Composizione di modelli Decision Forest e Neural Network

Visualizza su TensorFlow.org Esegui in Google Colab Visualizza su GitHub Scarica taccuino API funzionale Keras

introduzione

Benvenuti al modello composizione tutorial per foreste decisione tensorflow (TF-DF). Questo notebook vi mostra come comporre più insiemi di strutture decisionali e modelli di reti neurali insieme con uno strato di pre-elaborazione comune e l' API funzionale Keras .

Potresti voler comporre insieme modelli per migliorare le prestazioni predittive (assemblaggio), per ottenere il meglio da diverse tecnologie di modellazione (assemblaggio di modelli eterogenei), per addestrare parti diverse del modello su diversi set di dati (ad esempio pre-addestramento) o per creare un modello impilato (es. un modello opera sulle previsioni di un altro modello).

Questo tutorial copre un caso d'uso avanzato della composizione del modello utilizzando l'API funzionale. Si possono trovare esempi per gli scenari più semplici di composizione modello nella sezione di questo "funzione di pre-elaborazione" esercitazione e nella sezione "usando un testo preaddestrato embedding" di questa esercitazione .

Ecco la struttura del modello che costruirai:

svg

Il tuo modello composto ha tre fasi:

  1. La prima fase è un livello di pre-elaborazione composto da una rete neurale e comune a tutti i modelli nella fase successiva. In pratica, un tale livello di pre-elaborazione potrebbe essere un inserimento pre-addestrato per la messa a punto o una rete neurale inizializzata in modo casuale.
  2. La seconda fase è un insieme di due foreste decisionali e due modelli di rete neurale.
  3. L'ultima fase fa la media delle previsioni dei modelli nella seconda fase. Non contiene pesi apprendibili.

Le reti neurali sono addestrati utilizzando l' algoritmo di backpropagation e la discesa del gradiente. Questo algoritmo ha due proprietà importanti: (1) Il livello della rete neurale può essere addestrato se riceve un gradiente di perdita (più precisamente, il gradiente di perdita in base all'output del livello), e (2) l'algoritmo "trasmette" il gradiente di perdita dall'output del livello all'input del livello (questa è la "regola della catena"). Per questi due motivi, Backpropagation può addestrare insieme più livelli di reti neurali impilati uno sopra l'altro.

In questo esempio, le foreste di decisione sono addestrati con la foresta casuale algoritmo (RF). A differenza della retropropagazione, l'addestramento della RF non "trasmette" il gradiente di perdita dalla sua uscita al suo ingresso. Per questo motivo, il classico algoritmo RF non può essere utilizzato per addestrare o mettere a punto una rete neurale sottostante. In altre parole, le fasi della "foresta decisionale" non possono essere utilizzate per addestrare il "blocco di pre-elaborazione NN apprendibile".

  1. Addestrare la fase di pre-elaborazione e reti neurali.
  2. Allena le fasi decisionali della foresta.

Installa TensorFlow Decision Forests

Installa TF-DF eseguendo la seguente cella.

pip install tensorflow_decision_forests -U --quiet

Installare Wurlitzer per mostrare i registri dettagliati di tale formazione. Questo è necessario solo nei notebook.

pip install wurlitzer -U --quiet

Importa librerie

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.

set di dati

In questo tutorial utilizzerai un semplice set di dati sintetico per semplificare l'interpretazione del modello finale.

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)

Genera alcuni esempi:

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]))

Puoi anche tracciarli per avere un'idea del modello sintetico:

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

Si noti che questo modello è liscio e non allineato all'asse. Ciò andrà a vantaggio dei modelli di rete neurale. Questo perché è più facile per una rete neurale che per un albero decisionale avere confini decisionali rotondi e non allineati.

D'altra parte, addestreremo il modello su un piccolo set di dati con 2500 esempi. Ciò andrà a vantaggio dei modelli di foresta decisionale. Questo perché le foreste decisionali sono molto più efficienti, utilizzando tutte le informazioni disponibili dagli esempi (le foreste decisionali sono "campione efficiente").

Il nostro insieme di reti neurali e foreste decisionali utilizzerà il meglio di entrambi i mondi.

Creiamo un treno e di prova 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)

Struttura del modello

Definire la struttura del modello come segue:

# 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)

Prima di addestrare il modello, puoi tracciarlo per verificare se è simile al diagramma iniziale.

from keras.utils.vis_utils import plot_model

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

png

Formazione modello

Per prima cosa addestrare il preprocessing e due livelli di rete neurale utilizzando l'algoritmo di backpropagation.

%%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>

Valutiamo la preelaborazione e la parte con le sole due reti neurali:

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

Addestriamo i due componenti Decision Forest (uno dopo l'altro).

%%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>

E valutiamo individualmente le Decision Forests.

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

Valutiamo l'intera composizione del modello:

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

Per finire, perfezioniamo un po' di più il livello della rete neurale. Notare che non ottimizziamo l'incorporamento pre-addestrato poiché i modelli DF dipendono da esso (a meno che non li riaddestraremmo anche in seguito).

In sintesi, hai:

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

Qui puoi vedere che il modello composto funziona meglio delle sue singole parti. Questo è il motivo per cui gli ensemble funzionano così bene.

Qual è il prossimo?

In questo esempio, hai visto come combinare foreste decisionali con reti neurali. Un ulteriore passo sarebbe quello di addestrare ulteriormente la rete neurale e le foreste decisionali insieme.

Inoltre, per motivi di chiarezza, le foreste decisionali hanno ricevuto solo l'input preelaborato. Tuttavia, le foreste decisionali sono generalmente grandi e consumano dati grezzi. Il modello verrebbe migliorato fornendo anche le caratteristiche grezze ai modelli della foresta decisionale.

In questo esempio, il modello finale è la media delle previsioni dei singoli modelli. Questa soluzione funziona bene se tutto il modello esegue più o meno con lo stesso. Tuttavia, se uno dei sottomodelli è molto buono, aggregarlo con altri modelli potrebbe effettivamente essere dannoso (o viceversa; ad esempio prova a ridurre il numero di esempi da 1k e vedi come danneggia molto le reti neurali; o abilitare lo SPARSE_OBLIQUE spaccatura nel secondo modello Foresta casuale).