE,
    verbose=False)
print('Test Set Evaluation...')
print(crystals_ensemble_model.evaluate(heart_test_xs, heart_test_ys))