In [1]:
%pylab
%matplotlib inline
In [2]:
cd ..
In [3]:
import sys
import numpy as np
import skimage
import cv2
import sklearn
import imp
In [4]:
import holoviews
In [24]:
import neukrill_net.utils
import neukrill_net.image_features
import neukrill_net.highlevelfeatures
import neukrill_net.stacked
In [17]:
import skimage.feature
In [21]:
import sklearn.ensemble
In [6]:
import time
In [7]:
#%pdb
In [8]:
settings = neukrill_net.utils.Settings('settings.json')
In [9]:
X,y = settings.flattened_train_paths(settings.classes)
In [13]:
reload(neukrill_net.highlevelfeatures)
Out[13]:
In [25]:
reload(neukrill_net.image_features)
Out[25]:
In [14]:
attrlst = ['height','width','numpixels','aspectratio','mean','std','stderr',
'numwhite','propwhite','numnonwhite','propnonwhite','numblack','propblack','numbool','propbool']
hlf = neukrill_net.highlevelfeatures.BasicAttributes(attrlst)
hlf += neukrill_net.highlevelfeatures.Haralick()
hlf += neukrill_net.highlevelfeatures.ThresholdAdjacency()
hlf += neukrill_net.highlevelfeatures.ContourMoments()
hlf += neukrill_net.highlevelfeatures.ContourHistogram()
hlf += neukrill_net.highlevelfeatures.CoocurProps()
In [18]:
hlf.preprocess_and_extract_image(neukrill_net.highlevelfeatures.loadimage(X[0]))
Out[18]:
In [20]:
neukrill_net.highlevelfeatures.ContourHistogram().preprocess_and_extract_image(neukrill_net.highlevelfeatures.loadimage(X[0]))
Out[20]:
In [23]:
kprf_base = sklearn.ensemble.RandomForestClassifier(n_estimators=1000, max_depth=25,
min_samples_leaf=20, n_jobs=12, random_state=42)
In [26]:
max_num_kp = 150
detector_list = [lambda image: neukrill_net.image_features.get_ORB_keypoints(image, n=max_num_kp, patchSize=9),
lambda image: neukrill_net.image_features.get_BRISK_keypoints(image, n=max_num_kp),
lambda image: neukrill_net.image_features.get_MSER_keypoints(image, n=max_num_kp)]
describer_list = [neukrill_net.image_features.get_ORB_descriptions,
neukrill_net.image_features.get_BRISK_descriptions,
neukrill_net.image_features.get_ORB_descriptions]
In [27]:
for index,detector in enumerate(detector_list):
hlf += neukrill_net.highlevelfeatures.KeypointEnsembleClassifier(detector, describer_list[index], kprf_base,
return_num_kp=True, summary_method='vote')
In [28]:
rf_clf = sklearn.ensemble.RandomForestClassifier(n_estimators=2500, max_depth=30,
min_samples_leaf=1, n_jobs=12, random_state=42)
In [30]:
reload(neukrill_net.stacked)
Out[30]:
In [32]:
import sklearn.pipeline
In [33]:
selector = sklearn.feature_selection.SelectPercentile(sklearn.feature_selection.f_classif, percentile=33)
In [39]:
stack_pipe = sklearn.pipeline.Pipeline([('filter', selector), ('clf', rf_clf)])
stacked_clf = neukrill_net.stacked.StackedClassifier(hlf, stack_pipe, inner_prop=0.25, random_state=42)
In [37]:
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(X, y, test_size=0.5, random_state=42)
In [41]:
clf = stacked_clf
t0 = time.time()
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(X, y, test_size=0.5, random_state=42)
clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
print("Accuracy={}".format(clf.score(X_test, y_test)))
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, clf.predict_proba(X_test))))
In [44]:
kp_hlf = neukrill_net.highlevelfeatures.KeypointEnsembleClassifier(detector_list[1], describer_list[1], kprf_base,
return_num_kp=True, summary_method='vote')
In [45]:
kp_stacked_clf = neukrill_net.stacked.StackedClassifier(kp_hlf, stack_pipe, inner_prop=0.25, random_state=42)
In [46]:
clf = kp_stacked_clf
t0 = time.time()
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(X, y, test_size=0.5, random_state=42)
clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
print("Accuracy={}".format(clf.score(X_test, y_test)))
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, clf.predict_proba(X_test))))
In [ ]:
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, clf.predict_proba(X_test))))
This is similar to just the Contour Moments and Haralick features
On reduced
In [22]:
my_X = X_new
clf = sklearn.ensemble.RandomForestClassifier(n_estimators=1000, max_depth=20, min_samples_leaf=5, n_jobs=12, random_state=42)
t0 = time.time()
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(
sklearn.preprocessing.StandardScaler().fit_transform(my_X), y, test_size=0.5, random_state=42)
clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
print("Accuracy={}".format(clf.score(X_test, y_test)))
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, clf.predict_proba(X_test))))
Does slightly worse with fewer features.
Maybe it was too few?
In [23]:
my_X = sklearn.feature_selection.SelectKBest(sklearn.feature_selection.f_classif, k=100).fit_transform(XF.squeeze(0), y)
clf = sklearn.ensemble.RandomForestClassifier(n_estimators=1000, max_depth=20, min_samples_leaf=5, n_jobs=12, random_state=42)
t0 = time.time()
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(
sklearn.preprocessing.StandardScaler().fit_transform(my_X), y, test_size=0.5, random_state=42)
clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
print("Accuracy={}".format(clf.score(X_test, y_test)))
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, clf.predict_proba(X_test))))
In [24]:
import neukrill_net.taxonomy
In [29]:
reload(neukrill_net.stacked)
Out[29]:
In [36]:
reload(neukrill_net.taxonomy)
Out[36]:
In [37]:
neukrill_net.taxonomy.taxonomy
Out[37]:
In [38]:
settings.classes
Out[38]:
In [39]:
marked_taxonomy = neukrill_net.stacked.propagate_labels_to_leaves(neukrill_net.taxonomy.taxonomy, settings.classes)
In [40]:
marked_taxonomy
Out[40]:
In [41]:
base_clf = sklearn.ensemble.RandomForestClassifier(n_estimators=1000, max_depth=20, min_samples_leaf=5, n_jobs=12, random_state=42)
hier_clf = neukrill_net.stacked.HierarchyClassifier(marked_taxonomy, base_clf)
t0 = time.time()
hier_clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
t0 = time.time()
p = hier_clf.predict_proba(X_test)
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, p)))
print("Time={}".format(time.time()-t0))
print("Accuracy={}".format(sklearn.metrics.accuracy_score(y_test,np.argmax(p,1))))
In [42]:
base_clf = sklearn.ensemble.RandomForestClassifier(n_estimators=1000, max_depth=20, min_samples_leaf=5, n_jobs=12, random_state=42)
hier_clf = neukrill_net.stacked.HierarchyClassifier(marked_taxonomy, base_clf)
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(
sklearn.preprocessing.StandardScaler().fit_transform(XF.squeeze(0)), y, test_size=0.5, random_state=42)
t0 = time.time()
hier_clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
t0 = time.time()
p = hier_clf.predict_proba(X_test)
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, p)))
print("Time={}".format(time.time()-t0))
print("Accuracy={}".format(sklearn.metrics.accuracy_score(y_test,np.argmax(p,1))))
Try with a pipline to reduce the number of features at each level
In [44]:
import sklearn.pipeline
In [47]:
base_clf = sklearn.ensemble.RandomForestClassifier(n_estimators=1000, max_depth=20, min_samples_leaf=5, n_jobs=12, random_state=42)
best_filter = sklearn.feature_selection.SelectKBest(sklearn.feature_selection.f_classif, k=100)
base_pipe = sklearn.pipeline.Pipeline([('filter', best_filter), ('clf', base_clf)])
hier_clf = neukrill_net.stacked.HierarchyClassifier(marked_taxonomy, base_pipe)
my_X = sklearn.preprocessing.StandardScaler().fit_transform(XF.squeeze(0))
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(my_X, y, test_size=0.5, random_state=42)
t0 = time.time()
hier_clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
t0 = time.time()
p = hier_clf.predict_proba(X_test)
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, p)))
print("Time={}".format(time.time()-t0))
print("Accuracy={}".format(sklearn.metrics.accuracy_score(y_test,np.argmax(p,1))))
In [47]:
base_clf = sklearn.ensemble.RandomForestClassifier(n_estimators=1000, max_depth=20, min_samples_leaf=5, n_jobs=12, random_state=42)
best_filter = sklearn.feature_selection.SelectKBest(sklearn.feature_selection.f_classif, k=100)
base_pipe = sklearn.pipeline.Pipeline([('filter', best_filter), ('clf', base_clf)])
hier_clf = neukrill_net.stacked.HierarchyClassifier(marked_taxonomy, base_pipe)
my_X = sklearn.preprocessing.StandardScaler().fit_transform(XF.squeeze(0))
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(my_X, y, test_size=0.5, random_state=42)
t0 = time.time()
hier_clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
t0 = time.time()
p = hier_clf.predict_proba(X_test)
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, p)))
print("Time={}".format(time.time()-t0))
print("Accuracy={}".format(sklearn.metrics.accuracy_score(y_test,np.argmax(p,1))))
In [48]:
clf = sklearn.linear_model.LogisticRegression(random_state=42)
In [49]:
t0 = time.time()
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(
sklearn.preprocessing.StandardScaler().fit_transform(XF.squeeze(0)), y, test_size=0.5, random_state=42)
clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
print("Accuracy={}".format(clf.score(X_test, y_test)))
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, clf.predict_proba(X_test))))
In [50]:
clf = sklearn.linear_model.LogisticRegression(random_state=42)
my_X = sklearn.feature_selection.SelectKBest(sklearn.feature_selection.f_classif, k=100).fit_transform(XF.squeeze(0), y)
t0 = time.time()
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(
sklearn.preprocessing.StandardScaler().fit_transform(my_X), y, test_size=0.5, random_state=42)
clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
print("Accuracy={}".format(clf.score(X_test, y_test)))
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, clf.predict_proba(X_test))))
In [55]:
base_clf = sklearn.linear_model.LogisticRegression(random_state=42)
hier_clf = neukrill_net.stacked.HierarchyClassifier(marked_taxonomy, base_clf)
my_X = sklearn.preprocessing.StandardScaler().fit_transform(XF.squeeze(0))
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(my_X, y, test_size=0.5, random_state=42)
t0 = time.time()
hier_clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
t0 = time.time()
p = hier_clf.predict_proba(X_test)
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, p)))
print("Time={}".format(time.time()-t0))
print("Accuracy={}".format(sklearn.metrics.accuracy_score(y_test,np.argmax(p,1))))
In [52]:
base_clf = sklearn.linear_model.LogisticRegression(random_state=42)
best_filter = sklearn.feature_selection.SelectKBest(sklearn.feature_selection.f_classif, k=100)
base_pipe = sklearn.pipeline.Pipeline([('filter', best_filter), ('clf', base_clf)])
hier_clf = neukrill_net.stacked.HierarchyClassifier(marked_taxonomy, base_pipe)
my_X = sklearn.preprocessing.StandardScaler().fit_transform(XF.squeeze(0))
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(my_X, y, test_size=0.5, random_state=42)
t0 = time.time()
hier_clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
t0 = time.time()
p = hier_clf.predict_proba(X_test)
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, p)))
print("Time={}".format(time.time()-t0))
print("Accuracy={}".format(sklearn.metrics.accuracy_score(y_test,np.argmax(p,1))))
In [56]:
clf = sklearn.svm.SVC(kernel='linear', probability=True, random_state=42)
t0 = time.time()
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(
sklearn.preprocessing.StandardScaler().fit_transform(XF.squeeze(0)), y, test_size=0.5, random_state=42)
clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
print("Accuracy={}".format(clf.score(X_test, y_test)))
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, clf.predict_proba(X_test))))
In [57]:
clf = sklearn.svm.SVC(kernel='linear', probability=True, random_state=42)
my_X = sklearn.feature_selection.SelectKBest(sklearn.feature_selection.f_classif, k=100).fit_transform(XF.squeeze(0), y)
t0 = time.time()
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(
sklearn.preprocessing.StandardScaler().fit_transform(my_X), y, test_size=0.5, random_state=42)
clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
print("Accuracy={}".format(clf.score(X_test, y_test)))
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, clf.predict_proba(X_test))))
In [58]:
base_clf = sklearn.svm.SVC(kernel='linear', probability=True, random_state=42)
hier_clf = neukrill_net.stacked.HierarchyClassifier(marked_taxonomy, base_clf)
my_X = sklearn.preprocessing.StandardScaler().fit_transform(XF.squeeze(0))
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(my_X, y, test_size=0.5, random_state=42)
t0 = time.time()
hier_clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
t0 = time.time()
p = hier_clf.predict_proba(X_test)
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, p)))
print("Time={}".format(time.time()-t0))
print("Accuracy={}".format(sklearn.metrics.accuracy_score(y_test,np.argmax(p,1))))
In [59]:
base_clf = sklearn.svm.SVC(kernel='linear', probability=True, random_state=42)
best_filter = sklearn.feature_selection.SelectKBest(sklearn.feature_selection.f_classif, k=100)
base_pipe = sklearn.pipeline.Pipeline([('filter', best_filter), ('clf', base_clf)])
hier_clf = neukrill_net.stacked.HierarchyClassifier(marked_taxonomy, base_pipe)
my_X = sklearn.preprocessing.StandardScaler().fit_transform(XF.squeeze(0))
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(my_X, y, test_size=0.5, random_state=42)
t0 = time.time()
hier_clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
t0 = time.time()
p = hier_clf.predict_proba(X_test)
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, p)))
print("Time={}".format(time.time()-t0))
print("Accuracy={}".format(sklearn.metrics.accuracy_score(y_test,np.argmax(p,1))))
one-vs-one
In [60]:
clf = sklearn.svm.SVC(probability=True, random_state=42)
t0 = time.time()
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(
sklearn.preprocessing.StandardScaler().fit_transform(XF.squeeze(0)), y, test_size=0.5, random_state=42)
clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
print("Accuracy={}".format(clf.score(X_test, y_test)))
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, clf.predict_proba(X_test))))
In [61]:
clf = sklearn.svm.SVC(kernel='rbf', probability=True, random_state=42)
my_X = sklearn.feature_selection.SelectKBest(sklearn.feature_selection.f_classif, k=100).fit_transform(XF.squeeze(0), y)
t0 = time.time()
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(
sklearn.preprocessing.StandardScaler().fit_transform(my_X), y, test_size=0.5, random_state=42)
clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
print("Accuracy={}".format(clf.score(X_test, y_test)))
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, clf.predict_proba(X_test))))
In [62]:
base_clf = sklearn.svm.SVC(kernel='rbf', probability=True, random_state=42)
hier_clf = neukrill_net.stacked.HierarchyClassifier(marked_taxonomy, base_clf)
my_X = sklearn.preprocessing.StandardScaler().fit_transform(XF.squeeze(0))
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(my_X, y, test_size=0.5, random_state=42)
t0 = time.time()
hier_clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
t0 = time.time()
p = hier_clf.predict_proba(X_test)
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, p)))
print("Time={}".format(time.time()-t0))
print("Accuracy={}".format(sklearn.metrics.accuracy_score(y_test,np.argmax(p,1))))
In [63]:
base_clf = sklearn.svm.SVC(kernel='rbf', probability=True, random_state=42)
best_filter = sklearn.feature_selection.SelectKBest(sklearn.feature_selection.f_classif, k=100)
base_pipe = sklearn.pipeline.Pipeline([('filter', best_filter), ('clf', base_clf)])
hier_clf = neukrill_net.stacked.HierarchyClassifier(marked_taxonomy, base_pipe)
my_X = sklearn.preprocessing.StandardScaler().fit_transform(XF.squeeze(0))
X_train, X_test, y_train, y_test = sklearn.cross_validation.train_test_split(my_X, y, test_size=0.5, random_state=42)
t0 = time.time()
hier_clf.fit(X_train, y_train)
t1 = time.time()
total = t1-t0
print("Time={}".format(total))
t0 = time.time()
p = hier_clf.predict_proba(X_test)
print("Logloss={}".format(sklearn.metrics.log_loss(y_test, p)))
print("Time={}".format(time.time()-t0))
print("Accuracy={}".format(sklearn.metrics.accuracy_score(y_test,np.argmax(p,1))))
In [ ]: