Based on: https://github.com/seg/2016-ml-contest. Co-ordinated by Matt Hall wFollowing Brendon Hall's tutorial in the leading edge.
A question to consider at the start of a data science exercise is "can this problem be solved by an experienced human?". From a geoscience background, the initial hunch is no. This is also reflected in the moderate scores, no team has over 65% yet.
We want to define a function y = h(X) Where y is originaly based on y = h(U) Where: h(U) is a geological interpretation of core data and will not be reproducible in this experiment. It is not assured that y can be predicted based only on these well logs.
This evaluation investigated this idea and consider what could be done to allow for a higher score. It followed:
Still at this time no signifcant feature engineering has been undertaken. Drafts have been made for a machine learning prediction for PE, and other categories and for feature selection based on random forest. However given the high number of quickly changing results after prediction the result looks like it needs to be smoothed. The expectation is this will improve with time-series features like lag and averaging over a window.
The results suggest that there is significant overlap in the wireline well log responses for a number of facies. That even with feature engineering it can be difficult to differentiate. It is beyond the scope of this submission to discuss the geological and petrophysical reasons for this. One key example is non-marine coarse siltstone vs. non-marine fine siltstone, this produces the most errors of all models tested. Even in core these two can be challenging to interpret. Perhaps gouping into one non-marine siltstone would be more suitable for a modelling exercise.
The learning curves and other QC plots do not seem to suggest that obtaining more data will neccisarily make this better as they all flatten out early. The main problem here is recall. Either the question should be reconfigured to ask for rock property groups that can be defined by a petrophysicist or a greater variety of well logs should be used. E.g. spectral gamma ray.
Once starting to fit models facies 9 (Bafflestones) tended to dominate the responses and overprint other carbonate facies that appear to be the right choice when submitting testing. By deleting F9 entirely this could be proven in the test scores. At this stage OneVsOne, OneVsRest and a standard model were fit all using a boosted tree model. When including F9 they would all demonstrate this issue and result in a large number of F9 facies and lower scores.
"The splitting of data into training and validation sets “must” be done according to labels. In case of any kind of classification problem, use stratified splitting. In python, you can do this using scikit-learn very easily." Abhishek Thakur.
Yet after exhaustive testing it was only using Leave Groups Out and leaving 2 wells out at a time that the models started to perform well or with K-Fold without stratification. When looking at the plot showing the distribution of facies it can be seen that the distribution is mixed across wells so perhaps this method is succesfully presenting only models that can be fitted to very different outputs of wells and therefore perform better against the test data.
The CRAWFORD well also has an interpolation gap (3025-3040 in the well logs where they go diagonally from one value to the next. This will not be easy to predict as the values should not correspond. CHURCHMAN BIBLE also has a gap in the well logs that has been interpolated. However the prediction does a decent job at having an attempt, it will have low accuracy but it is impressive to see that it can make a prediction in an interval which appears to have false log values. This may be more of a reflection of overfitting than something that will succesfully generalize.
A score around 90-95% may be needed to give confidence to be implemented in a producing field. The conclusions for this evaluation is that the experimental design needs to changed to be able to achieve this. This competition and the original paper has been a fantastic way to start a discussion within the geoscience community around machine learning, the lower scores appear to be a limitation of this specific challenge rather than the machine learning methodology.
In [3]:
# Initializing
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
import sys #only needed to determine Python version number
import matplotlib #only needed to determine Matplotlib version number
import seaborn as sns
#Enable inline plotting
%matplotlib inline
print ("This has been run using:")
print('Python version ' + sys.version)
print('Pandas version ' + pd.__version__)
print('Matplotlib version ' + matplotlib.__version__)
## Set colour template
sns.set(style="ticks") ## Set style to white background and ticks
sns.set_context("paper") ## Set size of labels
# Load file
try:
print("Loading dataset...")
CSV_Dataset = r"facies_vectors.csv"
Dataset = pd.read_csv(CSV_Dataset)
except:
print ("An error has occured")
print ("Please ensure that you have downloaded the dataset and")
print ("entered the file location and file path for your machine")
finally:
print ("Expected input is the file facies_vectors.csv")
Dataset.head()
## Remove Recruit F9
Dataset = Dataset[Dataset["Well Name"] != "Recruit F9"]
Dataset = Dataset[Dataset["Well Name"] != "CROSS H CATTLE"]
Dataset = Dataset[Dataset["Facies"] != 9]
In [144]:
#### Make a new array that consists of the facies organised in columns for each well
Facies_Plot = Dataset.iloc[:,0]
Well = Dataset.iloc[:,2]
Facies_Plot = pd.concat([Facies_Plot, Well], axis=1)
grouped = Facies_Plot.groupby('Well Name')
Facies_Plot = Facies_Plot.rename(columns = {"Well Name":"Well_Name"}) # To be able to use the header in the .unique method
List = Facies_Plot.Well_Name.unique()
u = np.arange(501) #make a dummy key based on the legth of the longest well - Should be automated.
b = pd.DataFrame()
b["key"]=u
for i in List:
a = grouped.get_group(i)
a = a.rename(columns = {"Facies":i})
a = a.drop("Well Name", 1)
[c, d] = a.shape
e = np.arange(c)
a["key"]=e
#b = pd.concat([a, b], axis=1)
b = pd.merge(a, b, on='key')
Facies_Plot = b
Facies_Plot = Facies_Plot.drop("key", 1)
cmap_facies = ['#F4D03F', '#F5B041','#DC7633','#6E2C00', '#1B4F72','#2E86C1', '#AED6F1', '#A569BD', '#196F3D']
from matplotlib.colors import ListedColormap
cmap_list = ListedColormap(cmap_facies)
## Plot all facies to give a visual impression of the distribution of facies
f, ax = plt.subplots(figsize=(11, 9))
ax = sns.heatmap(Facies_Plot, yticklabels=False, cmap=cmap_list, linewidths=0, vmin=1, vmax=9, ax=ax, cbar_kws={"shrink": .5})
In [175]:
### Create new categories ####
ytemp= Dataset.iloc[:,0]
def new_categories(y):
def create_category(facies_names, y):
a = 1
for i in facies_names:
y = y.replace(to_replace=a,value=i)
a+=1
return y
##Relable numerics with String label
ytemp_1=y
facies_names = ['SS', 'CSiS', 'FSiS', 'SiSh', 'MS', 'WS', 'D', 'PS', 'BS']
ytemp_1=create_category(facies_names, ytemp_1)
## Make a new column of depositional environments
ytemp_2=ytemp
ytemp_2 = ytemp_2.rename(columns = {"Facies":"Dep Environment"})
facies_names = ['Clastic Non-marine', 'Clastic Non-marine', 'Clastic Non-marine', 'Clastic Non-marine', 'Clastic Marine', 'Carb Platform', 'Carb Platform', 'Carb Platform', 'Carb Platform']
ytemp_2=create_category(facies_names, ytemp_2)
## Make a new column of clastic vs. carbonate
ytemp_3=ytemp
ytemp_3.rename(columns={"Facies":"Lithology"}, inplace=True)
facies_names = ['Clastic', 'Clastic', 'Clastic', 'Clastic', 'Clastic', 'Carb', 'Carb', 'Carb', 'Carb']
ytemp_3=create_category(facies_names, ytemp_3)
## Make a new column of non-marine vs. marine
ytemp_4=ytemp
ytemp_4.rename(columns={"Facies":"Marine vs Non-marine"}, inplace=True)
facies_names = ['Non-marine', 'Non-marine', 'Non-marine', 'Marine', 'Marine', 'Marine', 'Marine', 'Marine', 'Marine']
ytemp_4=create_category(facies_names, ytemp_4)
## Merge the results into a new table
y = pd.concat([ytemp_1, ytemp_2, ytemp_3, ytemp_4], axis=1)
y.rename(columns={0:"Dep Environment"}, inplace=True)
y.rename(columns={1:"Lithology"}, inplace=True)
y.rename(columns={2:"M vs NM"}, inplace=True)
new_classes = y
return new_classes
new_classes = new_categories(ytemp)
X, y, y_one_column = prepare_X_Y(Dataset)
## Merge together new class labels, y results as one hot vecotr and X
df_full = pd.concat([new_classes, y, X], axis=1)
df = pd.concat([new_classes, X], axis=1)
In [47]:
### Set some colour parameters.
cmap_facies = ['#F4D03F', '#F5B041','#DC7633','#6E2C00', '#1B4F72','#2E86C1', '#AED6F1', '#A569BD', '#196F3D']
cmap_m_nm = ["sage", "royalblue"]
cmap_clas_carb = ["gold", "slategrey"]
##### Principal Component Analysis #####
# Drop well name and binary features
drop = ["Well Name", "NM_M", "Depth", "Formation"]
X_temp = X
X_temp = X.drop(drop, 1)
# Data should be preprocessed using mean normalisation before input to principal component analysis.
# After mean normalization each parameter should have zero mean. (mean=0 and variance=1).
from sklearn.preprocessing import normalize, StandardScaler, RobustScaler
#X_Scaler = StandardScaler()
X_Scaler = RobustScaler()
X_Scaled = X_Scaler.fit_transform(X_temp)
# Project onto the linear subspace spanned by k number of vectors.
K = 2 # K is called number of n_components in Sci Kit learn.
from sklearn.decomposition import PCA, IncrementalPCA
ipca = IncrementalPCA(n_components=K, batch_size=10)
X_ipca_K2 = ipca.fit_transform(X_Scaled)
# Choose K by looking to retain 99% (0.01) of variance. K should be the smallest value that will give a 99% variance.
# There can be some variations for example 95% (0.05).
cum_var_exp = np.cumsum(ipca.explained_variance_ratio_)
with plt.style.context('seaborn-whitegrid'):
plt.figure(figsize=(6, 4))
plt.bar(range((ipca.n_components_)), ipca.explained_variance_ratio_, alpha=0.5, align='center',
label='individual explained variance')
plt.step(range(ipca.n_components_), cum_var_exp, where='mid',
label='cumulative explained variance')
plt.ylabel('Explained variance ratio')
plt.xlabel('Principal components')
plt.legend(loc='best')
plt.tight_layout()
####### Plots ########
## Make a data frame using the 2 dimensional result from principal component analysis
df_ipca = pd.DataFrame(X_ipca_K2, columns=["X_pca_1", "X_pca_2"])
df_ipca = df = pd.concat([new_classes, df_ipca], axis=1)
with sns.axes_style("white"):
sns.jointplot(x="X_pca_1", y="X_pca_2", data=df_ipca, kind="hex", color="k");
sns.set(style="ticks")
facies_colors = ['#F5B041','#DC7633','#6E2C00', '#1B4F72','#2E86C1', '#AED6F1', '#A569BD', '#196F3D', '#F4D03F']
cmap=facies_colors
pal_facies=sns.color_palette(cmap)
def pca_plot(List):
for i in List:
Lithology_pca = sns.FacetGrid(df_ipca, col=i, palette=pal_facies, hue="Facies")
Lithology_pca.map(plt.scatter, "X_pca_1", "X_pca_2", alpha=.3)
Lithology_pca.add_legend();
### Plot facies split against different categories ###
List = ["Lithology", "Dep Environment", "M vs NM"]
pca_plot(List)
### Plot all facies ###
Facies = sns.FacetGrid(df_ipca, col="Facies", palette=pal_facies)
Facies.map(plt.scatter, "X_pca_1", "X_pca_2", alpha=.8)
Facies.add_legend();
In [1]:
##############################################################
###### Define a set of functions to run the evaluation #######
##############################################################
def Start_up():
print ("This has been run using:")
print('Python version ' + sys.version)
print('Pandas version ' + pd.__version__)
print('Matplotlib version ' + matplotlib.__version__)
## Set colour template
sns.set(style="ticks") ## Set style to white background and ticks
sns.set_context("paper") ## Set size of labels
try:
print("Loading dataset...")
CSV_Dataset = r"facies_vectors.csv"
Dataset = pd.read_csv(CSV_Dataset)
except:
print ("An error has occured")
print ("Please ensure that you have downloaded the dataset and")
print ("entered the file location and file path for your machine")
finally:
print ("Expected input is the file facies_vectors.csv")
return Dataset
#######################################################
################### Prepare X and y ###################
#######################################################
def prepare_X_Y(Dataset):
## Remove Recruit F9
##########################################################
####### Impute missing values with mean ##################
##########################################################
## Replace all missing values with the mean values
Dataset = Dataset.fillna(Dataset.mean())
##########################################################
####### Remove wells or facies deemed distracting ########
#########################################################
well_list_remove = ["Recruit F9", "CROSS H CATTLE", "CHURCHMAN BIBLE"]
for i in well_list_remove:
Dataset = Dataset[Dataset["Well Name"] != i]
Dataset = Dataset[Dataset["Facies"] != 9]
###########################################################
########################## Make y #########################
###########################################################
ytemp = Dataset.iloc[:,0] #Note 0 index is used in python for the first position.
#print (("m={0}").format(ytemp.shape))
## Keep the original version where all classifiers are stored in one columnn
y_one_column = ytemp
####################################################################################
########## Make one hot vector version of y in case using OneVsRest ################
####################################################################################
## One hot vector for each valye of y. In case this is to be used as input for OneVsRest
## Get all the elements of y
ySet = set(ytemp)
Yn = len(ySet)
#print (("K={0}").format(Yn))
# Each classifier should have a sperate column and be measured only in ones and zeros
one_hot_y = ytemp
y = pd.get_dummies(one_hot_y)
y = y.rename(columns={1: "NM Coarse Sandstone", 2: "NM Coarse Siltstone", 3: "NM Fine Siltstone", 4:"Marine Siltstone", 5:"Mud Stone", 6:"Wacke Stone", 7:"Dolomite", 8:"Packe Stone", 9:"Baffle Stone"})
[Dm , Dn] = Dataset.shape
X = Dataset.iloc[:,1:Dn] #where Dn is the number of columns in the original dataset
Dataset_out = Dataset
## X, y (One hot vector version), y (flat including intergers from 1-9).
####################################################################################
############### Make new categories based on grouping facies #######################
####################################################################################
def new_categories(y):
def create_category(facies_names, y):
a = 1
for i in facies_names:
y = y.replace(to_replace=a,value=i)
a+=1
return y
##Relable numerics with String label
ytemp_1=y
facies_names = ['SS', 'CSiS', 'FSiS', 'SiSh', 'MS', 'WS', 'D', 'PS', 'BS']
ytemp_1=create_category(facies_names, ytemp_1)
## Make a new column of depositional environments
ytemp_2=ytemp
ytemp_2 = ytemp_2.rename(columns = {"Facies":"Dep Environment"})
facies_names = ['Clastic Non-marine', 'Clastic Non-marine', 'Clastic Non-marine', 'Clastic Non-marine', 'Clastic Marine', 'Carb Platform', 'Carb Platform', 'Carb Platform', 'Carb Platform']
ytemp_2=create_category(facies_names, ytemp_2)
## Make a new column of clastic vs. carbonate
ytemp_3=ytemp
ytemp_3.rename(columns={"Facies":"Lithology"}, inplace=True)
facies_names = ['Clastic', 'Clastic', 'Clastic', 'Clastic', 'Clastic', 'Carb', 'Carb', 'Carb', 'Carb']
ytemp_3=create_category(facies_names, ytemp_3)
## Make a new column of non-marine vs. marine
ytemp_4=ytemp
ytemp_4.rename(columns={"Facies":"Marine vs Non-marine"}, inplace=True)
facies_names = ['Non-marine', 'Non-marine', 'Non-marine', 'Marine', 'Marine', 'Marine', 'Marine', 'Marine', 'Marine']
ytemp_4=create_category(facies_names, ytemp_4)
## Merge the results into a new table
y = pd.concat([ytemp_1, ytemp_2, ytemp_3, ytemp_4], axis=1)
y.rename(columns={0:"Dep Environment"}, inplace=True)
y.rename(columns={1:"Lithology"}, inplace=True)
y.rename(columns={2:"M vs NM"}, inplace=True)
new_classes = y
return new_classes
new_classes = new_categories(y_one_column)
return X, y, y_one_column, new_classes
############# ###################
#####################################################################
####################### Feature Engineering #########################
#####################################################################
def prepare_X(X):
## Merge together new class labels, y results as one hot vecotr and X
#df_full = pd.concat([y, X], axis=1)
#df = pd.concat([X], axis=1)
####################################################################
################# Time Series Features #############################
####################################################################
print("start", X.shape)
Well_List = X['Well Name'].unique()
# Function to calculate mean
def rolling_mean(trX, window, Log_List):
df_time_series = trX
## List the logs that should go through time-series feature engineering
time_series_method = "rolling_mean"
win = window
Append_to = []
for i in Well_List:
df_well = df_time_series[df_time_series['Well Name']==i].copy(deep=True)
df_well.sort_values('Depth', inplace=True)
for l in Log_List:
column_name = (time_series_method+"_"+l+"_"+str(win))
df_well[column_name]= pd.rolling_mean(df_well[l], window = win)
### Infill missing values at the top and base
### pad / ffill ---- Fill values forward ---- bfill / backfill
df_well[column_name].fillna(method="ffill", inplace=True)
df_well[column_name].fillna(method="bfill", inplace=True)
## Append back to list
Append_to.append(df_well)
## Concate to dataframe
df_time_series= pd.concat(Append_to)
return df_time_series
# Function to call rolling mean with different windows
def run_rolling_mean_series(trX, mean_values, Log_List):
mean_df = trX
for mean in mean_values:
mean_temp = rolling_mean(mean_df, mean, Log_List)
mean_df=pd.merge(mean_df, mean_temp)
return mean_df
# Function to make lag features (shift)
def lag_series(trX, lag_value, Log_List):
df_lag_series = trX
time_series_method = "lag"
value = lag_value
Append_to = []
for i in Well_List:
df_well = df_time_series[df_time_series['Well Name']==i].copy(deep=True)
df_well.sort_values('Depth', inplace=True)
for l in Log_List:
column_name = (time_series_method+"_"+l+"_"+str(value))
df_well[column_name]= df_well[l].shift(value)
### Infill missing values at the top and base
### pad / ffill ---- Fill values forward ---- bfill / backfill
df_well[column_name].fillna(method="ffill", inplace=True)
df_well[column_name].fillna(method="bfill", inplace=True)
Append_to.append(df_well)
## Concate to dataframe
df_lag_series= pd.concat(Append_to)
## return
return df_lag_series
# Function to call lag series with different amounts of shift
def run_lag_series(trX, lag_values, Log_List):
lag_df = trX
Append_to = []
for lag in lag_values:
lag_temp = lag_series(lag_df, lag, log_list)
lag_df=pd.merge(lag_df, lag_temp)
return lag_df
### Define which logs should go through feature engineering.
log_list = ["GR", "PE","ILD_log10", "DeltaPHI", "PHIND"]
################ Run mean #################
### Set values for mean window to ###
Mean_Values = [3]
### ###
#X = run_rolling_mean_series(X, Mean_Values, log_list)
print ("after mean", X.shape)
############ Run lag ######################
### Set values to apply lag to ###
Lag_Values = [1, -1, 2, -2, 3, -3]
### ###
# X = run_lag_series(X, Lag_Values, log_list)
# print("after lag", X.shape)
## Feature engineering. All features have shown importance in PCA
## Take advantage of this by creating new features based on the relationships of key featuers.
## Make relationships between key features
###################################################################
###### Features based on the relationships with each other ########
###################################################################
X["DeltaPHI_PHIND"]=(X["DeltaPHI"]*X["PHIND"])
X["GR_PE"]=(X["GR"]*X["PE"])
X["PHIND_PE"]=(X["PHIND"]*X["PE"])
X["DeltaPHI_PE"]=(X["DeltaPHI"]*X["PE"])
X["ILD_log10_PE"]=(X["ILD_log10"]*X["PE"])
X["DeltaPHI_PHIND2"]=(X["DeltaPHI"]/X["PHIND"])
X["GR_ILD_log10"]=(X["GR"]*X["ILD_log10"])
X["GR_PHIND"]=(X["GR"]*X["PHIND"])
X["GR_PHIND2"]=(X["GR"]/X["PHIND"])
X["GR_DeltaPHI"]=(X["GR"]*X["DeltaPHI"])
X["ILD_log10_PHIND"]=(X["ILD_log10"]*X["PHIND"])
X["ILD_log10_DeltaPHI"]=(X["ILD_log10"]*X["DeltaPHI"])
X["PHIND_PE2"]=(X["PHIND"]/X["PE"])
X["DeltaPHI_PE2"]=(X["DeltaPHI"]/X["PE"])
#### N_MN is a powerfull feature, however it could also be misleading ####
#### It is an interpreted feature not a raw log ####
#### Keep seperate in case it needs to be blocked out ####
X["M_PE"]=(X["NM_M"]*X["PE"])
X["M_PE2"]=(X["NM_M"]/X["PE"])
X["M_PHIND"]=(X["NM_M"]*X["PHIND"])
X["M_PHIND2"]=(X["NM_M"]/X["PHIND"])
X["M_GR"]=(X["NM_M"]*X["GR"])
X["M_GR2"]=(X["NM_M"]/X["GR"])
X["M_GR_PHIND"]=(X["NM_M"]*X["GR_PHIND"])
#### ####
####################################################################
###### Make new features out of linear trends of key features ######
####################################################################
def make_linear_features(X):
def Linear_regression_features(a,b):
from sklearn.linear_model import LinearRegression
LinReg = LinearRegression()
LinReg.fit(a,b)
f_lin = LinReg.predict(a)
return f_lin
In1 = [X["PE"], X["PE"], X["PHIND"]]
In2 = [X["PHIND"], X["ILD_log10"], X["ILD_log10"]]
m = 0
for i in In1:
a = i
b = In2[m]
[l,]=a.shape
a=a.values.reshape((l,1))
b=b.values.reshape((l,1))
f_lin = Linear_regression_features(a,b)
X["f_lin"+str(m)]=f_lin
m+=1
return X
X = make_linear_features(X)
####################################################################
########### Convert Formations into Features #######################
####################################################################
def Formation_Features(df_X):
## Get formations
X_Fm = df_X
a = X_Fm["Formation"]
## Convert raw text into tf–idf (term frequency-inverse document frequency)
from sklearn.feature_extraction.text import TfidfVectorizer
tfv = TfidfVectorizer()
Formation_Vector = tfv.fit_transform(a)
## Convert to a DataFrame - Convert Sparse matrix into a dense matrix for pandas to read
Formation_Vector = pd.DataFrame(list(Formation_Vector.toarray()))
New_features_List = list(Formation_Vector.columns.values)
for i in New_features_List:
X_Fm["tf_idf_"+str(i)]=Formation_Vector[i]
X_Fm = X_Fm.fillna(0)
#X_Fm = X_Fm.drop(["FV"], 1)
return X_Fm
X = Formation_Features(X)
#### Return X back with new features ####
return X
############################################################################################
########################## Make PE from machine learning ##################################
############################################################################################
def train_ml_pe(Xpe):
from sklearn.ensemble import RandomForestRegressor
X_pe = Xpe
## Remove wells without PE logs
X_pe_train = X_pe
X_pe_train = X_pe_train[X_pe["Well Name"] != "ALEXANDER D"]
X_pe_train = X_pe_train[X_pe_train["Well Name"] != "KIMZEY A"]
## Drop strings
X_pe_train = drop_features(X_pe_train)
# Seprate out a y
y_pe = X_pe_train["PE"]
# Drop PE
X_pe_train = X_pe_train.drop("PE", 1)
# Cross validation
cv = inner_loop_cv("KFold", 4)
# Set inputs for GridSearchCV
from sklearn.metrics import mean_squared_error
scores = ["neg_mean_squared_error"]
parameters = {}
estimator = RandomForestRegressor(n_estimators=120)
best_result_list, Full_dict, Best_Estimator, clf_pe = optimise_parameters(X_pe_train, y_pe, estimator, scores, cv, parameters)
return clf_pe
def predict_pe(clf_pe, Xpe):
X_pe = Xpe
X_pe = drop_features(X_pe)
X_pe = X_pe.drop("PE", 1)
# Predict PE
p_y_pe = clf_pe.predict(X_pe)
[a,]=(p_y_pe.shape)
p_y_pe = p_y_pe.reshape(a, 1)
X_pe["PE_pred"]=p_y_pe
X_pe["PE"]=Xpe["PE"]
X_pe["Well Name"]=Xpe["Well Name"]
X_pe["Formation"]=Xpe["Formation"]
X_pe = X_pe.fillna(X_pe["PE_pred"])
X_pe = X_pe.drop("PE_pred", 1)
return X_pe
#clf_pe = train_ml_pe(trX)
#test = predict_pe(clf_pe, trX)
############################################################################################
############### Use Machine Learning to make extra features ##############################
############################################################################################
######################################################################
############ Predict features using machine learning ################
######################################################################
######################################################################
############# Drop features ##########################################
######################################################################
def drop_features(X):
## List which features should be dropped from the training data
Dropped_Features = ["Formation", "Well Name"]
X = X.drop(Dropped_Features, 1)
return X
def drop_features2(X, y):
## List which features should be dropped from the training data
Dropped_Features = ["Formation", "Well Name"]
X = X.drop(Dropped_Features, 1)
X["y"]=y
print ("before", X.shape)
# X = X.fillna(10)
y = X["y"]
X = X.drop("y", 1)
print ("after", X.shape)
return X, y
def drop_features_test(X):
## List which features should be dropped from the training data
Dropped_Features = ["Formation", "Well Name"]
X = X.drop(Dropped_Features, 1)
print ("before", X.shape)
# X = X.fillna(10)
return X
#############################################################################
########## Functions to run grid search cv and calibrate model ############
#############################################################################
def X_y_inputs(X, y):
X = X
y = y
return X, y
##################################################################
###################### Scaling stratergy #########################
##################################################################
def select_scaler(Scaler_name):
if Scaler_name == "Robust":
from sklearn.preprocessing import RobustScaler
Scaler = RobustScaler()
elif Scaler_name == "Standard":
from sklearn.preprocessing import StandardScaler
Scaler = StandardScaler(with_mean=False)
return Scaler
##################################################################
################ Cross Validation Stratergy ######################
##################################################################
def inner_loop_cv(cv_name, n):
if cv_name == "KFold":
from sklearn.model_selection import KFold
cv = KFold(n_splits=n, shuffle=False)
elif cv_name == "StratifiedKFold":
from sklearn.model_selection import StratifiedKFold
cv = StratifiedKFold(n_splits=n, shuffle=False)
return cv
def inner_loop_cv_wells(cv_name, n, Dataset_out, X, y):
if cv_name == "LeaveWellOut":
from sklearn.model_selection import LeavePGroupsOut
groups=Dataset_out["Well Name"]
cv=list(LeavePGroupsOut(n_groups=n).split(X,y,groups))
return cv
##################################################################
############# Choose a model - Includes old attempts #############
##################################################################
def select_estimator(model_name):
## XGB
if model_name == "XGB":
import xgboost as xgb
xgb = xgb.XGBClassifier(gamma=0.05, subsample = 0.7, colsample_bytree = 0.9, min_child_weight = 1, max_depth = 3)
estimator = xgb
scores = ["f1_weighted"]
parameters = [{
"gamma":[0.05, 0.1, 0.3, 0.5, 0.7],
"max_depth":[2, 3, 4, 5, 6, 7, 8],
"min_child_weight":[1],
"subsample":[0.7, 0.8],
"colsample_bytree":[0.9]}]
elif model_name == "LogisticRegression":
from sklearn.linear_model import LogisticRegression
estimator = LogisticRegression()
parameters = [{"C": [1, 10, 100, 1000]}]
scores = ["f1_weighted"]
## Model to use for make extra features based on categories
elif model_name == "XGB_Litho":
import xgboost as xgb
xgb = xgb.XGBClassifier(gamma=0.05, subsample = 0.7, colsample_bytree = 0.9, min_child_weight = 1, max_depth = 3)
estimator = xgb
scores = ["f1_weighted"]
parameters = [{
"gamma":[0.1],
"max_depth":[7],
"min_child_weight":[1],
"subsample":[0.7],
"colsample_bytree":[0.9]}]
## Best score for attempt 4
elif model_name == "XGB_Best_v3":
import xgboost as xgb
xgb = xgb.XGBClassifier(gamma=0.05, subsample = 0.7, colsample_bytree = 0.9, min_child_weight = 1, max_depth = 3)
estimator = xgb
scores = ["f1_macro"]
parameters = [{'min_child_weight': [1], 'gamma': [0.7], 'subsample': [0.7], 'max_depth': [3], 'colsample_bytree': [0.9]}]
## OneVsOne attempt using scoping __ which alllows for GridCV and OnevsOne
elif model_name == "XGB_OneVsOne":
import xgboost as xgb
from sklearn.multiclass import OneVsOneClassifier
estimator = OneVsOneClassifier(xgb.XGBClassifier())
scores = ["f1_micro"]
parameters = [{'estimator__min_child_weight': [1, 3, 5, 7],
'estimator__gamma': [0.7, 0.5, 0.3, 0.1, 0.05],
'estimator__subsample': [0.7],
'estimator__max_depth': [2, 3, 5, 7, 10],
'estimator__colsample_bytree': [0.9],
"estimator__reg_alpha":[0.8, 1]}]
## OneVsOne attempt best score for attempt 5
elif model_name == "XGB_Best_v5":
import xgboost as xgb
from sklearn.multiclass import OneVsOneClassifier
estimator = OneVsOneClassifier(xgb.XGBClassifier())
scores = ["f1_macro"]
parameters = [{'estimator__subsample': [0.8], 'estimator__max_depth': [3], 'estimator__reg_alpha': [1], 'estimator__colsample_bytree': [0.9], 'estimator__min_child_weight': [5], 'estimator__gamma': [0.1]}]
## OneVsOne with Leave well out CV - missing two wells and F9
elif model_name == "XGB_Best_v6":
import xgboost as xgb
from sklearn.multiclass import OneVsOneClassifier
estimator = OneVsOneClassifier(xgb.XGBClassifier())
scores = ["f1_micro"]
parameters = [{'estimator__subsample': [0.7], 'estimator__max_depth': [2], 'estimator__reg_alpha': [1], 'estimator__colsample_bytree': [0.9], 'estimator__min_child_weight': [5], 'estimator__gamma': [0.7]}]
## OneVsOne with Leave well out CV all wells and F9 included
elif model_name == "XGB_Best_v7":
import xgboost as xgb
from sklearn.multiclass import OneVsOneClassifier
estimator = OneVsOneClassifier(xgb.XGBClassifier())
scores = ["f1_micro"]
parameters = [{'estimator__subsample': [0.7], 'estimator__max_depth': [2], 'estimator__reg_alpha': [0.5], 'estimator__colsample_bytree': [0.9], 'estimator__min_child_weight': [7], 'estimator__gamma': [0.7]}]
# Try 9
#Best classifier score: 0.56416854294 : {'estimator__gamma': 0.5, 'estimator__min_child_weight': 1, 'estimator__reg_alpha': 1, 'estimator__colsample_bytree': 0.9, 'estimator__subsample': 0.7, 'estimator__max_depth': 3}
elif model_name == "XGB_Best_v8":
import xgboost as xgb
from sklearn.multiclass import OneVsOneClassifier
estimator = OneVsOneClassifier(xgb.XGBClassifier())
scores = ["f1_micro"]
parameters = [{'estimator__subsample': [0.7], 'estimator__max_depth': [3], 'estimator__reg_alpha': [1], 'estimator__colsample_bytree': [0.9], 'estimator__min_child_weight': [1], 'estimator__gamma': [0.5]}]
return estimator, scores, parameters
#####################################################################################
####################### Grid Search for best parameters #############################
#####################################################################################
def optimise_parameters(X, y, estimator, scores, cv, parameters):
from sklearn.model_selection import GridSearchCV
from sklearn.metrics import classification_report
Full_dict = {}
best_result_list = []
for score in scores:
clf = GridSearchCV(estimator=estimator, param_grid=parameters, cv=cv,
scoring=score)
## Fit the model
clf.fit(X, y)
## Store the best parameters
result = clf.best_params_
best_result_list.append(result)
## Capture all results data as a dataframe and store in a dictionary
Full_dict[str(score)]=(pd.DataFrame(clf.cv_results_))
Best_Estimator = clf.best_estimator_
print ("Best classifier score:", clf.best_score_, ":", clf.best_params_)
return best_result_list, Full_dict, Best_Estimator, clf
#################################################################
############# Feature selection ###################
#################################################################
def Feature_Importance(X, y):
# Feature Importance Using RandomForestClassifier
from sklearn import metrics
from sklearn.ensemble import RandomForestClassifier
tree_X = X
tree_y = y
print (tree_X.shape)
print (tree_y.shape)
# fit a RandomForest Model
clf_f = RandomForestClassifier(n_estimators=100)
clf_f.fit(tree_X, tree_y)
X_selected = clf_f.transform(tree_X)
# Turn into Data Frame
X_df = pd.DataFrame(Input_X)
Feature_List = list(X_df.columns.values)
df_F_L = pd.DataFrame(Feature_List)
df_F_L["Feature_importance"]=clf_f.feature_importances_
return df_F_L
#################################################################
############ Methods to plot well bores #########################
#################################################################
### Code taken from other entries to make a well log display ###
def make_facies_log_plot(logs, facies_colors):
#### Import to make def well plots work ####
from mpl_toolkits.axes_grid1 import make_axes_locatable
import matplotlib.colors as colors
import matplotlib as mpl
#make sure logs are sorted by depth
logs = logs.sort_values(by='Depth')
cmap_facies = colors.ListedColormap(
facies_colors[0:len(facies_colors)], 'indexed')
ztop=logs.Depth.min(); zbot=logs.Depth.max()
cluster=np.repeat(np.expand_dims(logs['Predicted_y'].values,1), 100, 1)
f, ax = plt.subplots(nrows=1, ncols=6, figsize=(8, 12))
ax[0].plot(logs.GR, logs.Depth, '-g')
ax[1].plot(logs.ILD_log10, logs.Depth, '-')
ax[2].plot(logs.DeltaPHI, logs.Depth, '-', color='0.5')
ax[3].plot(logs.PHIND, logs.Depth, '-', color='r')
ax[4].plot(logs.PE, logs.Depth, '-', color='black')
im=ax[5].imshow(cluster, interpolation='none', aspect='auto',
cmap=cmap_facies,vmin=1,vmax=9)
divider = make_axes_locatable(ax[5])
cax = divider.append_axes("right", size="20%", pad=0.05)
cbar=plt.colorbar(im, cax=cax)
cbar.set_label((17*' ').join([' SS ', 'CSiS', 'FSiS',
'SiSh', ' MS ', ' WS ', ' D ',
' PS ', ' BS ']))
cbar.set_ticks(range(0,1)); cbar.set_ticklabels('')
for i in range(len(ax)-1):
ax[i].set_ylim(ztop,zbot)
ax[i].invert_yaxis()
ax[i].grid()
ax[i].locator_params(axis='x', nbins=3)
ax[0].set_xlabel("GR")
ax[0].set_xlim(logs.GR.min(),logs.GR.max())
ax[1].set_xlabel("ILD_log10")
ax[1].set_xlim(logs.ILD_log10.min(),logs.ILD_log10.max())
ax[2].set_xlabel("DeltaPHI")
ax[2].set_xlim(logs.DeltaPHI.min(),logs.DeltaPHI.max())
ax[3].set_xlabel("PHIND")
ax[3].set_xlim(logs.PHIND.min(),logs.PHIND.max())
ax[4].set_xlabel("PE")
ax[4].set_xlim(logs.PE.min(),logs.PE.max())
ax[5].set_xlabel('Predicted_y')
ax[1].set_yticklabels([]); ax[2].set_yticklabels([]); ax[3].set_yticklabels([])
ax[4].set_yticklabels([]); ax[5].set_yticklabels([])
ax[5].set_xticklabels([])
f.suptitle('Well: %s'%logs.iloc[0]['Well Name'], fontsize=14,y=0.94)
def compare_facies_plot(logs, facies_colors):
#make sure logs are sorted by depth
from mpl_toolkits.axes_grid1 import make_axes_locatable
import matplotlib.colors as colors
import matplotlib as mpl
logs = logs.sort_values(by='Depth')
cmap_facies = colors.ListedColormap(
facies_colors[0:len(facies_colors)], 'indexed')
ztop=logs.Depth.min(); zbot=logs.Depth.max()
cluster1 = np.repeat(np.expand_dims(logs['Facies'].values,1), 100, 1)
cluster2 = np.repeat(np.expand_dims(logs["Predicted_y"].values,1), 100, 1)
f, ax = plt.subplots(nrows=1, ncols=7, figsize=(9, 12))
ax[0].plot(logs.GR, logs.Depth, '-g')
ax[1].plot(logs.ILD_log10, logs.Depth, '-')
ax[2].plot(logs.DeltaPHI, logs.Depth, '-', color='0.5')
ax[3].plot(logs.PHIND, logs.Depth, '-', color='r')
ax[4].plot(logs.PE, logs.Depth, '-', color='black')
imA = ax[5].imshow(cluster1, interpolation='none', aspect='auto',
cmap=cmap_facies,vmin=1,vmax=9)
imB = ax[6].imshow(cluster2, interpolation='none', aspect='auto',
cmap=cmap_facies,vmin=1,vmax=9)
divider = make_axes_locatable(ax[6])
cax = divider.append_axes("right", size="20%", pad=0.05)
cbar=plt.colorbar(imB, cax=cax)
cbar.set_label((17*' ').join([' SS ', 'CSiS', 'FSiS',
'SiSh', ' MS ', ' WS ', ' D ',
' PS ', ' BS ']))
cbar.set_ticks(range(0,1)); cbar.set_ticklabels('')
for i in range(len(ax)-2):
ax[i].set_ylim(ztop,zbot)
ax[i].invert_yaxis()
ax[i].grid()
ax[i].locator_params(axis='x', nbins=3)
ax[0].set_xlabel("GR")
ax[0].set_xlim(logs.GR.min(),logs.GR.max())
ax[1].set_xlabel("ILD_log10")
ax[1].set_xlim(logs.ILD_log10.min(),logs.ILD_log10.max())
ax[2].set_xlabel("DeltaPHI")
ax[2].set_xlim(logs.DeltaPHI.min(),logs.DeltaPHI.max())
ax[3].set_xlabel("PHIND")
ax[3].set_xlim(logs.PHIND.min(),logs.PHIND.max())
ax[4].set_xlabel("PE")
ax[4].set_xlim(logs.PE.min(),logs.PE.max())
ax[5].set_xlabel('Facies')
ax[6].set_xlabel("Predicted_y")
ax[1].set_yticklabels([]); ax[2].set_yticklabels([]); ax[3].set_yticklabels([])
ax[4].set_yticklabels([]); ax[5].set_yticklabels([])
ax[5].set_xticklabels([])
ax[6].set_xticklabels([])
f.suptitle('Well: %s'%logs.iloc[0]['Well Name'], fontsize=14,y=0.94)
#####################################################################################
#####################################################################################
print("Ready")
In [2]:
######## Initalize and load file ##########
import matplotlib.pyplot as plt
import pandas as pd
import numpy as np
import sys #only needed to determine Python version number
import matplotlib #only needed to determine Matplotlib version number
import seaborn as sns
#Enable inline plotting
%matplotlib inline
############################################
################### Run ####################
############################################
# 0.
################ Initialize and load dataset ##################
Dataset = Start_up()
####### #######
# 1.
################ Prepare X and y. Produces y in one column or as one hot vector format. ##################
trX, t_y, t_y_one_column, new_classes = prepare_X_Y(Dataset)
####### #######
# Extra - Use a machine learning algorithm to add to PE.
# Appears to lead to overfitting
clf_pe = train_ml_pe(trX)
trX2 = predict_pe(clf_pe, trX)
# 2.
################ Feature engineering ##################
df_X = prepare_X(trX)
####### #######
# 3.
############### Drop uneeded features ##################
Input_X = drop_features(df_X)
####### #######
# 4.
############### Set inputs to X and y variable ##################
X, y = X_y_inputs(Input_X, t_y_one_column)
####### #######
# 5.
############### Fit a scaler and transform X before training ##################
Scaler_name ="Standard"
Scaler = select_scaler(Scaler_name)
X = Scaler.fit_transform(X)
####### #######
# 6.
############### Make dataframe showing feature importance ##################
df_F_L = Feature_Importance(X, y)
####### #######
print("Begin Grid Search .......")
# 7.
############## Select estimator, scoring methodology and parameter grid to be used in GridSearchCV ###########
# model_name = "XGB"
model_name = "XGB_Best_v8"
estimator, scores, parameters = select_estimator(model_name)
####### ########
# 8.
############## Choose Cross Validation Stratergy ###############
n=6 # number of splits or number of wells to leave out.
cv_name = "KFold"
cv = inner_loop_cv(cv_name, n)
#cv_name = "LeaveWellOut"
#cv = inner_loop_cv_wells(cv_name, n, Dataset_out, X, y)
####### ########
# 9.
########### Run GridSearchCV return the best classifier to be used to predict on test results ##############
best_result_list, Full_dict, Best_Estimator, clf = optimise_parameters(X, y, estimator, scores, cv, parameters)
####### ########
################################################
# Finished - Clf to be used to predict on test #
################################################
In [3]:
plt.rcParams['figure.figsize']=(5,15)
ax = sns.barplot(x="Feature_importance", y=0, data=df_F_L, palette="GnBu_r")
In [4]:
p_y=clf.predict(X)
def Plot_Confusion_Matrix(p_y, y):
#### Print Score ####
from sklearn.metrics import f1_score
print ()
print ("F1 score:", f1_score(y, p_y, average="weighted"))
#### Make Confusion Matrix ####
from sklearn.metrics import confusion_matrix
cnf_matrix = confusion_matrix(y, p_y)
sns.heatmap(cnf_matrix, annot=True, fmt="d")
plt.ylabel('True label')
plt.xlabel('Predicted label')
plt.rcParams['figure.figsize']=(6,5)
Plot_Confusion_Matrix(p_y, y)
In [5]:
##### Load test data #####
df_test = pd.read_csv('validation_data_nofacies.csv')
#### Feature Engineering #####
Xts1 = prepare_X(df_test)
#Xts2 = drop_features(Xts1)
Xts2 = Xts1
Xts2 = Xts2.drop(["Formation", "Well Name"], 1)
print(Xts2.shape)
#### Scale X for test data ####
Xts3 = Scaler.transform(Xts2)
#### Predict y on the test data ###
p_y_ts=clf.predict(Xts3)
#### Merge predicted y with the test data ####
df_test_result=df_test
df_test_result["Predicted_y"]=p_y_ts
#### Save results to a csv file ####
df_test_result.to_csv("ADMC_Prediction_XGB_Submission_9")
df_test_result.head()
Out[5]:
In [7]:
facies_colors = ['#F4D03F', '#F5B041','#DC7633','#6E2C00', '#1B4F72','#2E86C1', '#AED6F1', '#A569BD', '#196F3D']
##### Plot result Wells ####
make_facies_log_plot(
df_test_result[df_test_result['Well Name'] == 'STUART'],
facies_colors)
make_facies_log_plot(
df_test_result[df_test_result['Well Name'] == 'CRAWFORD'],
facies_colors)
Consider what and why is creating errors. Previous investigations have used ROC curves and Precision vs. Recall plots. Additional testing has shown that slightly higher training scores do not always give higher test scores. To further optimize the results, visual comparisson of where the classifier is succesful and where there are errors. One outcome of this was the decision to remove CHURCHMAN BIBLE - This could be refined by only eliminating the specific depth intervals (3060-3070) where there is log interpolation which should not be inlcuded in training.
In [15]:
#### Merge predicted y with the test data ####
facies_colors = ['#F4D03F', '#F5B041','#DC7633','#6E2C00', '#1B4F72','#2E86C1', '#AED6F1', '#A569BD', '#196F3D']
Facies_Plot = trX
grouped = Facies_Plot.groupby('Well Name')
Facies_Plot = Facies_Plot.rename(columns = {"Well Name":"Well_Name"}) # To be able to use the header in the .unique method
List = Facies_Plot.Well_Name.unique()
df_test_result=trX
df_test_result["Facies"]=Dataset["Facies"]
df_test_result["Predicted_y"]=p_y
for i in List:
compare_facies_plot(
df_test_result[df_test_result['Well Name'] == i],
facies_colors)
In [ ]: