Could anyone suggest an elegant way to eliminate the redundant reruns in a pipeline. The example below shows that MyLogistic.fit() is called 37 times (36 for the CV + 1 on the whole data set). For instance, if we take the hyperparameters:
(0.2, 1, 0.01)
(0.2, 1, 0.5)
then it's not necessary to rerun the first step, MyLogistic, with C = 0.2, because its result will be the same (let's ignore the random seed issue). The same reasoning applies to any combination of the first k hyperparameters: once we got the result from the first k steps, we can recycle it.
Thanks to Ben's suggestion, I used the "memory" argument in Pipeline and the count of calls got reduced to 12 (13). Apparently that's because there are 4 combinations of select_k* parameters, times 3 folds. But, for C = 0.02 there is no need to call MyLogistic() twice for max_features 1 and 2. Is there a more or less elegant way to avoid
those redundant calls also?
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import (GridSearchCV, StratifiedKFold)
from sklearn.feature_selection import SelectFromModel
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
from sklearn.datasets import load_iris
import numpy as np
class MyLogistic(LogisticRegression):
__call_counter = 0
def fit(self, X, y, sample_weight=None):
print("MyLogistic fit is called.")
MyLogistic._MyLogistic__call_counter += 1
# fit() returns self.
return super().fit(X, y, sample_weight)
# If I use this "extension", everything works fine.
#class MyLogistic(LogisticRegression):
# pass
initial_logistic = MyLogistic(solver="liblinear", random_state = np.random.RandomState(18))
final_logistic = LogisticRegression(solver="liblinear", random_state = np.random.RandomState(20))
# prefit = False by default
select_best = SelectFromModel(estimator = initial_logistic, threshold = -np.inf)
select_k_best_pipeline = Pipeline(steps=[
('first_scaler', StandardScaler(with_mean = False)),
# initial_logistic will be called from select_best, prefit = false by default.
('select_k_best', select_best),
('final_logit', final_logistic)
])
select_best_grid = {'select_k_best__estimator__C' : [0.02, 0.03],
'select_k_best__max_features': [1, 2],
'final_logit__C' : [0.01, 0.5, 1.0]}
skf = StratifiedKFold(n_splits = 3, shuffle = True, random_state = 17)
# Use n_jobs = 1 to get the correct count of calls
logit_best_searcher = GridSearchCV(estimator = select_k_best_pipeline, param_grid = select_best_grid, cv = skf,
scoring = "roc_auc", n_jobs = 1, verbose = 4)
X, y = load_iris(return_X_y=True)
logit_best_searcher.fit(X, y > 0)
print("Best hyperparams: ", logit_best_searcher.best_params_)
print("The count of calls is: ", MyLogistic._MyLogistic__call_counter)
Related
When trying to train/evaluate a support vector machine in scikit-learn, I am experiencing some unexpected behaviour and I am wondering whether I am doing something wrong or that this is a possible bug.
In a very specific subset of circumstances, nested cross-validation using GridSearchCV and SVM, provides inflated predictive results, even with randomly generated data.
For instance, see this code:
from sklearn import svm
from sklearn.linear_model import LogisticRegression
import numpy as np
from sklearn.model_selection import GridSearchCV, StratifiedKFold, LeaveOneOut
from sklearn.metrics import roc_auc_score, brier_score_loss
from tqdm import tqdm
import pandas as pd
N = 20
N_FEATURES = 50
param_grid = {'C': [1e-5, 1e-3, 1, 1e3, 1e5]}
scores = []
for z in tqdm(range(100)):
X = np.random.uniform(size=(N, N_FEATURES))
y = np.random.binomial(1, 0.5, size=N)
if z < 10:
y = np.array([0, 1] * int(N/2))
y = np.random.permutation(y)
for skf_outer in [StratifiedKFold(n_splits=5), LeaveOneOut()]:
for skf_inner in [5, LeaveOneOut()]:
for model in [svm.SVC(probability=True), LogisticRegression()]:
y_pred, y_real = [], []
for train_index, test_index in skf_outer.split(X, y):
X_train, X_test = X[train_index], X[test_index, :]
y_train, y_test = y[train_index], y[test_index]
clf = GridSearchCV(
model, param_grid, cv=skf_inner, n_jobs=-1, scoring='neg_brier_score'
)
clf.fit(X_train, y_train)
predictions = clf.predict_proba(X_test)[:, 1]
y_pred.extend(predictions)
y_real.extend(y_test)
scores.append([str(skf_outer), str(skf_inner), str(model), np.mean(y), brier_score_loss(np.array(y_real), np.array(y_pred)), roc_auc_score(np.array(y_real), np.array(y_pred))])
df_scores = pd.DataFrame(scores)
df_scores.columns = ['skf_outer', 'skf_inner', 'model', 'y_label', 'brier', 'auc']
df_scores['y_0.5'] = df_scores['y_label'] == 0.5
df_scores = df_scores.groupby(['skf_outer', 'skf_inner', 'model', 'y_0.5']).mean()
print(df_scores)
In the following circumstances:
Both in the inner- and outerloop of the CV, LeaveOneOut() is used
The SVM is used
The y labels are balanced (i.e. the mean of y is 0.5)
The predictions are much better than expected by random chance (AUC>0.9, sometimes even 1, Brier of 0.15 or lower). I can replicate this generating more samples, more features etc - the issue stays the same. Swapping the SVM for LogisticRegression (as shown in the analysis above), leads to expected results (AUC 0.5, Brier of 0.25). And for the other scenario's (no LOO-CV in either inner or outer loop, or a different distribution of y labels), the results are as expected.
Can anyone replicate this? Am I missing something obvious?
I've replicated this with an older version of sklearn (0.24.0) and the newest one (1.2.0).
I have a set of training data that consists of X, which is a set of n columns of data (features), and Y, which is one column of target variable.
I am trying to train my model with logistic regression using the following pipeline:
pipeline = sklearn.pipeline.Pipeline([
('logistic_regression', LogisticRegression(penalty = 'none', C = 10))
])
My goal is to obtain the values of each of the n coefficients corresponding to the features, under the assumption of a linear model (y = coeff_0 + coeff_1*x1 + ... + coeff_n*xn).
What I tried was to train this pipeline on my data with model = pipeline.fit(X, Y). So I think that I now have the model that contains the coefficients that I want. However, I don't know how to access them. I'm looking for something like mode.best_params_('logistic_regression').
Does anyone know how to extract the fitted coefficients from a model like this?
Have a look at the scikit-learn documentation for Pipeline, this example is inspired by it:
from sklearn import svm
from sklearn.datasets import make_classification
from sklearn.feature_selection import SelectKBest
from sklearn.feature_selection import f_regression
from sklearn.pipeline import Pipeline
# generate some data to play with
X, y = make_classification(n_informative=5, n_redundant=0, random_state=42)
# ANOVA SVM-C
anova_filter = SelectKBest(f_regression, k=5)
clf = svm.SVC(kernel='linear')
anova_svm = Pipeline([('anova', anova_filter), ('svc', clf)])
anova_svm.set_params(anova__k=10, svc__C=.1).fit(X, y)
# access coefficients
print(anova_svm['svc'].coef_)
model.coef_ does the job, .best_params_ is usualy associated with GridSearch, i.e. hyperparameter optimization.
In your specific case try: model['logistic_regression'].coefs_.
Example to get the coefs from a pipeline.
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import load_iris
from sklearn.pipeline import Pipeline
X, y = load_iris(return_X_y=True)
pipeline = Pipeline([('lr', LogisticRegression(penalty = 'l2',
C = 10))])
pipeline.fit(X, y)
pipeline['lr'].coef_
array([[-0.42923513, 2.08235619, -4.28084811, -1.97174699],
[ 1.06321671, -0.08077595, -0.46911772, -2.3221883 ],
[-0.63398158, -2.00158024, 4.74996583, 4.29393529]])
here is how to visualize the coefficients and measure model accuracy. I used the baby weight and height and gestation period to predict preterm
pipeline = Pipeline([('lr', LogisticRegression(penalty='l2',C=10))])
scaler=StandardScaler()
#X=np.array(df['gestation_wks']).reshape(-1,1)
X=scaler.fit_transform(df[['bwt_lbs','height_ft','gestation_wks']])
y=np.array(df['PreTerm'])
X_train,X_test, y_train,y_test=train_test_split(X,y,test_size=0.3,random_state=123)
pipeline.fit(X_train,y_train)
y_pred_prob=pipeline.predict_proba(X_test)
predictions=pipeline.predict(X_test)
print(predictions)
sns.countplot(x=predictions, orient='h')
plt.show()
#print(predictions[:,0])
print(pipeline['lr'].coef_)
print(pipeline['lr'].intercept_)
print('Coefficients close to zero will contribute little to the end result')
num_err = np.sum(y != pipeline.predict(X))
print("Number of errors:", num_err)
def my_loss(y,w):
s = 0
for i in range(y.size):
# Get the true and predicted target values for example 'i'
y_i_true = y[i]
y_i_pred = w[i]
s = s + (y_i_true - y_i_pred)**2
return s
print("Loss:",my_loss(y_test,predictions))
fpr, tpr, threshholds = roc_curve(y_test,y_pred_prob[:,1])
plt.plot([0, 1], [0, 1], 'k--')
plt.plot(fpr, tpr)
plt.xlabel('False Positive Rate')
plt.ylabel('True Positive Rate')
plt.title('ROC Curve')
plt.show()
accuracy=round(pipeline['lr'].score(X_train, y_train) * 100, 2)
print("Model Accuracy={accuracy}".format(accuracy=accuracy))
cm=confusion_matrix(y_test,predictions)
print(cm)
Im new to Tensorflow&ML and following this example:
https://www.tensorflow.org/get_started/tflearn
It works very well until change hidden_units parameter here:
classifier = tf.contrib.learn.DNNClassifier(feature_columns=feature_columns,
hidden_units=[10, 20, 10],
n_classes=3,
model_dir="/tmp/iris_model")
When i try anything, for example hidden_units = [20, 40, 20] or hidden_units = [20] it throws an error.
I tried to find out on my own but unsuccessfully so far and thought someone here can help.
The question is how to chose a number of hidden layers for DNN Classifier and why two my examples above do not work?
Here is a full code:
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import urllib
import tensorflow as tf
import numpy as np
IRIS_TRAINING = "iris_training.csv"
IRIS_TRAINING_URL = "http://download.tensorflow.org/data/iris_training.csv"
IRIS_TEST = "iris_test.csv"
IRIS_TEST_URL = "http://download.tensorflow.org/data/iris_test.csv"
if not os.path.exists(IRIS_TRAINING):
raw = urllib.request.urlopen(IRIS_TRAINING_URL).read()
with open(IRIS_TRAINING,'wb') as f:
f.write(raw)
if not os.path.exists(IRIS_TEST):
raw = urllib.request.urlopen(IRIS_TEST_URL).read()
with open(IRIS_TEST,'wb') as f:
f.write(raw)
# Load datasets.
training_set = tf.contrib.learn.datasets.base.load_csv_with_header(
filename=IRIS_TRAINING,
target_dtype=np.int,
features_dtype=np.float32)
test_set = tf.contrib.learn.datasets.base.load_csv_with_header(
filename=IRIS_TEST,
target_dtype=np.int,
features_dtype=np.float32)
# Specify that all features have real-value data
feature_columns = [tf.contrib.layers.real_valued_column("", dimension=4)]
# Build 3 layer DNN with 10, 20, 10 units respectively.
classifier = tf.contrib.learn.DNNClassifier(feature_columns=feature_columns,
hidden_units=[10, 20, 10],
n_classes=3,
model_dir="/tmp/iris_model")
# Define the training inputs
def get_train_inputs():
x = tf.constant(training_set.data)
y = tf.constant(training_set.target)
return x, y
# Fit model.
classifier.fit(input_fn=get_train_inputs, steps=2000)
# Define the test inputs
def get_test_inputs():
x = tf.constant(test_set.data)
y = tf.constant(test_set.target)
return x, y
# Evaluate accuracy.
accuracy_score = classifier.evaluate(input_fn=get_test_inputs,
steps=1)["accuracy"]
print("\nTest Accuracy: {0:f}\n".format(accuracy_score))
Found it,
if model_dir is not specified than moel works just fine with new hidden_units
I am trying to use LSTM Recurrent Neural Net using Keras to forecast future purchase. My input variables are time-window of purchases for previous 5 days, and a categorical variable which I encoded as dummy variables A, B, ...,I. My input data looks like following:
>>> dataframe.head()
day price A B C D E F G H I TS_bigHolidays
0 2015-06-16 7.031160 1 0 0 0 0 0 0 0 0 0
1 2015-06-17 10.732429 1 0 0 0 0 0 0 0 0 0
2 2015-06-18 21.312692 1 0 0 0 0 0 0 0 0 0
My problem is my forecasts/fitted values (both for trained and test data) seem to be shifted forward. Here is a plot:
My question is what parameter in LSTM Keras should I change to correct this issue? Or do I need to change anything in my input data?
Here is my code:
import numpy as np
import os
import matplotlib.pyplot as plt
import pandas
import math
import time
import csv
from keras.models import Sequential
from keras.layers.core import Dense, Activation, Dropout
from keras.layers.recurrent import LSTM
from sklearn.preprocessing import MinMaxScaler
np.random.seed(1234)
exo_feature = ["A","B","C","D","E","F","G","H","I", "TS_bigHolidays"]
look_back = 5 #this is number of days we are looking back for sliding window of time series
forecast_period_length = 40
# load the dataset
dataframe = pandas.read_csv('processedDataframeGameSphere.csv', header = 0, engine='python', skipfooter=6)
dataframe["price"] = dataframe['price'].astype('float32')
scaler = MinMaxScaler(feature_range=(0, 100))
dataframe["price"] = scaler.fit_transform(dataframe['price'])
# this function is used to make sliding window for time series data
def create_dataframe(dataframe, look_back=1):
dataX, dataY = [], []
for i in range(dataframe.shape[0]-look_back-1):
price_lookback = dataframe['price'][i: (i + look_back)] #i+look_back is exclusive here
exog_feature = dataframe[exo_feature].ix[i + look_back - 1] #Y is i+ look_back ,that's why
row_i = price_lookback.append(exog_feature)
dataX.append(row_i)
dataY.append(dataframe["price"][i + look_back])
return np.array(dataX), np.array(dataY)
window_dataframe, Y = create_dataframe(dataframe, look_back)
# split into train and test sets
train_size = int(dataframe.shape[0] - forecast_period_length) #28 is the number of days we want to forecast , 4 weeks
test_size = dataframe.shape[0] - train_size
test_size_start_point_with_lookback = train_size - look_back
trainX, trainY = window_dataframe[0:train_size,:], Y[0:train_size]
print(trainX.shape)
print(trainY.shape)
#below changed datawindowY indexing, since it's just array.
testX, testY = window_dataframe[train_size:dataframe.shape[0],:], Y[train_size:dataframe.shape[0]]
# reshape input to be [samples, time steps, features]
trainX = np.reshape(trainX, (trainX.shape[0], 1, trainX.shape[1]))
testX = np.reshape(testX, (testX.shape[0], 1, testX.shape[1]))
print(trainX.shape)
print(testX.shape)
# create and fit the LSTM network
dimension_input = testX.shape[2]
model = Sequential()
layers = [dimension_input, 50, 100, 1]
epochs = 100
model.add(LSTM(
input_dim=layers[0],
output_dim=layers[1],
return_sequences=True))
model.add(Dropout(0.2))
model.add(LSTM(
layers[2],
return_sequences=False))
model.add(Dropout(0.2))
model.add(Dense(
output_dim=layers[3]))
model.add(Activation("linear"))
start = time.time()
model.compile(loss="mse", optimizer="rmsprop")
print "Compilation Time : ", time.time() - start
model.fit(
trainX, trainY,
batch_size= 10, nb_epoch=epochs, validation_split=0.05,verbose =2)
# Estimate model performance
trainScore = model.evaluate(trainX, trainY, verbose=0)
trainScore = math.sqrt(trainScore)
trainScore = scaler.inverse_transform(np.array([[trainScore]]))
print('Train Score: %.2f RMSE' % (trainScore))
testScore = model.evaluate(testX, testY, verbose=0)
testScore = math.sqrt(testScore)
testScore = scaler.inverse_transform(np.array([[testScore]]))
print('Test Score: %.2f RMSE' % (testScore))
# generate predictions for training
trainPredict = model.predict(trainX)
testPredict = model.predict(testX)
# shift train predictions for plotting
np_price = np.array(dataframe["price"])
print(np_price.shape)
np_price = np_price.reshape(np_price.shape[0],1)
trainPredictPlot = np.empty_like(np_price)
trainPredictPlot[:, :] = np.nan
trainPredictPlot[look_back:len(trainPredict)+look_back, :] = trainPredict
testPredictPlot = np.empty_like(np_price)
testPredictPlot[:, :] = np.nan
testPredictPlot[len(trainPredict)+look_back+1:dataframe.shape[0], :] = testPredict
# plot baseline and predictions
plt.plot(dataframe["price"])
plt.plot(trainPredictPlot)
plt.plot(testPredictPlot)
plt.show()
It's not a problem of LSTM, if you use just simple feed-forward network, the effect will be the same.
the problem is the network tend to mimic yesterday value instead of 'forecasting' you expect.
(it is nice strategy in term of reducing MSE loss)
you need more 'care' to avoid this issue and it's not a simple issue.
I'm having a difficult time visualizing what this Tensorflow class creates. I want to implement a LSTM RNN that handles 3D data.
class Grid3LSTMCell(GridRNNCell):
"""3D BasicLSTM cell
This creates a 2D cell which receives input and gives output in the first dimension.
The first dimension can optionally be non-recurrent if `non_recurrent_fn` is specified.
The second and third dimensions are LSTM.
"""
def __init__(self, num_units, tied=False, non_recurrent_fn=None,
use_peepholes=False, forget_bias=1.0):
super(Grid3LSTMCell, self).__init__(num_units=num_units, num_dims=3,
input_dims=0, output_dims=0, priority_dims=0, tied=tied,
non_recurrent_dims=None if non_recurrent_fn is None else 0,
cell_fn=lambda n, i: rnn_cell.LSTMCell(
num_units=n, input_size=i, forget_bias=forget_bias,
use_peepholes=use_peepholes),
non_recurrent_fn=non_recurrent_fn)
The class is found in `from tensorflow.contrib.grid_rnn.python.ops import grid_rnn_cell`.
This is difficult to explain, so I've provided a drawing. Here is what I want it to do...
However the comment sounds like it isn't doing this. The comment makes it sound like the RNN is still a flat RNN, where the first dimension is outputting to, what is commonly called, the outputs variable (see below). The second dimension is outputting to the next step in the RNN, and the third dimension is outputting to the next hidden layer.
outputs, states = rnn.rnn(lstm_cell, x, dtype=tf.float32)
If this is the case, what is the point in having the first and second dimensions? Aren't they essentially the same thing? The BasicLSTMCell sends the output to the next step into outputs -- in other words they are one in the same.
Clarity?
For reference, here is my example code...
import tensorflow as tf
from tensorflow.python.ops import rnn, rnn_cell
from tensorflow.contrib.grid_rnn.python.ops import grid_rnn_cell
import numpy as np
#define parameters
learning_rate = 0.01
batch_size = 2
n_input_x = 10
n_input_y = 10
n_input_z = 10
n_hidden = 128
n_classes = 2
n_output = n_input_x * n_classes
x = tf.placeholder("float", [n_input_x, n_input_y, n_input_z])
y = tf.placeholder("float", [n_input_x, n_input_y, n_input_z, n_classes])
weights = {}
biases = {}
for i in xrange(n_input_y * n_input_z):
weights[i] = tf.Variable(tf.random_normal([n_hidden, n_output]))
biases[i] = tf.Variable(tf.random_normal([n_output]))
#generate random data
input_data = np.random.rand(n_input_x, n_input_y, n_input_z)
ground_truth = np.random.rand(n_input_x, n_input_y, n_input_z, n_classes)
#build GridLSTM
def GridLSTM_network(x):
x = tf.reshape(x, [-1,n_input_x])
x = tf.split(0, n_input_y * n_input_z, x)
lstm_cell = grid_rnn_cell.Grid3LSTMCell(n_hidden)
outputs, states = rnn.rnn(lstm_cell, x, dtype=tf.float32)
output = []
for i in xrange(n_input_y * n_input_z):
output.append(tf.matmul(outputs[i], weights[i]) + biases[i])
return output
#initialize network, cost, optimizer and all variables
pred = GridLSTM_network(x)
# import pdb
# pdb.set_trace()
pred = tf.pack(pred)
pred = tf.transpose(pred,[1,0,2])
pred= tf.reshape(pred, [-1, n_input_x, n_input_y, n_input_z, n_classes])
temp_pred = tf.reshape(pred, [-1,n_classes])
temp_y = tf.reshape(y,[-1, n_classes])
cost = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(temp_pred, temp_y))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)
# Evaluate model
correct_pred = tf.equal(0,tf.cast(tf.sub(tf.nn.sigmoid(temp_pred),temp_y), tf.int32))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
# Initializing the variables
init = tf.initialize_all_variables()
# Launch the graph
with tf.Session() as sess:
sess.run(init)
step = 0
while 1:
print step
step = step + 1
# pdb.set_trace
sess.run(optimizer, feed_dict={x: input_data, y: ground_truth})