Nouveau! Utilisez Simple ML for Sheets pour appliquer l'apprentissage automatique aux données de vos feuilles de calcul Google Lire la suite

Cette page a été traduite par l'API Cloud Translation.

Composition de modèles de forêt de décision et de réseau neuronal

Voir sur TensorFlow.org

Exécuter dans Google Colab

Voir sur GitHub

Télécharger le cahier

API fonctionnelle Keras

introduction

Bienvenue dans le tutoriel composition de modèle pour les forêts tensorflow décision (TF-DF). Ce portable vous montre comment composer la forêt de décisions multiples et modèles de réseaux de neurones ensemble en utilisant une couche de pré - traitement commun et l' API fonctionnelle Keras .

Vous souhaiterez peut-être composer des modèles ensemble pour améliorer les performances prédictives (assemblage), pour tirer le meilleur parti de différentes technologies de modélisation (assemblage de modèles hétérogènes), pour entraîner différentes parties du modèle sur différents ensembles de données (par exemple, pré-entraînement), ou pour créer un modèle empilé (par exemple, un modèle fonctionne sur les prédictions d'un autre modèle).

Ce didacticiel couvre un cas d'utilisation avancé de la composition de modèle à l'aide de l'API fonctionnelle. Vous pouvez trouver des exemples pour les scénarios plus simples de la composition de modèle dans la section « fonction de pré - traitement » de ce tutoriel et dans la section « à l' aide d' un texte intégrant pré - entraîné » de ce tutoriel .

Voici la structure du modèle que vous allez construire :

!pip install graphviz -U --quiet

from graphviz import Source

Source("""
digraph G {
  raw_data [label="Input features"];
  preprocess_data [label="Learnable NN pre-processing", shape=rect];

  raw_data -> preprocess_data

  subgraph cluster_0 {
    color=grey;
    a1[label="NN layer", shape=rect];
    b1[label="NN layer", shape=rect];
    a1 -> b1;
    label = "Model #1";
  }

   subgraph cluster_1 {
    color=grey;
    a2[label="NN layer", shape=rect];
    b2[label="NN layer", shape=rect];
    a2 -> b2;
    label = "Model #2";
  }

  subgraph cluster_2 {
    color=grey;
    a3[label="Decision Forest", shape=rect];
    label = "Model #3";
  }

  subgraph cluster_3 {
    color=grey;
    a4[label="Decision Forest", shape=rect];
    label = "Model #4";
  }

  preprocess_data -> a1;
  preprocess_data -> a2;
  preprocess_data -> a3;
  preprocess_data -> a4;

  b1  -> aggr;
  b2  -> aggr;
  a3 -> aggr;
  a4 -> aggr;

  aggr [label="Aggregation (mean)", shape=rect]
  aggr -> predictions
}
""")

svg

Votre modèle composé comporte trois étapes :

La première étape est une couche de prétraitement composée d'un réseau de neurones et commune à tous les modèles de l'étape suivante. En pratique, une telle couche de prétraitement pourrait être soit une intégration pré-entraînée à affiner, soit un réseau de neurones initialisé de manière aléatoire.
La deuxième étape est un ensemble de deux modèles de forêt de décision et de deux réseaux de neurones.
La dernière étape fait la moyenne des prédictions des modèles de la deuxième étape. Il ne contient aucun poids pouvant être appris.

Les réseaux de neurones sont formés en utilisant l' algorithme de rétropropagation et descente de gradient. Cet algorithme a deux propriétés importantes : (1) La couche de réseau de neurones peut être entraînée si elle reçoit un gradient de perte (plus précisément, le gradient de perte en fonction de la sortie de la couche), et (2) l'algorithme "transmet" le gradient de perte de la sortie de la couche à l'entrée de la couche (c'est la "règle de la chaîne"). Pour ces deux raisons, la rétropropagation peut entraîner ensemble plusieurs couches de réseaux de neurones empilés les uns sur les autres.

Dans cet exemple, les forêts de décision sont formés à la forêt aléatoire algorithme (RF). Contrairement à la rétropropagation, l'apprentissage de RF ne "transmet" pas le gradient de perte de sa sortie à son entrée. Pour cette raison, l'algorithme RF classique ne peut pas être utilisé pour entraîner ou affiner un réseau de neurones en dessous. En d'autres termes, les étapes "forêt de décision" ne peuvent pas être utilisées pour former le "bloc de pré-traitement NN apprenant".

Former le prétraitement et l'étape des réseaux de neurones.
Former les étapes de la forêt de décision.

Installer les forêts de décision TensorFlow

Installez TF-DF en exécutant la cellule suivante.

pip install tensorflow_decision_forests -U --quiet

Installer Wurlitzer pour afficher les journaux de formation détaillés. Cela n'est nécessaire que dans les ordinateurs portables.

pip install wurlitzer -U --quiet

Importer des bibliothèques

import tensorflow_decision_forests as tfdf

import os
import numpy as np
import pandas as pd
import tensorflow as tf
import math
import matplotlib.pyplot as plt

try:
  from wurlitzer import sys_pipes
except:
  from colabtools.googlelog import CaptureLog as sys_pipes

from IPython.core.magic import register_line_magic
from IPython.display import Javascript

WARNING:root:Failure to load the custom c++ tensorflow ops. This error is likely caused the version of TensorFlow and TensorFlow Decision Forests are not compatible.
WARNING:root:TF Parameter Server distributed training not available.

Base de données

Vous utiliserez un jeu de données synthétique simple dans ce didacticiel pour faciliter l'interprétation du modèle final.

def make_dataset(num_examples, num_features, seed=1234):
  np.random.seed(seed)
  features = np.random.uniform(-1, 1, size=(num_examples, num_features))
  noise = np.random.uniform(size=(num_examples))

  left_side = np.sqrt(
      np.sum(np.multiply(np.square(features[:, 0:2]), [1, 2]), axis=1))
  right_side = features[:, 2] * 0.7 + np.sin(
      features[:, 3] * 10) * 0.5 + noise * 0.0 + 0.5

  labels = left_side <= right_side
  return features, labels.astype(int)

Générez quelques exemples :

make_dataset(num_examples=5, num_features=4)

(array([[-0.6169611 ,  0.24421754, -0.12454452,  0.57071717],
        [ 0.55995162, -0.45481479, -0.44707149,  0.60374436],
        [ 0.91627871,  0.75186527, -0.28436546,  0.00199025],
        [ 0.36692587,  0.42540405, -0.25949849,  0.12239237],
        [ 0.00616633, -0.9724631 ,  0.54565324,  0.76528238]]),
 array([0, 0, 0, 1, 0]))

Vous pouvez également les tracer pour avoir une idée du motif synthétique :

plot_features, plot_label = make_dataset(num_examples=50000, num_features=4)

plt.rcParams["figure.figsize"] = [8, 8]
common_args = dict(c=plot_label, s=1.0, alpha=0.5)

plt.subplot(2, 2, 1)
plt.scatter(plot_features[:, 0], plot_features[:, 1], **common_args)

plt.subplot(2, 2, 2)
plt.scatter(plot_features[:, 1], plot_features[:, 2], **common_args)

plt.subplot(2, 2, 3)
plt.scatter(plot_features[:, 0], plot_features[:, 2], **common_args)

plt.subplot(2, 2, 4)
plt.scatter(plot_features[:, 0], plot_features[:, 3], **common_args)

<matplotlib.collections.PathCollection at 0x7f6b78d20e90>

png

Notez que ce motif est lisse et non aligné sur l'axe. Cela profitera aux modèles de réseaux de neurones. En effet, il est plus facile pour un réseau de neurones que pour un arbre de décision d'avoir des limites de décision rondes et non alignées.

D'autre part, nous allons entraîner le modèle sur un petit ensemble de données avec 2500 exemples. Cela avantagera les modèles de forêt décisionnelle. Cela est dû au fait que les forêts de décision sont beaucoup plus efficaces, en utilisant toutes les informations disponibles dans les exemples (les forêts de décision sont « sample efficientes »).

Notre ensemble de réseaux de neurones et de forêts de décision utilisera le meilleur des deux mondes.

Créons un train et un test tf.data.Dataset :

def make_tf_dataset(batch_size=64, **args):
  features, labels = make_dataset(**args)
  return tf.data.Dataset.from_tensor_slices(
      (features, labels)).batch(batch_size)


num_features = 10

train_dataset = make_tf_dataset(
    num_examples=2500, num_features=num_features, batch_size=64, seed=1234)
test_dataset = make_tf_dataset(
    num_examples=10000, num_features=num_features, batch_size=64, seed=5678)

Structure du modèle

Définissez la structure du modèle comme suit :

# Input features.
raw_features = tf.keras.layers.Input(shape=(num_features,))

# Stage 1
# =======

# Common learnable pre-processing
preprocessor = tf.keras.layers.Dense(10, activation=tf.nn.relu6)
preprocess_features = preprocessor(raw_features)

# Stage 2
# =======

# Model #1: NN
m1_z1 = tf.keras.layers.Dense(5, activation=tf.nn.relu6)(preprocess_features)
m1_pred = tf.keras.layers.Dense(1, activation=tf.nn.sigmoid)(m1_z1)

# Model #2: NN
m2_z1 = tf.keras.layers.Dense(5, activation=tf.nn.relu6)(preprocess_features)
m2_pred = tf.keras.layers.Dense(1, activation=tf.nn.sigmoid)(m2_z1)


def seed_advanced_argument(seed):
  """Create a seed argument for a TF-DF model.

  TODO(gbm): Surface the "seed" argument to the model constructor directly.
  """
  return tfdf.keras.AdvancedArguments(
      yggdrasil_training_config=tfdf.keras.core.YggdrasilTrainingConfig(
          random_seed=seed))


# Model #3: DF
model_3 = tfdf.keras.RandomForestModel(
    num_trees=1000, advanced_arguments=seed_advanced_argument(1234))
m3_pred = model_3(preprocess_features)

# Model #4: DF
model_4 = tfdf.keras.RandomForestModel(
    num_trees=1000,
    #split_axis="SPARSE_OBLIQUE", # Uncomment this line to increase the quality of this model
    advanced_arguments=seed_advanced_argument(4567))
m4_pred = model_4(preprocess_features)

# Since TF-DF uses deterministic learning algorithms, you should set the model's
# training seed to different values otherwise both
# `tfdf.keras.RandomForestModel` will be exactly the same.

# Stage 3
# =======

mean_nn_only = tf.reduce_mean(tf.stack([m1_pred, m2_pred], axis=0), axis=0)
mean_nn_and_df = tf.reduce_mean(
    tf.stack([m1_pred, m2_pred, m3_pred, m4_pred], axis=0), axis=0)

# Keras Models
# ============

ensemble_nn_only = tf.keras.models.Model(raw_features, mean_nn_only)
ensemble_nn_and_df = tf.keras.models.Model(raw_features, mean_nn_and_df)

WARNING:tensorflow:AutoGraph could not transform <bound method Socket.send of <zmq.Socket(zmq.PUSH) at 0x7f6ba21b62f0>> and will run it as-is.
Please report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.
Cause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method
To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
WARNING:absl:The model was called directly (i.e. using `model(data)` instead of using `model.predict(data)`) before being trained. The model will only return zeros until trained. The output shape might change after training Tensor("inputs:0", shape=(None, 10), dtype=float32)
WARNING:tensorflow:AutoGraph could not transform <bound method Socket.send of <zmq.Socket(zmq.PUSH) at 0x7f6ba21b62f0>> and will run it as-is.
Please report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.
Cause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method
To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
WARNING: AutoGraph could not transform <bound method Socket.send of <zmq.Socket(zmq.PUSH) at 0x7f6ba21b62f0>> and will run it as-is.
Please report this to the TensorFlow team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output.
Cause: module, class, method, function, traceback, frame, or code object was expected, got cython_function_or_method
To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
WARNING:absl:The model was called directly (i.e. using `model(data)` instead of using `model.predict(data)`) before being trained. The model will only return zeros until trained. The output shape might change after training Tensor("inputs:0", shape=(None, 10), dtype=float32)

Avant d'entraîner le modèle, vous pouvez le tracer pour vérifier s'il est similaire au diagramme initial.

from keras.utils.vis_utils import plot_model

plot_model(ensemble_nn_and_df, to_file="/tmp/model.png", show_shapes=True)

png

Formation de modèle

Entraînez d'abord le prétraitement et les deux couches de réseau de neurones à l'aide de l'algorithme de rétropropagation.

%%time
ensemble_nn_only.compile(
        optimizer=tf.keras.optimizers.Adam(),
        loss=tf.keras.losses.BinaryCrossentropy(),
        metrics=["accuracy"])

ensemble_nn_only.fit(train_dataset, epochs=20, validation_data=test_dataset)

Epoch 1/20
40/40 [==============================] - 1s 13ms/step - loss: 0.6115 - accuracy: 0.7308 - val_loss: 0.5857 - val_accuracy: 0.7407
Epoch 2/20
40/40 [==============================] - 0s 9ms/step - loss: 0.5645 - accuracy: 0.7484 - val_loss: 0.5487 - val_accuracy: 0.7391
Epoch 3/20
40/40 [==============================] - 0s 9ms/step - loss: 0.5310 - accuracy: 0.7496 - val_loss: 0.5237 - val_accuracy: 0.7392
Epoch 4/20
40/40 [==============================] - 0s 9ms/step - loss: 0.5074 - accuracy: 0.7500 - val_loss: 0.5055 - val_accuracy: 0.7391
Epoch 5/20
40/40 [==============================] - 0s 9ms/step - loss: 0.4887 - accuracy: 0.7496 - val_loss: 0.4901 - val_accuracy: 0.7397
Epoch 6/20
40/40 [==============================] - 0s 9ms/step - loss: 0.4725 - accuracy: 0.7520 - val_loss: 0.4763 - val_accuracy: 0.7440
Epoch 7/20
40/40 [==============================] - 0s 9ms/step - loss: 0.4585 - accuracy: 0.7584 - val_loss: 0.4644 - val_accuracy: 0.7542
Epoch 8/20
40/40 [==============================] - 0s 9ms/step - loss: 0.4470 - accuracy: 0.7700 - val_loss: 0.4544 - val_accuracy: 0.7682
Epoch 9/20
40/40 [==============================] - 0s 9ms/step - loss: 0.4374 - accuracy: 0.7804 - val_loss: 0.4462 - val_accuracy: 0.7789
Epoch 10/20
40/40 [==============================] - 0s 9ms/step - loss: 0.4297 - accuracy: 0.7848 - val_loss: 0.4395 - val_accuracy: 0.7865
Epoch 11/20
40/40 [==============================] - 0s 9ms/step - loss: 0.4232 - accuracy: 0.7904 - val_loss: 0.4339 - val_accuracy: 0.7933
Epoch 12/20
40/40 [==============================] - 0s 10ms/step - loss: 0.4176 - accuracy: 0.7952 - val_loss: 0.4289 - val_accuracy: 0.7963
Epoch 13/20
40/40 [==============================] - 0s 9ms/step - loss: 0.4126 - accuracy: 0.7992 - val_loss: 0.4243 - val_accuracy: 0.8010
Epoch 14/20
40/40 [==============================] - 0s 9ms/step - loss: 0.4078 - accuracy: 0.8052 - val_loss: 0.4199 - val_accuracy: 0.8033
Epoch 15/20
40/40 [==============================] - 0s 9ms/step - loss: 0.4029 - accuracy: 0.8096 - val_loss: 0.4155 - val_accuracy: 0.8067
Epoch 16/20
40/40 [==============================] - 0s 9ms/step - loss: 0.3981 - accuracy: 0.8132 - val_loss: 0.4109 - val_accuracy: 0.8099
Epoch 17/20
40/40 [==============================] - 0s 9ms/step - loss: 0.3932 - accuracy: 0.8152 - val_loss: 0.4061 - val_accuracy: 0.8129
Epoch 18/20
40/40 [==============================] - 0s 9ms/step - loss: 0.3883 - accuracy: 0.8208 - val_loss: 0.4012 - val_accuracy: 0.8149
Epoch 19/20
40/40 [==============================] - 0s 9ms/step - loss: 0.3832 - accuracy: 0.8232 - val_loss: 0.3963 - val_accuracy: 0.8168
Epoch 20/20
40/40 [==============================] - 0s 10ms/step - loss: 0.3783 - accuracy: 0.8276 - val_loss: 0.3912 - val_accuracy: 0.8203
CPU times: user 12.1 s, sys: 2.14 s, total: 14.2 s
Wall time: 8.54 s
<keras.callbacks.History at 0x7f6b181d7450>

Évaluons le prétraitement et la partie avec les deux réseaux de neurones uniquement :

evaluation_nn_only = ensemble_nn_only.evaluate(test_dataset, return_dict=True)
print("Accuracy (NN #1 and #2 only): ", evaluation_nn_only["accuracy"])
print("Loss (NN #1 and #2 only): ", evaluation_nn_only["loss"])

157/157 [==============================] - 0s 2ms/step - loss: 0.3912 - accuracy: 0.8203
Accuracy (NN #1 and #2 only):  0.8202999830245972
Loss (NN #1 and #2 only):  0.39124569296836853

Entraînons les deux composants de la forêt de décision (l'un après l'autre).

%%time
train_dataset_with_preprocessing = train_dataset.map(lambda x,y: (preprocessor(x), y))
test_dataset_with_preprocessing = test_dataset.map(lambda x,y: (preprocessor(x), y))

model_3.fit(train_dataset_with_preprocessing)
model_4.fit(train_dataset_with_preprocessing)

WARNING:tensorflow:AutoGraph could not transform <function <lambda> at 0x7f6b86bc3dd0> and will run it as-is.
Cause: could not parse the source code of <function <lambda> at 0x7f6b86bc3dd0>: no matching AST found among candidates:

To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
WARNING:tensorflow:AutoGraph could not transform <function <lambda> at 0x7f6b86bc3dd0> and will run it as-is.
Cause: could not parse the source code of <function <lambda> at 0x7f6b86bc3dd0>: no matching AST found among candidates:

To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
WARNING: AutoGraph could not transform <function <lambda> at 0x7f6b86bc3dd0> and will run it as-is.
Cause: could not parse the source code of <function <lambda> at 0x7f6b86bc3dd0>: no matching AST found among candidates:

To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
WARNING:tensorflow:AutoGraph could not transform <function <lambda> at 0x7f6b783a9320> and will run it as-is.
Cause: could not parse the source code of <function <lambda> at 0x7f6b783a9320>: no matching AST found among candidates:

To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
WARNING:tensorflow:AutoGraph could not transform <function <lambda> at 0x7f6b783a9320> and will run it as-is.
Cause: could not parse the source code of <function <lambda> at 0x7f6b783a9320>: no matching AST found among candidates:

To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
WARNING: AutoGraph could not transform <function <lambda> at 0x7f6b783a9320> and will run it as-is.
Cause: could not parse the source code of <function <lambda> at 0x7f6b783a9320>: no matching AST found among candidates:

To silence this warning, decorate the function with @tf.autograph.experimental.do_not_convert
23/40 [================>.............] - ETA: 0s
[INFO kernel.cc:736] Start Yggdrasil model training
[INFO kernel.cc:737] Collect training examples
[INFO kernel.cc:392] Number of batches: 40
[INFO kernel.cc:393] Number of examples: 2500
[INFO kernel.cc:759] Dataset:
Number of records: 2500
Number of columns: 11

Number of columns by type:
    NUMERICAL: 10 (90.9091%)
    CATEGORICAL: 1 (9.09091%)

Columns:

NUMERICAL: 10 (90.9091%)
    0: "data:0.0" NUMERICAL mean:0.356465 min:0 max:2.37352 sd:0.451418
    1: "data:0.1" NUMERICAL mean:0.392088 min:0 max:2.3411 sd:0.470499
    2: "data:0.2" NUMERICAL mean:0.382386 min:0 max:2.11809 sd:0.483672
    3: "data:0.3" NUMERICAL mean:0.290395 min:0 max:2.27481 sd:0.400102
    4: "data:0.4" NUMERICAL mean:0.210684 min:0 max:1.35897 sd:0.281379
    5: "data:0.5" NUMERICAL mean:0.4008 min:0 max:2.06561 sd:0.453018
    6: "data:0.6" NUMERICAL mean:0.289166 min:0 max:2.0263 sd:0.407337
    7: "data:0.7" NUMERICAL mean:0.277971 min:0 max:1.77561 sd:0.363215
    8: "data:0.8" NUMERICAL mean:0.41254 min:0 max:2.79804 sd:0.553333
    9: "data:0.9" NUMERICAL mean:0.197082 min:0 max:1.60773 sd:0.298194

CATEGORICAL: 1 (9.09091%)
    10: "__LABEL" CATEGORICAL integerized vocab-size:3 no-ood-item

Terminology:
    nas: Number of non-available (i.e. missing) values.
    ood: Out of dictionary.
    manually-defined: Attribute which type is manually defined by the user i.e. the type was not automatically inferred.
    tokenized: The attribute value is obtained through tokenization.
    has-dict: The attribute is attached to a string dictionary e.g. a categorical attribute stored as a string.
    vocab-size: Number of unique values.

[INFO kernel.cc:762] Configure learner
[INFO kernel.cc:787] Training config:
learner: "RANDOM_FOREST"
features: "data:0\\.0"
features: "data:0\\.1"
features: "data:0\\.2"
features: "data:0\\.3"
features: "data:0\\.4"
features: "data:0\\.5"
features: "data:0\\.6"
features: "data:0\\.7"
features: "data:0\\.8"
features: "data:0\\.9"
label: "__LABEL"
task: CLASSIFICATION
random_seed: 1234
[yggdrasil_decision_forests.model.random_forest.proto.random_forest_config] {
  num_trees: 1000
  decision_tree {
    max_depth: 16
    min_examples: 5
    in_split_min_examples_check: true
    missing_value_policy: GLOBAL_IMPUTATION
    allow_na_conditions: false
    categorical_set_greedy_forward {
      sampling: 0.1
      max_num_items: -1
      min_item_frequency: 1
    }
    growing_strategy_local {
    }
    categorical {
      cart {
      }
    }
    num_candidate_attributes_ratio: -1
    axis_aligned_split {
    }
    internal {
      sorting_strategy: PRESORTED
    }
  }
  winner_take_all_inference: true
  compute_oob_performances: true
  compute_oob_variable_importances: false
  adapt_bootstrap_size_ratio_for_maximum_training_duration: false
}

[INFO kernel.cc:790] Deployment config:
num_threads: 6

[INFO kernel.cc:817] Train model
[INFO random_forest.cc:315] Training random forest on 2500 example(s) and 10 feature(s).
[INFO random_forest.cc:628] Training of tree  1/1000 (tree index:1) done accuracy:0.781996 logloss:7.85767
[INFO random_forest.cc:628] Training of tree  11/1000 (tree index:8) done accuracy:0.79895 logloss:2.7263
[INFO random_forest.cc:628] Training of tree  21/1000 (tree index:20) done accuracy:0.8012 logloss:1.26831
[INFO random_forest.cc:628] Training of tree  31/1000 (tree index:30) done accuracy:0.8076 logloss:0.898323
[INFO random_forest.cc:628] Training of tree  41/1000 (tree index:37) done accuracy:0.8084 logloss:0.736323
[INFO random_forest.cc:628] Training of tree  51/1000 (tree index:51) done accuracy:0.8072 logloss:0.612984
[INFO random_forest.cc:628] Training of tree  61/1000 (tree index:63) done accuracy:0.8104 logloss:0.55782
[INFO random_forest.cc:628] Training of tree  71/1000 (tree index:69) done accuracy:0.81 logloss:0.544938
[INFO random_forest.cc:628] Training of tree  81/1000 (tree index:80) done accuracy:0.814 logloss:0.532167
[INFO random_forest.cc:628] Training of tree  91/1000 (tree index:89) done accuracy:0.8144 logloss:0.530892
[INFO random_forest.cc:628] Training of tree  101/1000 (tree index:100) done accuracy:0.814 logloss:0.516588
[INFO random_forest.cc:628] Training of tree  111/1000 (tree index:108) done accuracy:0.8128 logloss:0.490739
[INFO random_forest.cc:628] Training of tree  121/1000 (tree index:118) done accuracy:0.8124 logloss:0.490544
[INFO random_forest.cc:628] Training of tree  131/1000 (tree index:134) done accuracy:0.8112 logloss:0.451653
[INFO random_forest.cc:628] Training of tree  141/1000 (tree index:140) done accuracy:0.8136 logloss:0.437757
[INFO random_forest.cc:628] Training of tree  151/1000 (tree index:150) done accuracy:0.8144 logloss:0.424328
[INFO random_forest.cc:628] Training of tree  161/1000 (tree index:159) done accuracy:0.8132 logloss:0.42426
[INFO random_forest.cc:628] Training of tree  171/1000 (tree index:168) done accuracy:0.814 logloss:0.411061
[INFO random_forest.cc:628] Training of tree  181/1000 (tree index:184) done accuracy:0.8136 logloss:0.411324
[INFO random_forest.cc:628] Training of tree  191/1000 (tree index:190) done accuracy:0.8148 logloss:0.410002
[INFO random_forest.cc:628] Training of tree  201/1000 (tree index:200) done accuracy:0.8144 logloss:0.409526
[INFO random_forest.cc:628] Training of tree  211/1000 (tree index:208) done accuracy:0.814 logloss:0.40944
[INFO random_forest.cc:628] Training of tree  221/1000 (tree index:218) done accuracy:0.8152 logloss:0.409039
[INFO random_forest.cc:628] Training of tree  231/1000 (tree index:234) done accuracy:0.8144 logloss:0.409254
[INFO random_forest.cc:628] Training of tree  241/1000 (tree index:242) done accuracy:0.8144 logloss:0.40879
[INFO random_forest.cc:628] Training of tree  251/1000 (tree index:251) done accuracy:0.8152 logloss:0.395703
[INFO random_forest.cc:628] Training of tree  261/1000 (tree index:259) done accuracy:0.8168 logloss:0.395747
[INFO random_forest.cc:628] Training of tree  271/1000 (tree index:268) done accuracy:0.814 logloss:0.394959
[INFO random_forest.cc:628] Training of tree  281/1000 (tree index:283) done accuracy:0.8148 logloss:0.395202
[INFO random_forest.cc:628] Training of tree  291/1000 (tree index:292) done accuracy:0.8136 logloss:0.395536
[INFO random_forest.cc:628] Training of tree  301/1000 (tree index:300) done accuracy:0.8128 logloss:0.39472
[INFO random_forest.cc:628] Training of tree  311/1000 (tree index:308) done accuracy:0.8124 logloss:0.394763
[INFO random_forest.cc:628] Training of tree  321/1000 (tree index:318) done accuracy:0.8132 logloss:0.394732
[INFO random_forest.cc:628] Training of tree  331/1000 (tree index:334) done accuracy:0.8136 logloss:0.394822
[INFO random_forest.cc:628] Training of tree  341/1000 (tree index:343) done accuracy:0.812 logloss:0.395051
[INFO random_forest.cc:628] Training of tree  351/1000 (tree index:350) done accuracy:0.8132 logloss:0.39492
[INFO random_forest.cc:628] Training of tree  361/1000 (tree index:358) done accuracy:0.8132 logloss:0.395054
[INFO random_forest.cc:628] Training of tree  371/1000 (tree index:368) done accuracy:0.812 logloss:0.395588
[INFO random_forest.cc:628] Training of tree  381/1000 (tree index:384) done accuracy:0.8104 logloss:0.395576
[INFO random_forest.cc:628] Training of tree  391/1000 (tree index:390) done accuracy:0.8132 logloss:0.395713
[INFO random_forest.cc:628] Training of tree  401/1000 (tree index:400) done accuracy:0.8088 logloss:0.383693
[INFO random_forest.cc:628] Training of tree  411/1000 (tree index:408) done accuracy:0.8088 logloss:0.383575
[INFO random_forest.cc:628] Training of tree  421/1000 (tree index:417) done accuracy:0.8096 logloss:0.383934
[INFO random_forest.cc:628] Training of tree  431/1000 (tree index:434) done accuracy:0.81 logloss:0.384001
[INFO random_forest.cc:628] Training of tree  441/1000 (tree index:442) done accuracy:0.808 logloss:0.384118
[INFO random_forest.cc:628] Training of tree  451/1000 (tree index:450) done accuracy:0.8096 logloss:0.384076
[INFO random_forest.cc:628] Training of tree  461/1000 (tree index:458) done accuracy:0.8104 logloss:0.383208
[INFO random_forest.cc:628] Training of tree  471/1000 (tree index:468) done accuracy:0.812 logloss:0.383298
[INFO random_forest.cc:628] Training of tree  481/1000 (tree index:482) done accuracy:0.81 logloss:0.38358
[INFO random_forest.cc:628] Training of tree  491/1000 (tree index:492) done accuracy:0.812 logloss:0.383453
[INFO random_forest.cc:628] Training of tree  501/1000 (tree index:500) done accuracy:0.8128 logloss:0.38317
[INFO random_forest.cc:628] Training of tree  511/1000 (tree index:508) done accuracy:0.812 logloss:0.383369
[INFO random_forest.cc:628] Training of tree  521/1000 (tree index:518) done accuracy:0.8132 logloss:0.383461
[INFO random_forest.cc:628] Training of tree  531/1000 (tree index:532) done accuracy:0.8124 logloss:0.38342
[INFO random_forest.cc:628] Training of tree  541/1000 (tree index:542) done accuracy:0.8128 logloss:0.383376
[INFO random_forest.cc:628] Training of tree  551/1000 (tree index:550) done accuracy:0.8128 logloss:0.383663
[INFO random_forest.cc:628] Training of tree  561/1000 (tree index:558) done accuracy:0.812 logloss:0.383574
[INFO random_forest.cc:628] Training of tree  571/1000 (tree index:568) done accuracy:0.8116 logloss:0.383529
[INFO random_forest.cc:628] Training of tree  581/1000 (tree index:580) done accuracy:0.8128 logloss:0.383624
[INFO random_forest.cc:628] Training of tree  591/1000 (tree index:592) done accuracy:0.814 logloss:0.383599
[INFO random_forest.cc:628] Training of tree  601/1000 (tree index:601) done accuracy:0.8148 logloss:0.383524
[INFO random_forest.cc:628] Training of tree  611/1000 (tree index:608) done accuracy:0.8156 logloss:0.383555
[INFO random_forest.cc:628] Training of tree  621/1000 (tree index:619) done accuracy:0.8132 logloss:0.382847
[INFO random_forest.cc:628] Training of tree  631/1000 (tree index:632) done accuracy:0.8124 logloss:0.382872
[INFO random_forest.cc:628] Training of tree  641/1000 (tree index:641) done accuracy:0.8144 logloss:0.382728
[INFO random_forest.cc:628] Training of tree  651/1000 (tree index:648) done accuracy:0.8132 logloss:0.382554
[INFO random_forest.cc:628] Training of tree  661/1000 (tree index:658) done accuracy:0.8128 logloss:0.382705
[INFO random_forest.cc:628] Training of tree  671/1000 (tree index:670) done accuracy:0.8136 logloss:0.38288
[INFO random_forest.cc:628] Training of tree  681/1000 (tree index:682) done accuracy:0.8152 logloss:0.383007
[INFO random_forest.cc:628] Training of tree  691/1000 (tree index:690) done accuracy:0.8144 logloss:0.382971
[INFO random_forest.cc:628] Training of tree  701/1000 (tree index:698) done accuracy:0.8152 logloss:0.382869
[INFO random_forest.cc:628] Training of tree  711/1000 (tree index:708) done accuracy:0.8152 logloss:0.382792
[INFO random_forest.cc:628] Training of tree  721/1000 (tree index:722) done accuracy:0.8136 logloss:0.38274
[INFO random_forest.cc:628] Training of tree  731/1000 (tree index:732) done accuracy:0.8144 logloss:0.38268
[INFO random_forest.cc:628] Training of tree  741/1000 (tree index:740) done accuracy:0.814 logloss:0.382835
[INFO random_forest.cc:628] Training of tree  751/1000 (tree index:751) done accuracy:0.8152 logloss:0.38297
[INFO random_forest.cc:628] Training of tree  761/1000 (tree index:758) done accuracy:0.8152 logloss:0.382917
[INFO random_forest.cc:628] Training of tree  771/1000 (tree index:770) done accuracy:0.8156 logloss:0.370596
[INFO random_forest.cc:628] Training of tree  781/1000 (tree index:782) done accuracy:0.816 logloss:0.370687
[INFO random_forest.cc:628] Training of tree  791/1000 (tree index:789) done accuracy:0.8164 logloss:0.37068
[INFO random_forest.cc:628] Training of tree  801/1000 (tree index:798) done accuracy:0.8172 logloss:0.370535
[INFO random_forest.cc:628] Training of tree  811/1000 (tree index:809) done accuracy:0.816 logloss:0.370674
[INFO random_forest.cc:628] Training of tree  821/1000 (tree index:821) done accuracy:0.816 logloss:0.370929
[INFO random_forest.cc:628] Training of tree  831/1000 (tree index:829) done accuracy:0.8148 logloss:0.370904
[INFO random_forest.cc:628] Training of tree  841/1000 (tree index:841) done accuracy:0.8164 logloss:0.371016
[INFO random_forest.cc:628] Training of tree  851/1000 (tree index:849) done accuracy:0.8168 logloss:0.370914
[INFO random_forest.cc:628] Training of tree  861/1000 (tree index:860) done accuracy:0.8164 logloss:0.371043
[INFO random_forest.cc:628] Training of tree  871/1000 (tree index:871) done accuracy:0.8168 logloss:0.371094
[INFO random_forest.cc:628] Training of tree  881/1000 (tree index:878) done accuracy:0.8152 logloss:0.371054
[INFO random_forest.cc:628] Training of tree  891/1000 (tree index:888) done accuracy:0.8156 logloss:0.370908
[INFO random_forest.cc:628] Training of tree  901/1000 (tree index:900) done accuracy:0.8156 logloss:0.370831
[INFO random_forest.cc:628] Training of tree  911/1000 (tree index:910) done accuracy:0.8152 logloss:0.370775
[INFO random_forest.cc:628] Training of tree  921/1000 (tree index:922) done accuracy:0.814 logloss:0.370804
[INFO random_forest.cc:628] Training of tree  931/1000 (tree index:929) done accuracy:0.8148 logloss:0.370495
[INFO random_forest.cc:628] Training of tree  941/1000 (tree index:941) done accuracy:0.816 logloss:0.370443
[INFO random_forest.cc:628] Training of tree  951/1000 (tree index:948) done accuracy:0.8156 logloss:0.370486
[INFO random_forest.cc:628] Training of tree  961/1000 (tree index:960) done accuracy:0.8152 logloss:0.370519
[INFO random_forest.cc:628] Training of tree  971/1000 (tree index:971) done accuracy:0.8144 logloss:0.370543
[INFO random_forest.cc:628] Training of tree  981/1000 (tree index:983) done accuracy:0.8144 logloss:0.370629
[INFO random_forest.cc:628] Training of tree  991/1000 (tree index:991) done accuracy:0.814 logloss:0.370625
[INFO random_forest.cc:628] Training of tree  1000/1000 (tree index:998) done accuracy:0.8144 logloss:0.370667
[INFO random_forest.cc:696] Final OOB metrics: accuracy:0.8144 logloss:0.370667
[INFO kernel.cc:828] Export model in log directory: /tmp/tmp9izglk4r
[INFO kernel.cc:836] Save model in resources
[INFO kernel.cc:988] Loading model from path
40/40 [==============================] - 6s 66ms/step
[INFO decision_forest.cc:590] Model loaded with 1000 root(s), 324508 node(s), and 10 input feature(s).
[INFO abstract_model.cc:993] Engine "RandomForestOptPred" built
[INFO kernel.cc:848] Use fast generic engine
24/40 [=================>............] - ETA: 0s
[INFO kernel.cc:736] Start Yggdrasil model training
[INFO kernel.cc:737] Collect training examples
[INFO kernel.cc:392] Number of batches: 40
[INFO kernel.cc:393] Number of examples: 2500
[INFO kernel.cc:759] Dataset:
Number of records: 2500
Number of columns: 11

Number of columns by type:
    NUMERICAL: 10 (90.9091%)
    CATEGORICAL: 1 (9.09091%)

Columns:

NUMERICAL: 10 (90.9091%)
    0: "data:0.0" NUMERICAL mean:0.356465 min:0 max:2.37352 sd:0.451418
    1: "data:0.1" NUMERICAL mean:0.392088 min:0 max:2.3411 sd:0.470499
    2: "data:0.2" NUMERICAL mean:0.382386 min:0 max:2.11809 sd:0.483672
    3: "data:0.3" NUMERICAL mean:0.290395 min:0 max:2.27481 sd:0.400102
    4: "data:0.4" NUMERICAL mean:0.210684 min:0 max:1.35897 sd:0.281379
    5: "data:0.5" NUMERICAL mean:0.4008 min:0 max:2.06561 sd:0.453018
    6: "data:0.6" NUMERICAL mean:0.289166 min:0 max:2.0263 sd:0.407337
    7: "data:0.7" NUMERICAL mean:0.277971 min:0 max:1.77561 sd:0.363215
    8: "data:0.8" NUMERICAL mean:0.41254 min:0 max:2.79804 sd:0.553333
    9: "data:0.9" NUMERICAL mean:0.197082 min:0 max:1.60773 sd:0.298194

CATEGORICAL: 1 (9.09091%)
    10: "__LABEL" CATEGORICAL integerized vocab-size:3 no-ood-item

Terminology:
    nas: Number of non-available (i.e. missing) values.
    ood: Out of dictionary.
    manually-defined: Attribute which type is manually defined by the user i.e. the type was not automatically inferred.
    tokenized: The attribute value is obtained through tokenization.
    has-dict: The attribute is attached to a string dictionary e.g. a categorical attribute stored as a string.
    vocab-size: Number of unique values.

[INFO kernel.cc:762] Configure learner
[INFO kernel.cc:787] Training config:
learner: "RANDOM_FOREST"
features: "data:0\\.0"
features: "data:0\\.1"
features: "data:0\\.2"
features: "data:0\\.3"
features: "data:0\\.4"
features: "data:0\\.5"
features: "data:0\\.6"
features: "data:0\\.7"
features: "data:0\\.8"
features: "data:0\\.9"
label: "__LABEL"
task: CLASSIFICATION
random_seed: 4567
[yggdrasil_decision_forests.model.random_forest.proto.random_forest_config] {
  num_trees: 1000
  decision_tree {
    max_depth: 16
    min_examples: 5
    in_split_min_examples_check: true
    missing_value_policy: GLOBAL_IMPUTATION
    allow_na_conditions: false
    categorical_set_greedy_forward {
      sampling: 0.1
      max_num_items: -1
      min_item_frequency: 1
    }
    growing_strategy_local {
    }
    categorical {
      cart {
      }
    }
    num_candidate_attributes_ratio: -1
    axis_aligned_split {
    }
    internal {
      sorting_strategy: PRESORTED
    }
  }
  winner_take_all_inference: true
  compute_oob_performances: true
  compute_oob_variable_importances: false
  adapt_bootstrap_size_ratio_for_maximum_training_duration: false
}

[INFO kernel.cc:790] Deployment config:
num_threads: 6

[INFO kernel.cc:817] Train model
[INFO random_forest.cc:315] Training random forest on 2500 example(s) and 10 feature(s).
[INFO random_forest.cc:628] Training of tree  1/1000 (tree index:1) done accuracy:0.783262 logloss:7.81204
[INFO random_forest.cc:628] Training of tree  11/1000 (tree index:9) done accuracy:0.801127 logloss:2.73187
[INFO random_forest.cc:628] Training of tree  21/1000 (tree index:19) done accuracy:0.811449 logloss:1.1286
[INFO random_forest.cc:628] Training of tree  31/1000 (tree index:32) done accuracy:0.8132 logloss:0.910787
[INFO random_forest.cc:628] Training of tree  41/1000 (tree index:42) done accuracy:0.812 logloss:0.745694
[INFO random_forest.cc:628] Training of tree  51/1000 (tree index:48) done accuracy:0.8144 logloss:0.690226
[INFO random_forest.cc:628] Training of tree  61/1000 (tree index:59) done accuracy:0.8136 logloss:0.659137
[INFO random_forest.cc:628] Training of tree  71/1000 (tree index:72) done accuracy:0.8176 logloss:0.577357
[INFO random_forest.cc:628] Training of tree  81/1000 (tree index:79) done accuracy:0.814 logloss:0.565115
[INFO random_forest.cc:628] Training of tree  91/1000 (tree index:91) done accuracy:0.8156 logloss:0.56459
[INFO random_forest.cc:628] Training of tree  101/1000 (tree index:99) done accuracy:0.8148 logloss:0.564104
[INFO random_forest.cc:628] Training of tree  111/1000 (tree index:109) done accuracy:0.8172 logloss:0.537417
[INFO random_forest.cc:628] Training of tree  121/1000 (tree index:120) done accuracy:0.8156 logloss:0.524543
[INFO random_forest.cc:628] Training of tree  131/1000 (tree index:132) done accuracy:0.8152 logloss:0.511111
[INFO random_forest.cc:628] Training of tree  141/1000 (tree index:141) done accuracy:0.816 logloss:0.498209
[INFO random_forest.cc:628] Training of tree  151/1000 (tree index:150) done accuracy:0.8192 logloss:0.485477
[INFO random_forest.cc:628] Training of tree  161/1000 (tree index:160) done accuracy:0.8196 logloss:0.472341
[INFO random_forest.cc:628] Training of tree  171/1000 (tree index:171) done accuracy:0.818 logloss:0.459903
[INFO random_forest.cc:628] Training of tree  181/1000 (tree index:182) done accuracy:0.8172 logloss:0.459812
[INFO random_forest.cc:628] Training of tree  191/1000 (tree index:190) done accuracy:0.8192 logloss:0.459588
[INFO random_forest.cc:628] Training of tree  201/1000 (tree index:199) done accuracy:0.818 logloss:0.459855
[INFO random_forest.cc:628] Training of tree  211/1000 (tree index:209) done accuracy:0.8176 logloss:0.459088
[INFO random_forest.cc:628] Training of tree  221/1000 (tree index:221) done accuracy:0.8168 logloss:0.43377
[INFO random_forest.cc:628] Training of tree  231/1000 (tree index:233) done accuracy:0.8196 logloss:0.433567
[INFO random_forest.cc:628] Training of tree  241/1000 (tree index:241) done accuracy:0.8208 logloss:0.434371
[INFO random_forest.cc:628] Training of tree  251/1000 (tree index:250) done accuracy:0.8192 logloss:0.434301
[INFO random_forest.cc:628] Training of tree  261/1000 (tree index:260) done accuracy:0.8172 logloss:0.43402
[INFO random_forest.cc:628] Training of tree  271/1000 (tree index:271) done accuracy:0.818 logloss:0.433583
[INFO random_forest.cc:628] Training of tree  281/1000 (tree index:283) done accuracy:0.8184 logloss:0.420657
[INFO random_forest.cc:628] Training of tree  291/1000 (tree index:291) done accuracy:0.8168 logloss:0.420481
[INFO random_forest.cc:628] Training of tree  301/1000 (tree index:299) done accuracy:0.82 logloss:0.419901
[INFO random_forest.cc:628] Training of tree  311/1000 (tree index:312) done accuracy:0.8188 logloss:0.419881
[INFO random_forest.cc:628] Training of tree  321/1000 (tree index:319) done accuracy:0.8172 logloss:0.419582
[INFO random_forest.cc:628] Training of tree  331/1000 (tree index:332) done accuracy:0.8176 logloss:0.419608
[INFO random_forest.cc:628] Training of tree  341/1000 (tree index:341) done accuracy:0.816 logloss:0.419608
[INFO random_forest.cc:628] Training of tree  351/1000 (tree index:352) done accuracy:0.8152 logloss:0.419729
[INFO random_forest.cc:628] Training of tree  361/1000 (tree index:361) done accuracy:0.8152 logloss:0.419264
[INFO random_forest.cc:628] Training of tree  371/1000 (tree index:369) done accuracy:0.8148 logloss:0.418932
[INFO random_forest.cc:628] Training of tree  381/1000 (tree index:379) done accuracy:0.8156 logloss:0.419148
[INFO random_forest.cc:628] Training of tree  391/1000 (tree index:391) done accuracy:0.8164 logloss:0.419344
[INFO random_forest.cc:628] Training of tree  401/1000 (tree index:398) done accuracy:0.8156 logloss:0.419051
[INFO random_forest.cc:628] Training of tree  411/1000 (tree index:408) done accuracy:0.8168 logloss:0.406486
[INFO random_forest.cc:628] Training of tree  421/1000 (tree index:420) done accuracy:0.8168 logloss:0.406477
[INFO random_forest.cc:628] Training of tree  431/1000 (tree index:430) done accuracy:0.816 logloss:0.406362
[INFO random_forest.cc:628] Training of tree  441/1000 (tree index:440) done accuracy:0.8172 logloss:0.406377
[INFO random_forest.cc:628] Training of tree  451/1000 (tree index:448) done accuracy:0.8176 logloss:0.406083
[INFO random_forest.cc:628] Training of tree  461/1000 (tree index:458) done accuracy:0.8172 logloss:0.406205
[INFO random_forest.cc:628] Training of tree  471/1000 (tree index:474) done accuracy:0.8168 logloss:0.406437
[INFO random_forest.cc:628] Training of tree  481/1000 (tree index:482) done accuracy:0.8184 logloss:0.406287
[INFO random_forest.cc:628] Training of tree  491/1000 (tree index:490) done accuracy:0.8172 logloss:0.40588
[INFO random_forest.cc:628] Training of tree  501/1000 (tree index:498) done accuracy:0.816 logloss:0.406036
[INFO random_forest.cc:628] Training of tree  511/1000 (tree index:508) done accuracy:0.8164 logloss:0.406053
[INFO random_forest.cc:628] Training of tree  521/1000 (tree index:524) done accuracy:0.8168 logloss:0.405945
[INFO random_forest.cc:628] Training of tree  531/1000 (tree index:530) done accuracy:0.816 logloss:0.405778
[INFO random_forest.cc:628] Training of tree  541/1000 (tree index:540) done accuracy:0.8156 logloss:0.405737
[INFO random_forest.cc:628] Training of tree  551/1000 (tree index:552) done accuracy:0.8156 logloss:0.406028
[INFO random_forest.cc:628] Training of tree  561/1000 (tree index:559) done accuracy:0.8164 logloss:0.406081
[INFO random_forest.cc:628] Training of tree  571/1000 (tree index:569) done accuracy:0.8152 logloss:0.405734
[INFO random_forest.cc:628] Training of tree  581/1000 (tree index:579) done accuracy:0.8172 logloss:0.393451
[INFO random_forest.cc:628] Training of tree  591/1000 (tree index:591) done accuracy:0.816 logloss:0.393428
[INFO random_forest.cc:628] Training of tree  601/1000 (tree index:603) done accuracy:0.8156 logloss:0.393545
[INFO random_forest.cc:628] Training of tree  611/1000 (tree index:609) done accuracy:0.8156 logloss:0.3934
[INFO random_forest.cc:628] Training of tree  621/1000 (tree index:620) done accuracy:0.8148 logloss:0.393539
[INFO random_forest.cc:628] Training of tree  631/1000 (tree index:629) done accuracy:0.8156 logloss:0.393731
[INFO random_forest.cc:628] Training of tree  641/1000 (tree index:641) done accuracy:0.8164 logloss:0.39383
[INFO random_forest.cc:628] Training of tree  651/1000 (tree index:649) done accuracy:0.8152 logloss:0.393724
[INFO random_forest.cc:628] Training of tree  661/1000 (tree index:659) done accuracy:0.8152 logloss:0.393764
[INFO random_forest.cc:628] Training of tree  671/1000 (tree index:670) done accuracy:0.816 logloss:0.393834
[INFO random_forest.cc:628] Training of tree  681/1000 (tree index:680) done accuracy:0.8156 logloss:0.393894
[INFO random_forest.cc:628] Training of tree  691/1000 (tree index:689) done accuracy:0.8152 logloss:0.393746
[INFO random_forest.cc:628] Training of tree  701/1000 (tree index:698) done accuracy:0.814 logloss:0.393743
[INFO random_forest.cc:628] Training of tree  711/1000 (tree index:708) done accuracy:0.8152 logloss:0.393294
[INFO random_forest.cc:628] Training of tree  721/1000 (tree index:721) done accuracy:0.816 logloss:0.393451
[INFO random_forest.cc:628] Training of tree  731/1000 (tree index:733) done accuracy:0.8164 logloss:0.393486
[INFO random_forest.cc:628] Training of tree  741/1000 (tree index:739) done accuracy:0.8156 logloss:0.393553
[INFO random_forest.cc:628] Training of tree  751/1000 (tree index:751) done accuracy:0.816 logloss:0.393731
[INFO random_forest.cc:628] Training of tree  761/1000 (tree index:758) done accuracy:0.8172 logloss:0.393635
[INFO random_forest.cc:628] Training of tree  771/1000 (tree index:769) done accuracy:0.8164 logloss:0.393584
[INFO random_forest.cc:628] Training of tree  781/1000 (tree index:779) done accuracy:0.8184 logloss:0.393728
[INFO random_forest.cc:628] Training of tree  791/1000 (tree index:789) done accuracy:0.8192 logloss:0.393858
[INFO random_forest.cc:628] Training of tree  801/1000 (tree index:800) done accuracy:0.8184 logloss:0.381756
[INFO random_forest.cc:628] Training of tree  811/1000 (tree index:813) done accuracy:0.82 logloss:0.38174
[INFO random_forest.cc:628] Training of tree  821/1000 (tree index:819) done accuracy:0.8196 logloss:0.381865
[INFO random_forest.cc:628] Training of tree  831/1000 (tree index:829) done accuracy:0.8172 logloss:0.381929
[INFO random_forest.cc:628] Training of tree  841/1000 (tree index:838) done accuracy:0.8164 logloss:0.382007
[INFO random_forest.cc:628] Training of tree  851/1000 (tree index:850) done accuracy:0.8172 logloss:0.382099
[INFO random_forest.cc:628] Training of tree  861/1000 (tree index:863) done accuracy:0.8172 logloss:0.381937
[INFO random_forest.cc:628] Training of tree  871/1000 (tree index:869) done accuracy:0.8168 logloss:0.382131
[INFO random_forest.cc:628] Training of tree  881/1000 (tree index:879) done accuracy:0.8188 logloss:0.381963
[INFO random_forest.cc:628] Training of tree  891/1000 (tree index:889) done accuracy:0.8192 logloss:0.382052
[INFO random_forest.cc:628] Training of tree  901/1000 (tree index:901) done accuracy:0.8184 logloss:0.382174
[INFO random_forest.cc:628] Training of tree  911/1000 (tree index:913) done accuracy:0.8192 logloss:0.382273
[INFO random_forest.cc:628] Training of tree  921/1000 (tree index:919) done accuracy:0.82 logloss:0.382407
[INFO random_forest.cc:628] Training of tree  931/1000 (tree index:929) done accuracy:0.8216 logloss:0.382277
[INFO random_forest.cc:628] Training of tree  941/1000 (tree index:939) done accuracy:0.8204 logloss:0.382434
[INFO random_forest.cc:628] Training of tree  951/1000 (tree index:951) done accuracy:0.8192 logloss:0.382444
[INFO random_forest.cc:628] Training of tree  961/1000 (tree index:959) done accuracy:0.8192 logloss:0.382497
[INFO random_forest.cc:628] Training of tree  971/1000 (tree index:969) done accuracy:0.8188 logloss:0.382592
[INFO random_forest.cc:628] Training of tree  981/1000 (tree index:979) done accuracy:0.8192 logloss:0.382657
[INFO random_forest.cc:628] Training of tree  991/1000 (tree index:989) done accuracy:0.8188 logloss:0.382671
[INFO random_forest.cc:628] Training of tree  1000/1000 (tree index:997) done accuracy:0.8192 logloss:0.38269
[INFO random_forest.cc:696] Final OOB metrics: accuracy:0.8192 logloss:0.38269
[INFO kernel.cc:828] Export model in log directory: /tmp/tmp0r9hhl7d
[INFO kernel.cc:836] Save model in resources
[INFO kernel.cc:988] Loading model from path
40/40 [==============================] - 3s 64ms/step
[INFO decision_forest.cc:590] Model loaded with 1000 root(s), 324942 node(s), and 10 input feature(s).
[INFO kernel.cc:848] Use fast generic engine
CPU times: user 21.5 s, sys: 755 ms, total: 22.2 s
Wall time: 10.5 s
<keras.callbacks.History at 0x7f6b7874c4d0>

Et évaluons les forêts de décision individuellement.

model_3.compile(["accuracy"])
model_4.compile(["accuracy"])

evaluation_df3_only = model_3.evaluate(
    test_dataset_with_preprocessing, return_dict=True)
evaluation_df4_only = model_4.evaluate(
    test_dataset_with_preprocessing, return_dict=True)

print("Accuracy (DF #3 only): ", evaluation_df3_only["accuracy"])
print("Accuracy (DF #4 only): ", evaluation_df4_only["accuracy"])

157/157 [==============================] - 2s 8ms/step - loss: 0.0000e+00 - accuracy: 0.8218
157/157 [==============================] - 1s 8ms/step - loss: 0.0000e+00 - accuracy: 0.8223
Accuracy (DF #3 only):  0.8217999935150146
Accuracy (DF #4 only):  0.8223000168800354

Évaluons l'ensemble de la composition du modèle :

ensemble_nn_and_df.compile(
    loss=tf.keras.losses.BinaryCrossentropy(), metrics=["accuracy"])

evaluation_nn_and_df = ensemble_nn_and_df.evaluate(
    test_dataset, return_dict=True)

print("Accuracy (2xNN and 2xDF): ", evaluation_nn_and_df["accuracy"])
print("Loss (2xNN and 2xDF): ", evaluation_nn_and_df["loss"])

157/157 [==============================] - 2s 8ms/step - loss: 0.3707 - accuracy: 0.8236
Accuracy (2xNN and 2xDF):  0.8235999941825867
Loss (2xNN and 2xDF):  0.3706760108470917

Pour finir, affinons un peu plus la couche de réseau neuronal. Notez que nous n'affinons pas l'intégration pré-entraînée car les modèles DF en dépendent (à moins que nous ne les réformions également après).

En résumé, vous avez :

print(f"Accuracy (NN #1 and #2 only):\t{evaluation_nn_only['accuracy']:.6f}")
print(f"Accuracy (DF #3 only):\t\t{evaluation_df3_only['accuracy']:.6f}")
print(f"Accuracy (DF #4 only):\t\t{evaluation_df4_only['accuracy']:.6f}")
print("----------------------------------------")
print(f"Accuracy (2xNN and 2xDF):\t{evaluation_nn_and_df['accuracy']:.6f}")


def delta_percent(src_eval, key):
  src_acc = src_eval["accuracy"]
  final_acc = evaluation_nn_and_df["accuracy"]
  increase = final_acc - src_acc
  print(f"\t\t\t\t  {increase:+.6f} over {key}")


delta_percent(evaluation_nn_only, "NN #1 and #2 only")
delta_percent(evaluation_df3_only, "DF #3 only")
delta_percent(evaluation_df4_only, "DF #4 only")

Accuracy (NN #1 and #2 only): 0.820300
Accuracy (DF #3 only):        0.821800
Accuracy (DF #4 only):        0.822300
----------------------------------------
Accuracy (2xNN and 2xDF): 0.823600
                  +0.003300 over NN #1 and #2 only
                  +0.001800 over DF #3 only
                  +0.001300 over DF #4 only

Ici, vous pouvez voir que le modèle composé fonctionne mieux que ses parties individuelles. C'est pourquoi les ensembles fonctionnent si bien.

Et après?

Dans cet exemple, vous avez vu comment combiner des forêts de décision avec des réseaux de neurones. Une étape supplémentaire serait de poursuivre la formation du réseau de neurones et des forêts de décision ensemble.

De plus, par souci de clarté, les forêts de décision n'ont reçu que l'entrée prétraitée. Cependant, les forêts décisionnelles sont généralement grandes et consomment des données brutes. Le modèle serait amélioré en fournissant également les caractéristiques brutes aux modèles de forêt de décision.

Dans cet exemple, le modèle final est la moyenne des prédictions des modèles individuels. Cette solution fonctionne bien si tous les modèles fonctionnent plus ou moins avec le même. Cependant, si l'un des sous-modèles est très bon, son agrégation avec d'autres modèles peut en fait être préjudiciable (ou vice-versa ; par exemple, essayez de réduire le nombre d'exemples de 1k et voyez comment cela nuit beaucoup aux réseaux de neurones ; ou permettre à la SPARSE_OBLIQUE division du second modèle Random Forest).