we all know the common approach to define a pipeline with a dimensionality reduction technique and then a model for training and testing. Then we can apply the GridSearchCv for hyperparameter tuning.
grid = GridSearchCV(
Pipeline([
('reduce_dim', PCA()),
('classify', RandomForestClassifier(n_jobs = -1))
]),
param_grid=[
{
'reduce_dim__n_components': range(0.7,0.9,0.1),
'classify__n_estimators': range(10,50,5),
'classify__max_features': ['auto', 0.2],
'classify__min_samples_leaf': [40,50,60],
'classify__criterion': ['gini', 'entropy']
}
],
cv=5, scoring='f1')
grid.fit(X,y)
I can understand the above code.
Now i was going through the documentation today and there i found one part code which is little bit strange.
pipe = Pipeline([
# the reduce_dim stage is populated by the param_grid
('reduce_dim', 'passthrough'), # How does this work??
('classify', LinearSVC(dual=False, max_iter=10000))
])
N_FEATURES_OPTIONS = [2, 4, 8]
C_OPTIONS = [1, 10, 100, 1000]
param_grid = [
{
'reduce_dim': [PCA(iterated_power=7), NMF()],
'reduce_dim__n_components': N_FEATURES_OPTIONS, ### No PCA is used..??
'classify__C': C_OPTIONS
},
{
'reduce_dim': [SelectKBest(chi2)],
'reduce_dim__k': N_FEATURES_OPTIONS,
'classify__C': C_OPTIONS
},
]
reducer_labels = ['PCA', 'NMF', 'KBest(chi2)']
grid = GridSearchCV(pipe, n_jobs=1, param_grid=param_grid)
X, y = load_digits(return_X_y=True)
grid.fit(X, y)
First of all while defining a pipeline, it used a string 'passthrough' instead of a object.
('reduce_dim', 'passthrough'), ```
Then while defining different dimensionality reduction technique for the grid search, it used a different strategy. How does [PCA(iterated_power=7), NMF()] this work ?
'reduce_dim': [PCA(iterated_power=7), NMF()],
'reduce_dim__n_components': N_FEATURES_OPTIONS, # here
Please Someone explain the code to me .
Solved - in one line, the order is ['PCA', 'NMF', 'KBest(chi2)']
Courtesy of - seralouk (see answer below)
For Reference If someone looks for more details
1 2 3
It is equivalent as far as I know.
In the documentation you have this:
pipe = Pipeline([
# the reduce_dim stage is populated by the param_grid
('reduce_dim', 'passthrough'),
('classify', LinearSVC(dual=False, max_iter=10000))
])
N_FEATURES_OPTIONS = [2, 4, 8]
C_OPTIONS = [1, 10, 100, 1000]
param_grid = [
{
'reduce_dim': [PCA(iterated_power=7), NMF()],
'reduce_dim__n_components': N_FEATURES_OPTIONS,
'classify__C': C_OPTIONS
},
{
'reduce_dim': [SelectKBest(chi2)],
'reduce_dim__k': N_FEATURES_OPTIONS,
'classify__C': C_OPTIONS
},
]
Initially we have ('reduce_dim', 'passthrough'), and then 'reduce_dim': [PCA(iterated_power=7), NMF()]
The definition of the PCA is done in the second line.
You could define alternatively:
pipe = Pipeline([
# the reduce_dim stage is populated by the param_grid
('reduce_dim', PCA(iterated_power=7)),
('classify', LinearSVC(dual=False, max_iter=10000))
])
N_FEATURES_OPTIONS = [2, 4, 8]
C_OPTIONS = [1, 10, 100, 1000]
param_grid = [
{
'reduce_dim__n_components': N_FEATURES_OPTIONS,
'classify__C': C_OPTIONS
},
{
'reduce_dim': [SelectKBest(chi2)],
'reduce_dim__k': N_FEATURES_OPTIONS,
'classify__C': C_OPTIONS
},
]
Related
I am trying to run this code on my dataset. However, its taking too long (the code has been running since yesterday and my dataset is not too large)
# A parameter grid for XGBoost
params = {
'min_child_weight': [1, 5, 10],
'gamma': [0.5, 1, 1.5, 2, 5],
'subsample': [0.6, 0.8, 1.0],
'colsample_bytree': [0.6, 0.8, 1.0],
'max_depth': [3, 4, 5]
}
xgb_classifier = XGBClassifier(learning_rate=0.02, n_estimators=600,
objective='binary:logistic',silent=True, nthread=1)
folds = 3
param_comb = 5
skf = StratifiedKFold(n_splits=folds, shuffle = True, random_state = 1001)
random_search = RandomizedSearchCV(xgb_classifier, param_distributions=params,
n_iter=param_comb, scoring='roc_auc', n_jobs=4, cv=skf.split(trainx,trainy),
verbose=3, random_state=1001 )
random_search.fit(trainx, trainy)
I am new in the field. any ideas?
I am playing around with the XGBoostClassifier and tuning this with GridSearchCV. I first created the variable xgbc:
xgbc = xgb.XGBClassifier()
I did'nt use any parameters as I wanted to see the default model performance. This gave me accuracy_score = 85.65%, recall_score = 77.91% and roc_auc_score = 84.21%, using the following lines of code:
print("Accuracy: ", accuracy_score(y_test, xgbc.predict(X_test)))
print("Recall: ", recall_score(y_test, xgbc.predict(X_test)))
print("ROC_AUC: ", roc_auc_score(y_test, xgbc.predict(X_test)))
Next, I used GridSearchCV to try to tune the parameters, like this:
Setting up the parameter dictionary:
xgbc_params = {'max_depth': [5, 6, 7], #6
'learning_rate': [0.25, 0.300000012, 0.35], #0.300000012
'gamma':[0, 0.001, 0.1], #0
'reg_lambda': [0.8, 0.95, 1], #1
'scale_pos_weight': [0, 1, 2], #1
'n_estimators': [95, 100, 105]} #100
(The numbers after the # are the default values, which gave me the above scores.)
And now run the GridSearchCV like this:
xgbc_grid = GridSearchCV(xgbc, param_grid=xgbc_params, scoring = make_scorer(accuracy_score), cv = 10, n_jobs = -1)
Next, fit this to the training data:
xgbc_grid.fit(X_train, y_train, verbose = 1, early_stopping_rounds = 10, eval_metric = 'aucpr', eval_set = [(X_test, y_test)])
Finally, run the metrics again:
print("Best Reg estimators: ", xgbc_grid.best_params_)
print("Accuracy: ", accuracy_score(y_test, xgbc_grid.predict(X_test)))
print("Recall: ", recall_score(y_test, xgbc_grid.predict(X_test)))
print("ROC_AUC: ", roc_auc_score(y_test, xgbc_grid.predict(X_test)))
Now, the scores change: accuracy_score = 0.8340807174887892, recall_score = 0.7325581395348837 and roc_auc_score = 0.8420896282464777. Also, here is the best_params_ result:
Best Reg estimators: {'gamma': 0, 'learning_rate': 0.35, 'max_depth': 5, 'n_estimators': 95, 'reg_lambda': 0.8, 'scale_pos_weight': 1}
Here is my problem:
The parameter values that GridSearchCV returns through xgbc_grid.best_params_ are not the most optimal for accuracy, as the accuracy score decreases. Can you please help me figure out why this is happenning?
In the parameter dictionary above, I have provided the default values. If I set the parameters to only these single values, then I get the 85% accuracy, like, 'max_depth': [6]. However, as soon as I add other values, like 'max_depth': [5, 6, 7], then GridSearchCV gives the parameters that are not the highest on accuracy score. Full details below:
Base Reg estimators (acc = 85%): {'gamma': 0, 'learning_rate': 0.35, 'max_depth': 5, 'n_estimators': 95, 'reg_lambda': 0.8, 'scale_pos_weight': 1}
Best Reg estimators (acc = 83%): {'gamma': 0, 'learning_rate': 0.35, 'max_depth': 6, 'n_estimators': 100, 'reg_lambda': 1, 'scale_pos_weight': 1}
After using k-means i have 3 clusters.
I've used 10 features (marks) in k-means for this data set.
I'm understand that we can't draw 10D chart, but how can i visualize this clusters?
Should i separate data by 2 or 3 features instead 10?
What axises should i use in my case?
For drawing i'm using js and highcharts.js on client side.
Example of code (just for stackoverflow requirement), but I have 10 coordinates for every point
const kmeans = require('ml-kmeans');
let data = [[1, 1, 1, 1, 1], [1, 2, 1, 1, 1], [-1, -1, -1, 1, 1], [-1, -1, -1.5, 1, 1]];
let centers = [[1, 2, 1, 1, 1], [-1, -1, -1, 1, 1]];
let ans = kmeans(data, 2, { initialization: centers });
console.log(ans);
/*KMeansResult*/
{
clusters: [ 0, 0, 1, 1, 1 ]
centroids:
[ { centroid: [ 1, 1.5, 1, 1, 1 ], error: 0.25, size: 2 },
{ centroid: [ -1, -1, -1.25, 1, 1 ], error: 0.0625, size: 2 } ],
converged: true, iterations: 1
}
*/*
Use your favorite generic visualization approach. Clusterings do not have very special requirements.
E.g.
Scatterplot matrix
Dimensionality reduction with PCA
tSNE embeddings
MDS
UMAP
Boxplots
Violin plots
...
I'm trying to calculate probabilities for a multi-class dataset using scikit learn. However, for some reason, I'm getting a the same probabilities for every example. Any idea what's happening? Does this have to do with my model, my use of the library, or something else? Appreciate any help!
svm_model = svm.SVC(probability=True, kernel='rbf',C=1, decision_function_shape='ovr', gamma=0.001,verbose=100)
svm_model.fit(train_X,train_y)
preds= svm_model.predict_proba(test_X)
train_X looks like this
array([[2350, 5550, 2750.0, ..., 23478, 1, 3],
[2500, 5500, 3095.5, ..., 23674, 0, 3],
[3300, 6900, 3600.0, ..., 6529, 0, 3],
...,
[2150, 6175, 2500.0, ..., 11209, 0, 3],
[2095, 5395, 2595.4, ..., 10070, 0, 3],
[1650, 2850, 2000.0, ..., 25463, 1, 3]], dtype=object)
train_y looks like this
0 1
1 2
10 2
100 2
1000 2
10000 2
10001 2
10002 2
10003 2
10004 2
10005 2
10006 2
10007 2
10008 1
10009 1
1001 2
10010 2
test_X looks like this
array([[2190, 3937, 2200.5, ..., 24891, 1, 5],
[2695, 7000, 2850.0, ..., 5491, 1, 4],
[2950, 12000, 4039.5, ..., 22367, 0, 4],
...,
[2850, 5200, 3000.0, ..., 15576, 1, 1],
[3200, 16000, 4100.0, ..., 1320, 0, 3],
[2100, 3750, 2400.0, ..., 6022, 0, 1]], dtype=object)
My results look like
array([[ 0.07819139, 0.22727628, 0.69453233],
[ 0.07819139, 0.22727628, 0.69453233],
[ 0.07819139, 0.22727628, 0.69453233],
...,
[ 0.07819139, 0.22727628, 0.69453233],
[ 0.07819139, 0.22727628, 0.69453233],
[ 0.07819139, 0.22727628, 0.69453233]])
Start with preprocessing!.
It's very important to standardize your data to zero-mean and unit-variance.
The scikit-learn docs say this:
Support Vector Machine algorithms are not scale invariant, so it is highly recommended to scale your data. For example, scale each attribute on the input vector X to [0,1] or [-1,+1], or standardize it to have mean 0 and variance 1. Note that the same scaling must be applied to the test vector to obtain meaningful results. See section Preprocessing data for more details on scaling and normalization
sklearns Section on Preprocessing
sklearns StandardScaler.
The next step after this is parameter-tuning (C, gamma and co.). This is usually done by GridSearch. But i usually expect people to try a simple LinearSVM first before trying the Kernel-SVM (less hyper-parameters, less computation-time, better generalization for non-optimal parameter-chosings).
I am trying to set up a cost-sensitive binary classification learning in TensorFlow, which would put different penalties on false positives and false negatives. Does anyone know how to create a loss function from a set of penalty weights $(w_1, w_2, w_3, w_4)$ for (true positive, false positive, false negative, true negative).
I went over the standard cost functions offered, but can't figure out how to combine them to get something similar to the above.
Following #Cauchyzhou's answer, if you have the logits, and the sparse labels as well as a cost_matrix whose shape is [L, L], where L is the number of unique labels, you can simply use the function below to calculate the loss
def sparse_cost_sensitive_loss (logits, labels, cost_matrix):
batch_cost_matrix = tf.nn.embedding_lookup(cost_matrix, labels)
eps = 1e-6
probability = tf.clip_by_value(tf.nn.softmax(logits), eps, 1-eps)
cost_values = tf.log(1-probability)*batch_cost_matrix
loss = tf.reduce_mean(-tf.reduce_sum(cost_values, axis=1))
return loss
I am not aware of anyone who has built a cost sensitive neural network classifier but Alejandro Correa Bahnsen has published academic papers for cost sensitive logistic regression and cost sensitive decision trees and a very well documented python cost sensitive classification library named CostCla. CostCla is pretty easy to use if you are familiar with scikit-learn.
You should be able to use the Bayes minimum risk model in the library to minimize the cost of your neural network since it fits a cost model to output prediction probabilities of any classifier.
Note that CostCla is intended to work with potentially different costs for each sample. You give it a cost matrix for your training and test samples. However, you can just make all the rows in the cost matrix the same if that applies to your problem.
Here are a couple of additional academic papers on the subject:
The Foundations of Cost-Sensitive Learning
Optimal ROC Curve for a Combination of Classifiers
cost_matrix:
[[0,1,100],
[1,0,1],
[1,20,0]]
label:
[1,2]
y*:
[[0,1,0],
[0,0,1]]
y(prediction):
[[0.2,0.3,0.5],
[0.1,0.2,0.7]]
label,cost_matrix-->cost_embedding:
[[1,0,1],
[1,20,0]]
It obvious 0.3 in [0.2,0.3,0.5] refers to right lable probility of [0,1,0], so it should not contibute to loss.
0.7 in [0.1,0.2,0.7] is the same. In other words, the pos with value 1 in y* not contibute to loss.
So I have (1-y*):
[[1,0,1],
[1,1,0]]
Then the entropy is target*log(predict) + (1-target) * log(1-predict),and value 0 in y*,should use (1-target)*log(1-predict), so I use (1-predict) said (1-y)
1-y:
[[0.8,*0.7*,0.5],
[0.9,0.8,*0.3*]]
(italic num is useless)
the custom loss is
[[1,0,1], [1,20,0]] * log([[0.8,0.7,0.5],[0.9,0.8,0.3]]) *
[[1,0,1],[1,1,0]]
and you can see the (1-y*) can be drop here
so the loss is -tf.reduce_mean(cost_embedding*log(1-y))
,to make it applicable , should be:
-tf.reduce_mean(cost_embedding*log(tf.clip((1-y),1e-10)))
the demo is below
import tensorflow as tf
import numpy as np
hidden_units = 50
num_class = 3
class Model():
def __init__(self,name_scope,is_custom):
self.name_scope = name_scope
self.is_custom = is_custom
self.input_x = tf.placeholder(tf.float32,[None,hidden_units])
self.input_y = tf.placeholder(tf.int32,[None])
self.instantiate_weights()
self.logits = self.inference()
self.predictions = tf.argmax(self.logits,axis=1)
self.losses,self.train_op = self.opitmizer()
def instantiate_weights(self):
with tf.variable_scope(self.name_scope + 'FC'):
self.W = tf.get_variable('W',[hidden_units,num_class])
self.b = tf.get_variable('b',[num_class])
self.cost_matrix = tf.constant(
np.array([[0,1,100],[1,0,100],[20,5,0]]),
dtype = tf.float32
)
def inference(self):
return tf.matmul(self.input_x,self.W) + self.b
def opitmizer(self):
if not self.is_custom:
loss = tf.nn.sparse_softmax_cross_entropy_with_logits\
(labels=self.input_y,logits=self.logits)
else:
batch_cost_matrix = tf.nn.embedding_lookup(
self.cost_matrix,self.input_y
)
loss = - tf.log(1 - tf.nn.softmax(self.logits))\
* batch_cost_matrix
train_op = tf.train.AdamOptimizer().minimize(loss)
return loss,train_op
import random
batch_size = 128
norm_model = Model('norm',False)
custom_model = Model('cost',True)
split_point = int(0.9 * dataset_size)
train_set = datasets[:split_point]
test_set = datasets[split_point:]
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(100):
batch_index = random.sample(range(split_point),batch_size)
train_batch = train_set[batch_index]
train_labels = lables[batch_index]
_,eval_predict,eval_loss = sess.run([norm_model.train_op,
norm_model.predictions,norm_model.losses],
feed_dict={
norm_model.input_x:train_batch,
norm_model.input_y:train_labels
})
_,eval_predict1,eval_loss1 = sess.run([custom_model.train_op,
custom_model.predictions,custom_model.losses],
feed_dict={
custom_model.input_x:train_batch,
custom_model.input_y:train_labels
})
# print '默认',eval_predict,'\n自定义',eval_predict1
print np.sum(((eval_predict == train_labels)==True).astype(np.int)),\
np.sum(((eval_predict1 == train_labels)==True).astype(np.int))
if i%10 == 0:
print '默认测试',sess.run(norm_model.predictions,
feed_dict={
norm_model.input_x:test_set,
norm_model.input_y:lables[split_point:]
})
print '自定义测试',sess.run(custom_model.predictions,
feed_dict={
custom_model.input_x:test_set,
custom_model.input_y:lables[split_point:]
})
Here is other solution where you can use any tensorflow loss and make it cost sensitive using kwarg weights ... note that unlike most cases here you need to use cost as '1' instead of '0' when you want to keep loss as it is ...
Some advantages of this approach are:
it extends tf.losses.Loss and satisfies the call api
reduction kwarg of the original loss remains functional and the behaviour is propagated to CostSensitiveLoss
you can also pass your own extra weights to new loss instances. Note that internally generated weights are used by wrapped self.loss
import numpy as np
from keras.api._v2 import keras as tk
import tensorflow as tf
from keras.utils import losses_utils
import typing as t
class CostSensitiveLoss(tk.losses.Loss):
def __init__(
self,
cost_matrix: t.List, loss: tk.losses.Loss,
):
super().__init__(reduction=loss.reduction, name=loss.name)
self.loss = loss
self.cost_matrix = cost_matrix
self._cost_matrix = tf.constant(cost_matrix, dtype=tf.float32)
#classmethod
def from_config(cls, config):
config['loss'] = tk.losses.deserialize(config['loss'])
return cls(**config)
def get_config(self):
return {
'cost_matrix': self.cost_matrix,
'loss': tk.losses.serialize(self.loss),
'reduction': self.reduction, 'name': self.name
}
def call(self, y_true, y_pred):
# if y_true is one hot encoded then get integer indices
if y_true.ndim == 1:
y_true_index = y_true
elif y_true.ndim == 2:
y_true_index = tf.argmax(y_true, axis=1)
else:
raise Exception(f"`y_true.ndim` {y_true.ndim} not supported")
# get cost for batch
cost_for_batch = tf.nn.embedding_lookup(self._cost_matrix, y_true_index)
cost_for_batch *= y_pred
cost_for_batch = tf.reduce_sum(cost_for_batch, axis=1)
# get loss
return self.loss(y_true, y_pred, cost_for_batch)
if __name__ == '__main__':
# for debug purpose I have kept 'none' you can
# safely use other options like 'sum', 'auto'
_loss = tk.losses.MeanAbsoluteError(reduction='none')
# some cost matrices the first cost matrix is the case when you are
# not using cost sensitive weights
_cs_loss_1 = CostSensitiveLoss(
cost_matrix=[[1, 1, 1], [1, 1, 1], [1, 1, 1], ],
loss=_loss
)
_cs_loss_2 = CostSensitiveLoss(
cost_matrix=[[1, 2, 2], [4, 1, 4], [8, 8, 1], ],
loss=_loss
)
_cs_loss_3 = CostSensitiveLoss(
cost_matrix=[[1, 4, 8], [2, 1, 8], [2, 4, 1], ],
loss=_loss
)
_y_true = np.asarray(
[
[1, 0, 0],
[0, 1, 0],
[0, 0, 1],
[1, 0, 0],
[0, 1, 0],
[0, 0, 1],
[1, 0, 0],
[0, 1, 0],
[0, 0, 1],
]
)
_y_pred = np.asarray(
[
[0.8, 0.1, 0.1],
[0.1, 0.8, 0.1],
[0.1, 0.1, 0.8],
[0.1, 0.8, 0.1],
[0.1, 0.1, 0.8],
[0.8, 0.1, 0.1],
[0.1, 0.1, 0.8],
[0.8, 0.1, 0.1],
[0.1, 0.8, 0.1],
]
)
print("loss ........................")
print(_loss(_y_true, _y_pred).numpy())
print("cs_loss_1 ...................")
print(_cs_loss_1(_y_true, _y_pred).numpy())
print("cs_loss_2 ...................")
print(_cs_loss_2(_y_true, _y_pred).numpy())
print("cs_loss_3 ...................")
print(_cs_loss_3(_y_true, _y_pred).numpy())