gamma = [0.01, 0.001, 0.0001]
#Optimize the linear classifier first
f = open('./GridSearch_output', 'w')
f.write('PCA components: %d\n' % (n_components))
write2 = 'X_train: %d, X_test: %d\n' % (X_train_PCA.shape[0], X_test_PCA.shape[0])
f.write(write2)
f.write("Scores: mAP, accuracy_score, avg_Precision, avg_Recall\n\n")
for c in C:
#Optimize the linear kernel first
for lo in loss:
for pen in penalty:
classifier = OneVsRestClassifier(LinearSVC(random_state=0, C=c, loss=lo, penalty=pen)).fit(X_train_PCA, Y_train)
Scores = classify_library.metric_scores(classifier, X_test_PCA, Y_test)
setting = "Settings: Linear SVM, C: %d, loss: %s, penalty: %s\n" % (c,lo,pen)
score = "Scores: (%f, %f, %f, %f)\n" % (Scores[0], Scores[1], Scores[2], Scores[3])
f.write(setting)
f.write(score)
f.write('\n')
#Optimize the non-linear kernels
for ker in kernel:
for gam in gamma:
classifier = OneVsRestClassifier(svm.SVC(random_state=0, C=c, kernel=ker, gamma=gam)).fit(X_train_PCA, Y_train)
Scores = classify_library.metric_scores(classifier, X_test_PCA, Y_test)
setting = "Settings: SVM kernel: %s, C: %d, gamma: %f\n" % (ker,c,gam)
score = "Scores: (%f, %f, %f, %f)\n" % (Scores[0], Scores[1], Scores[2], Scores[3])
f.write(setting)
f.write(score)
f.write('\n')
best_ker = ker
best_gam = gam
print 'Non-linear: C = ' + str(
c) + ', kernel = ' + ker + ', Gamma = ' + str(gam)
print 'Precisions of CV = ' + str(scores)
print 'Mean Precision = ' + str(scores.mean())
print 'Confusion Matrix = '
print conf_mat
print '\n-----------------------------------------------------\n'
print '\n########################################################\n'
print 'BEST LINEAR SVM (CV precision = ' + str(best_linear * 100) + ' %)\n'
#Best Linear SVM
classifier_l = OneVsRestClassifier(
LinearSVC(random_state=0, C=best_c_l, loss=best_loss,
penalty=best_pen)).fit(X_train_final, Y_train)
Scores = classify_library.metric_scores(classifier_l, X_test_final, Y_test)
print "Settings: Linear SVM, C: %d, loss: %s, penalty: %s" % (
best_c_l, best_loss, best_pen)
print "Scores in test: (%f, %f, %f, %f)\n" % (Scores[0], Scores[1],
Scores[2], Scores[3])
print "Confusion Matrix = "
print conf_mat_l
print '\n########################################################\n'
print 'BEST NON-LINEAR SVM (CV precision = ' + str(
best_non_linear * 100) + ' %)\n'
#Best non-linear SVM
classifier_nl = OneVsRestClassifier(
svm.SVC(random_state=0, C=best_c_nl, kernel=best_ker,
gamma=best_gam)).fit(X_train_final, Y_train)
Scores = classify_library.metric_scores(classifier_nl, X_test_final,
X_train, Y_train = make_frame_matrix(training, training_output, class_index)
X_test, Y_test = make_frame_matrix(testing, testing_output, class_index)
### Reduced PCA dimension to 1000
# In[18]:
X_PCA, _ = make_frame_matrix(training_PCA, training_output, class_index)
pca = PCA(n_components=1000)
pca.fit(X_PCA)
X_train_PCA = pca.transform(X_train)
X_test_PCA = pca.transform(X_test)
# In[22]:
classifier = OneVsRestClassifier(
LinearSVC(random_state=0, C=1, loss='l2',
penalty='l2')).fit(X_train_PCA, Y_train)
classify_library.metric_scores(classifier, X_test_PCA, Y_test, verbose=True)
# In[27]:
baseline_file = "./baseline"
np.savez(baseline_file,
X_train_PCA=X_train_PCA,
X_test_PCA=X_test_PCA,
Y_train=Y_train,
Y_test=Y_test)
# In[ ]:
f = open('./GridSearch_output', 'w')
f.write('PCA components: %d\n' % (n_components))
write2 = 'X_train: %d, X_test: %d\n' % (X_train_PCA.shape[0],
X_test_PCA.shape[0])
f.write(write2)
f.write("Scores: mAP, accuracy_score, avg_Precision, avg_Recall\n\n")
for c in C:
#Optimize the linear kernel first
for lo in loss:
for pen in penalty:
classifier = OneVsRestClassifier(
LinearSVC(random_state=0, C=c, loss=lo,
penalty=pen)).fit(X_train_PCA, Y_train)
Scores = classify_library.metric_scores(classifier, X_test_PCA,
Y_test)
setting = "Settings: Linear SVM, C: %d, loss: %s, penalty: %s\n" % (
c, lo, pen)
score = "Scores: (%f, %f, %f, %f)\n" % (Scores[0], Scores[1],
Scores[2], Scores[3])
f.write(setting)
f.write(score)
f.write('\n')
#Optimize the non-linear kernels
for ker in kernel:
for gam in gamma:
classifier = OneVsRestClassifier(
svm.SVC(random_state=0, C=c, kernel=ker,
gamma=gam)).fit(X_train_PCA, Y_train)
Scores = classify_library.metric_scores(classifier, X_test_PCA,
Y_test)
X_test, Y_test = make_frame_matrix(testing,testing_output,class_index) ### Reduced PCA dimension to 1000 # In[18]: X_PCA, _ = make_frame_matrix(training_PCA, training_output, class_index) pca = PCA(n_components=1000) pca.fit(X_PCA) X_train_PCA = pca.transform(X_train) X_test_PCA = pca.transform(X_test) # In[22]: classifier = OneVsRestClassifier(LinearSVC(random_state=0, C=1, loss='l2', penalty='l2')).fit(X_train_PCA, Y_train) classify_library.metric_scores(classifier, X_test_PCA, Y_test, verbose=True) # In[27]: baseline_file = "./baseline" np.savez(baseline_file, X_train_PCA=X_train_PCA, X_test_PCA=X_test_PCA, Y_train=Y_train, Y_test=Y_test) # In[ ]:
flname_test = '/home/zhenyang/Workspace/data/UCF101/features/UCF101_test1.fv'
if os.path.exists(flname_test + '.npz'):
data = np.load(flname_test + '.npz')
X_test = data['X_test']
Y_test = data['Y_test']
else:
X_test, Y_test = make_FV_matrix(videos_test, fv_dir, labels_test)
np.savez(flname_test, X_test=X_test, Y_test=Y_test)
# TRAINING
model_file = '/home/zhenyang/Workspace/data/UCF101/models/UCF101_linearsvm_traintest1.model'
if os.path.exists(model_file + '.pkl'):
with open(model_file + '.pkl', 'r') as fp:
classifier = pickle.load(fp)
else:
estimator = OneVsRestClassifier(
LinearSVC(random_state=0, C=100, loss='l1', penalty='l2'))
classifier = estimator.fit(X_train, Y_train)
# store the model in a pickle file
with open(model_file + '.pkl', 'w') as fp:
pickle.dump(classifier, fp)
# TESTING
result_file = '/home/zhenyang/Workspace/data/UCF101/results/UCF101_linearsvm_traintest1.result'
results = classify_library.metric_scores(classifier,
X_test,
Y_test,
verbose=True)
print results
X_train, Y_train = make_FV_matrix(videos_train, fv_dir, labels_train)
np.savez(flname_train, X_train=X_train, Y_train=Y_train)
flname_test = '/home/zhenyang/Workspace/data/UCF101/features/UCF101_test1.fv'
if os.path.exists(flname_test+'.npz'):
data = np.load(flname_test+'.npz')
X_test = data['X_test']
Y_test = data['Y_test']
else:
X_test, Y_test = make_FV_matrix(videos_test, fv_dir, labels_test)
np.savez(flname_test, X_test=X_test, Y_test=Y_test)
# TRAINING
model_file = '/home/zhenyang/Workspace/data/UCF101/models/UCF101_linearsvm_traintest1.model'
if os.path.exists(model_file+'.pkl'):
with open(model_file+'.pkl', 'r') as fp:
classifier = pickle.load(fp)
else:
estimator = OneVsRestClassifier(LinearSVC(random_state=0, C=100, loss='l1', penalty='l2'))
classifier = estimator.fit(X_train, Y_train)
# store the model in a pickle file
with open(model_file+'.pkl', 'w') as fp:
pickle.dump(classifier, fp)
# TESTING
result_file = '/home/zhenyang/Workspace/data/UCF101/results/UCF101_linearsvm_traintest1.result'
results = classify_library.metric_scores(classifier, X_test, Y_test, verbose=True)
print results