Camuslu
Reputation: 123
sklearn Boosting: cross-validation to find optimal number of estimators without restarting everytime
In Python sklearn ensemble library, I want to train my data using some boosting method (say Adaboost). As I would like to know the optimal number of estimators, I plan to do a cv with different number of estimators each time. However, it seems doing it the following way is redundant:
for n in [50,100,150,200,250,300]:
model = AdaBoostClassifier(DecisionTreeClassifier(max_depth=1),n_estimators=n)
cross_val_score(model,x,y,k=5)
Because in AdaBoost, once I train the classifier on # of estimator=50, as I move along to train # of estimator=100, the first 50 classifiers and their weights don't change. I wonder if there is a way to start training directly with the 51st weak learner in this case.
Upvotes: 2
Views: 4628
Answers (2)
Matthew McGonagle
Reputation: 141
It is possible to use inheritance to make a "hack" of AdaBoostClassifier that doesn't retrain estimators and is compatible with many cross-validation functions in sklearn (must be cross-validation that doesn't shuffle data).
If you look at the source code in sklearn.ensemble.weight_boosting.py, you can see that you can get away with not needing to retrain estimators if you properly wrap the behavior of AdaBoostClassifier.fit() and AdaBoostClassifier._boost().
The problem with cross validation functions is that they make clones of the original estimator using sklearn.base.clone(), and in turn the function sklearn.base.clone() makes deep copies of the estimator's parameters. The deep copy nature makes it impossible for the estimator to "remember" its estimators between different cross-validation runs (clone() copies the contents of a reference and not the reference itself). The only way to do so (at least the only way I can think of) is to use global state to keep track of old estimators between runs. The catch here is that you have to compute a hash of your X features, which could be expensive!
Anyways, here is the hack to AdaBoostClassifier itself:
'''
adaboost_hack.py
Make a "hack" of AdaBoostClassifier in sklearn.ensemble.weight_boosting.py
that doesn't need to retrain estimators and is compatible with many sklearn
cross validation functions.
'''
import copy
import numpy as np
from sklearn.ensemble import AdaBoostClassifier
from sklearn.base import clone
# Used to hold important variables between runs of cross validation.
# Note that sklearn cross validation functions use sklearn.base.clone()
# to make copies of the estimator sent to it as a function. The function
# sklearn.base.clone() makes deep copies of parameters of an estimator, so
# the only way to provide a way to remember previous estimators between
# cross validation runs is to use a global variable.
#
# We will use hash values of the split of X[:, 0] as keys for remembering
# previous estimators of a cv fold. Note, you can NOT use cross validators
# that randomly shuffle the data before splitting. This will cause different
# hashes.
kfold_hash = {}
class WarmRestartAdaBoostClassifier(AdaBoostClassifier):
'''
Keep track of old estimators, estimator weights, the estimator errors, and
the next to last sample weight seen.
Note that AdaBoostClassifier._boost() does NOT boost the last seen sample
weight. Simple fix to this is to drop the last estimator and retrain it.
Wrap AdaBoostClassifier.fit() to decide whether to throw away estimators or add estimators
depending on the current number of estimators vs the number of old esimators.
Also look at the possibility of use the global kfold_hash to get old values if
use_kfold_hash == True.
Wrap AdaBoostClassifier._boost() with behavior to record the next to last sample weight.
'''
def __init__(self,
base_estimator=None,
n_estimators=50,
learning_rate=1.,
algorithm='SAMME.R',
random_state=None,
next_to_last_sample_weight = None,
old_estimators_ = [],
use_kfold_hash = False):
AdaBoostClassifier.__init__(self, base_estimator, n_estimators, learning_rate,
algorithm, random_state)
self.next_to_last_sample_weight = next_to_last_sample_weight
self._last_sample_weight = None
self.old_estimators_ = old_estimators_
self.use_kfold_hash = use_kfold_hash
def _boost(self, iboost, X, y, sample_weight, random_state):
'''
Record the sample weight.
Parameters and return behavior same as that of AdaBoostClassifier._boost() as
seen in sklearn.ensemble.weight_boosting.py
Parameters
----------
iboost : int
The index of the current boost iteration.
X : {array-like, sparse matrix} of shape = [n_samples, n_features]
The training input samples. Sparse matrix can be CSC, CSR, COO,
DOK, or LIL. COO, DOK, and LIL are converted to CSR.
y : array-like of shape = [n_samples]
The target values (class labels).
sample_weight : array-like of shape = [n_samples]
The current sample weights.
random_state : RandomState
The current random number generator
Returns
-------
sample_weight : array-like of shape = [n_samples] or None
The reweighted sample weights.
If None then boosting has terminated early.
estimator_weight : float
The weight for the current boost.
If None then boosting has terminated early.
error : float
The classification error for the current boost.
If None then boosting has terminated early.
'''
fit_info = AdaBoostClassifier._boost(self, iboost, X, y, sample_weight, random_state)
sample_weight, _, _ = fit_info
self.next_to_last_sample_weight = self._last_sample_weight
self._last_sample_weight = sample_weight
return fit_info
def fit(self, X, y):
hash_X = None
if self.use_kfold_hash:
# Use a hash of X features in this kfold to access the global information
# for this kfold.
hash_X = hash(bytes(X[:, 0]))
if hash_X in kfold_hash.keys():
self.old_estimators_ = kfold_hash[hash_X]['old_estimators_']
self.next_to_last_sample_weight = kfold_hash[hash_X]['next_to_last_sample_weight']
self.estimator_weights_ = kfold_hash[hash_X]['estimator_weights_']
self.estimator_errors_ = kfold_hash[hash_X]['estimator_errors_']
# We haven't done any fits yet.
if not self.old_estimators_:
AdaBoostClassifier.fit(self, X, y)
self.old_estimators_ = self.estimators_
# The case that we throw away estimators.
elif self.n_estimators < len(self.old_estimators_):
self.estimators_ = self.old_estimators_[:self.n_estimators]
self.estimator_weights_ = self.estimator_weights_[:self.n_estimators]
self.estimator_errors_ = self.estimator_errors_[:self.n_estimators]
# The case that we add new estimators.
elif self.n_estimators > len(self.old_estimators_):
n_more = self.n_estimators - len(self.old_estimators_)
self.fit_more(X, y, n_more)
# Record information in the global hash if necessary.
if self.use_kfold_hash:
kfold_hash[hash_X] = {'old_estimators_' : self.old_estimators_,
'next_to_last_sample_weight' : self.next_to_last_sample_weight,
'estimator_weights_' : self.estimator_weights_,
'estimator_errors_' : self.estimator_errors_}
return self
def fit_more(self, X, y, n_more):
'''
Fits additional estimators.
'''
# Since AdaBoostClassifier._boost() doesn't boost the last sample weight, we retrain the last estimator with
# its input sample weight.
self.n_estimators = n_more + 1
if self.old_estimators_ is None:
raise Exception('Should have already fit estimators before calling fit_more()')
self.old_estimators_ = self.old_estimators_[:-1]
old_estimator_weights = self.estimator_weights_[:-1]
old_estimator_errors = self.estimator_errors_[:-1]
sample_weight = self.next_to_last_sample_weight
AdaBoostClassifier.fit(self, X, y, sample_weight)
self.old_estimators_.extend(self.estimators_)
self.estimators_ = self.old_estimators_
self.n_estimators = len(self.estimators_)
self.estimator_weights_ = np.concatenate([old_estimator_weights, self.estimator_weights_])
self.estimator_errors_ = np.concatenate([old_estimator_errors, self.estimator_errors_])
And here is an example that allows you to compare timings/accuracies of the hack compared to the original AdaBoostClassifier. Note, that testing hack will have increasing time as we add estimators, but the training will not. I found the hack to run much faster than the original, but I'm not hashing large amount of X samples.
'''
example.py
Test the AdaBoost hack.
'''
import time # Used to get timing info.
import adaboost_hack
import numpy as np
import matplotlib.pyplot as plt
from sklearn.tree import DecisionTreeClassifier # We will use stumps for our classifiers.
from sklearn.ensemble import AdaBoostClassifier # Used to compare hack to original.
from sklearn.model_selection import (cross_val_score, KFold)
from sklearn.metrics import accuracy_score
my_random = np.random.RandomState(0) # For consistent results.
nSamples = 2000
# Make some sample data.
X = my_random.uniform(size = (nSamples, 2))
y = np.zeros(len(X), dtype = int)
# Decision boundary is the unit circle.
in_class = X[:, 0]**2 + X[:, 1]**2 > 1
y = np.zeros(len(X), dtype = int)
y[in_class] = 1
# Add some random error.
error_rate = 0.01
to_flip = my_random.choice(np.arange(len(y)), size = int(error_rate * len(y)), replace = False)
y[to_flip] = 1 - y[to_flip]
# Plot the data.
plt.scatter(X[:, 0], X[:, 1], c = y)
plt.title('Simulated Data')
plt.show()
# Make our hack solution. Initially do 2 estimators.
# Train the hack without testing. Should find nearly constant time per training session.
print('Training hack without testing.')
ada_boost_hack = adaboost_hack.WarmRestartAdaBoostClassifier(DecisionTreeClassifier(max_depth = 1,
random_state = my_random),
n_estimators = 1,
random_state = my_random)
nFit = 50
times = []
for i in range(nFit):
times.append(time.time())
ada_boost_hack.n_estimators += 1
ada_boost_hack.fit(X, y)
def get_differences(times):
times = np.array(times)
return times[1:] - times[:-1]
times_per_train = {'hack no test' : get_differences(times)}
# Now look at running tests while training the hack. Should have small linear growth between
# in time per training session.
print('Training hack with testing.')
ada_boost_hack = adaboost_hack.WarmRestartAdaBoostClassifier(DecisionTreeClassifier(max_depth = 1,
random_state = my_random),
n_estimators = 1,
random_state = my_random)
times = []
scores = []
for i in range(nFit):
times.append(time.time())
ada_boost_hack.n_estimators += 1
ada_boost_hack.fit(X, y)
y_predict = ada_boost_hack.predict(X)
new_score = accuracy_score(y, y_predict)
scores.append(new_score)
plt.plot(scores)
plt.title('Training scores for hack')
plt.ylabel('Accuracy')
plt.show()
times_per_train['hack with test'] = get_differences(times)
print('Now training hack with cross validation')
ada_boost_hack = adaboost_hack.WarmRestartAdaBoostClassifier(DecisionTreeClassifier(max_depth = 1,
random_state = my_random),
n_estimators = 1,
random_state = my_random,
use_kfold_hash = True)
# Now try cross_val_score().
scores = []
times = []
# We use KFold to make sure the hashes of X features of each fold are
# the same between each run.
for i in range(1, nFit + 1):
ada_boost_hack.set_params(n_estimators = i)
new_scores = cross_val_score(ada_boost_hack, X, y, cv = KFold(3))
scores.append(new_scores)
times.append(time.time())
def plot_cv_scores(scores):
scores = np.array(scores)
plt.plot(scores.mean(axis = 1))
plt.plot(scores.mean(axis = 1) + scores.std(axis = 1) * 2, color = 'red')
plt.plot(scores.mean(axis = 1) - scores.std(axis = 1) * 2, color = 'red')
plt.ylabel('Accuracy')
plot_cv_scores(scores)
plt.title('Cross validation scores for hack')
plt.show()
times_per_train['hack cross validation'] = get_differences(times)
# Double check that kfold_hash only has 3 keys since we used cv = 3.
print('adaboost_hack.keys() = ', adaboost_hack.kfold_hash.keys())
# Now get timings for original classifier.
print('Now doing cross validations of original')
ada_boost = AdaBoostClassifier(DecisionTreeClassifier(max_depth = 1,
random_state = np.random.RandomState(0)),
n_estimators = 1,
random_state = np.random.RandomState(0))
times = []
scores = []
# We use KFold to make sure the hashes of X features of each fold are
# the same between each run.
for i in range(1, nFit + 1):
ada_boost.set_params(n_estimators = i)
new_scores = cross_val_score(ada_boost, X, y, cv = KFold(3))
scores.append(new_scores)
times.append(time.time())
plot_cv_scores(scores)
plt.title('Cross validation scores for original')
plt.show()
times_per_train['original cross validation'] = get_differences(times)
# Plot all of the timing data.
for key in times_per_train.keys():
plt.plot(times_per_train[key])
plt.title('Time per training or cv score')
plt.ylabel('Time')
plt.xlabel('nth training or cv score')
plt.legend(times_per_train.keys())
plt.show()
Upvotes: 4
Matthew McGonagle
Reputation: 141
You can fit all 300 estimators and then use AdaBoostClassifier.staged_predict() to track how the error rate depends on the number of estimators. However, you will have to do the cross-validation splits yourself; I don't think it is compatible with cross_val_score().
For example,
from sklearn.ensemble import AdaBoostClassifier
from sklearn.tree import DecisionTreeClassifier # We will use simple stumps for individual estimators in AdaBoost.
from sklearn.metrics import accuracy_score
import numpy as np
import matplotlib.pyplot as plt
np.random.seed(0)
nSamples = {'train' : 2000, 'test' : 1000}
X = np.random.uniform(size = (nSamples['train'] + nSamples['test'], 2))
# Decision boundary is the unit circle.
in_class = X[:, 0]**2 + X[:, 1]**2 > 1
y = np.zeros(len(X), dtype = int)
y[in_class] = 1
# Add some random error.
error_rate = 0.01
to_flip = np.random.choice(np.arange(len(y)), size = int(error_rate * len(y)), replace = False)
y[to_flip] = 1 - y[to_flip]
# Split training and test.
X = {'train' : X[:nSamples['train']],
'test' : X[nSamples['train']:]}
y = {'train' : y[:nSamples['train']],
'test' : y[nSamples['train']:]}
# Make AdaBoost Classifier.
max_estimators = 50
ada_boost = AdaBoostClassifier(DecisionTreeClassifier(max_depth = 1, # Just a stump.
random_state = np.random.RandomState(0)),
n_estimators = max_estimators,
random_state = np.random.RandomState(0))
# Fit all estimators.
ada_boost.fit(X['train'], y['train'])
# Get the test accuracy for each stage of prediction.
scores = {'train' : [], 'test' : []}
for y_predict_train, y_predict_test in zip(ada_boost.staged_predict(X['train']),
ada_boost.staged_predict(X['test'])):
scores['train'].append(accuracy_score(y['train'], y_predict_train))
scores['test'].append(accuracy_score(y['test'], y_predict_test))
# Plot the results.
n_estimators = range(1, len(scores['train']) + 1)
for key in scores.keys():
plt.plot(n_estimators, scores[key])
plt.title('Staged Scores')
plt.ylabel('Accuracy')
plt.xlabel('N Estimators')
plt.legend(scores.keys())
plt.show()
Upvotes: 0