def launchsession2(num_slots, descriptor_type, randomSplits, levels_pyramid,
useKernelInter, useKernelPyr, rocCurveCM):
start = time.time()
# Read the train and test files
train_images_filenames, test_images_filenames, train_labels, test_labels = dataUtils.readData(
)
#Divide training into training and validation splits
train_percentage = 0.7 #70% training 30%validation
if randomSplits:
TrainingSplit, ValidationSplit = dataUtils.getRandomTrainingValidationSplit(
train_images_filenames, train_labels, train_percentage)
else:
TrainingSplit, ValidationSplit = dataUtils.getTrainingValidationSplit(
train_images_filenames, train_labels, train_percentage)
#Get descriptors D
if levels_pyramid != 0:
D, Train_descriptors, Train_label_per_descriptor, Train_keypoints, Train_image_size = descriptors.extractFeaturesPyramid(
TrainingSplit, descriptor_type, num_slots)
else:
D, Train_descriptors, Train_label_per_descriptor = descriptors.extractFeatures(
TrainingSplit, descriptor_type, num_slots)
#Computing bag of words using k-means and save codebook
k = 512
codebook = BoW.computeCodebook(D, k)
#Determine visual words
if levels_pyramid != 0:
visual_words = BoW.getVisualWordsSpatialPyramid(
codebook, k, Train_descriptors, Train_image_size, Train_keypoints,
levels_pyramid)
else:
visual_words = BoW.getVisualWords(codebook, k, Train_descriptors)
# Train a linear SVM classifier
if useKernelInter | useKernelPyr:
#Kernel intersection
clf, stdSlr, train_scaled = SVMClassifiers.trainSVMKernel(
visual_words,
Train_label_per_descriptor,
useKernelPyr,
levels_pyramid,
Cparam=1,
probabilities=rocCurveCM)
else:
clf, stdSlr = SVMClassifiers.trainSVM(visual_words,
Train_label_per_descriptor,
Cparam=1,
kernel_type='linear',
probabilities=rocCurveCM)
#For test set
if useKernelInter | useKernelPyr:
predictedLabels2 = SVMClassifiers.predictKernel(
test_images_filenames, descriptor_type, clf, stdSlr, train_scaled,
k, codebook, levels_pyramid, num_slots)
accuracy2 = Evaluation.computeAccuracyOld(predictedLabels2,
test_labels)
print 'Final Kernel intersection test accuracy: ' + str(accuracy2)
else:
# Get all the test data and predict their labels
predictedLabels = SVMClassifiers.predict(test_images_filenames,
descriptor_type, stdSlr,
codebook, k, levels_pyramid,
num_slots)
#Compute accuracy
accuracy = Evaluation.getMeanAccuracy(clf, predictedLabels,
test_labels)
print 'Final test accuracy: ' + str(accuracy)
#For validation set
validation_images_filenames, validation_labels = dataUtils.unzipTupleList(
ValidationSplit)
if useKernelInter | useKernelPyr:
predictedLabels2 = SVMClassifiers.predictKernel(
validation_images_filenames, descriptor_type, clf, stdSlr,
train_scaled, k, codebook, levels_pyramid, num_slots)
accuracy2 = Evaluation.computeAccuracyOld(predictedLabels2,
validation_labels)
print 'Final Kernel intersection validation accuracy: ' + str(
accuracy2)
else:
# Get all the test data and predict their labels
predictedLabels = SVMClassifiers.predict(validation_images_filenames,
descriptor_type, stdSlr,
codebook, k, levels_pyramid,
num_slots)
#Compute accuracy
validation_accuracy = Evaluation.getMeanAccuracy(
clf, predictedLabels, validation_labels)
print 'Final validation accuracy: ' + str(validation_accuracy)
#Roc curve and Confusion Matrix
if rocCurveCM:
graphs.rcurve(predictedLabels, validation_labels, clf)
graphs.plot_confusion_matrix(clf,
validation_labels,
stdSlr.transform(predictedLabels),
normalize=False,
title='Confusion matrix',
cmap=plt.cm.Blues)
end = time.time()
print 'Done in ' + str(end - start) + ' secs.'
def launchsession4(layer_taken, randomSplits, k, useServer, method_used):
start = time.time()
# Read the train and test files
if useServer:
train_images_filenames, test_images_filenames, train_labels, test_labels = dataUtils.readServerData(
)
else:
train_images_filenames, test_images_filenames, train_labels, test_labels = dataUtils.readData(
)
#Divide training into training and validation splits
train_percentage = 0.7 #70% training 30%validation
if randomSplits:
TrainingSplit, ValidationSplit = dataUtils.getRandomTrainingValidationSplit(
train_images_filenames, train_labels, train_percentage)
else:
TrainingSplit, ValidationSplit = dataUtils.getTrainingValidationSplit(
train_images_filenames, train_labels, train_percentage)
#Obtain information from VGG ConvNet
CNN_base_model = descriptors.getBaseModel() #Base model
#Compute features
print 'Extracting features'
D, Train_descriptors, Train_label_per_descriptor = descriptors.extractFeaturesMaps(
TrainingSplit, layer_taken, CNN_base_model, method_used)
not_use_BoW_other_layers = method_used[
'method_to_reduce_dim'] == 'Average' or method_used[
'method_to_reduce_dim'] == 'Max'
if layer_taken == 'fc1' or layer_taken == 'fc2' or layer_taken == 'flatten' or not_use_BoW_other_layers:
visual_words = D
codebook = None
else:
if method_used['usePCA'] > 0:
print 'Applying PCA'
D, Train_descriptors, pca = PCA_computing.PCA_to_data(
D, Train_descriptors, method_used['usePCA'])
else:
pca = None
#Computing bag of words using k-means and save codebook when necessary
codebook = BoW.computeCodebook(D, k)
#Determine visual words
visual_words = BoW.getVisualWords(codebook, k, Train_descriptors)
# Train a linear SVM classifier
clf, stdSlr = SVMClassifiers.trainSVM(visual_words,
Train_label_per_descriptor,
Cparam=1,
kernel_type='linear')
#For test set
TestSplit = zip(test_images_filenames, test_labels)
if layer_taken == 'fc1' or layer_taken == 'fc2' or layer_taken == 'flatten' or not_use_BoW_other_layers:
##Not using BoVW
predictedLabels = SVMClassifiers.predict(TestSplit, layer_taken,
stdSlr, clf, CNN_base_model,
method_used)
accuracy = Evaluation.computeAccuracyOld(predictedLabels, test_labels)
print 'Final test accuracy: ' + str(accuracy)
else:
#BoVW
predictedLabels = SVMClassifiers.predictBoVW(TestSplit, layer_taken,
stdSlr, codebook, k,
CNN_base_model, pca,
method_used)
accuracy = Evaluation.getMeanAccuracy(clf, predictedLabels,
test_labels)
print 'Final test accuracy: ' + str(accuracy)
#For validation set
validation_images_filenames, validation_labels = dataUtils.unzipTupleList(
ValidationSplit)
if layer_taken == 'fc1' or layer_taken == 'fc2' or layer_taken == 'flatten' or not_use_BoW_other_layers:
#Not using BoVW
predictedLabels = SVMClassifiers.predict(ValidationSplit, layer_taken,
stdSlr, clf, CNN_base_model,
method_used)
validation_accuracy = Evaluation.computeAccuracyOld(
predictedLabels, validation_labels)
print 'Final validation accuracy: ' + str(validation_accuracy)
else:
#BoVW
predictedLabels = SVMClassifiers.predictBoVW(ValidationSplit,
layer_taken, stdSlr,
codebook, k,
CNN_base_model, pca,
method_used)
validation_accuracy = Evaluation.getMeanAccuracy(
clf, predictedLabels, validation_labels)
print 'Final validation accuracy: ' + str(validation_accuracy)
end = time.time()
print 'Done in ' + str(end - start) + ' secs.'
def launchsession3(num_slots,descriptor_type,randomSplits,levels_pyramid,usePCA):
start = time.time()
# Read the train and test files
train_images_filenames,test_images_filenames,train_labels,test_labels=dataUtils.readData()
#Divide training into training and validation splits
train_percentage=0.7#70% training 30%validation
if randomSplits:
TrainingSplit, ValidationSplit=dataUtils.getRandomTrainingValidationSplit(train_images_filenames,train_labels,train_percentage)
else:
TrainingSplit, ValidationSplit=dataUtils.getTrainingValidationSplit(train_images_filenames,train_labels,train_percentage)
#Get descriptors D
if levels_pyramid>0:
D, Train_descriptors, Train_label_per_descriptor, Train_keypoints, Train_image_size = descriptors.extractFeaturesPyramid(TrainingSplit,descriptor_type,num_slots)
else:
D, Train_descriptors, Train_label_per_descriptor = descriptors.extractFeatures(TrainingSplit, descriptor_type,num_slots)
if usePCA>0:
print 'Applying PCA'
D, Train_descriptors, pca = PCA_computing.PCA_to_data(D, Train_descriptors, usePCA)
else:
pca = None
#Computing gmm
k = 64 # short codebooks (32, 64...)
gmm = fisherVectors.getGMM(D,k)
for idx,TrainDes in enumerate(Train_descriptors):
train_descriptor = np.float32(TrainDes)
Train_descriptors[idx]=train_descriptor
if levels_pyramid > 0:
fisher = fisherVectors.getFisherVectorsSpatialPyramid(Train_descriptors, k, gmm, Train_image_size, Train_keypoints, levels_pyramid)
else:
fisher = fisherVectors.getFisherVectors(Train_descriptors,k,gmm)
# Power-normalization
#fisher=fisherVectors.powerNormalization(fisher)
# L2 normalize
fisher=fisherVectors.normalizeL2(fisher)
# Train a linear SVM classifier
clf, stdSlr=SVMClassifiers.trainSVM(fisher,Train_label_per_descriptor,Cparam=1,kernel_type='linear')
#For test set
# Get all the test data and predict their labels
predictedLabels=SVMClassifiers.predict(test_images_filenames,descriptor_type,stdSlr,gmm, k, levels_pyramid,num_slots,pca)
#Compute accuracy
accuracy = Evaluation.getMeanAccuracy(clf,predictedLabels,test_labels)
print 'Final test accuracy: ' + str(accuracy)
#For validation set
validation_images_filenames,validation_labels=dataUtils.unzipTupleList(ValidationSplit)
# Get all the test data and predict their labels
predictedLabels=SVMClassifiers.predict(validation_images_filenames,descriptor_type,stdSlr, gmm, k, levels_pyramid,num_slots,pca)
#Compute accuracy
validation_accuracy = Evaluation.getMeanAccuracy(clf,predictedLabels,validation_labels)
print 'Final validation accuracy: ' + str(validation_accuracy)
end=time.time()
print 'Done in '+str(end-start)+' secs.'
return model
def catboost_predict(model,df):
col=df.columns.tolist()
test_pool=Pool(df,cat_features=[col.index(i) for i in cat_colums])
return model.predict(test_pool)
def prepare_submission(test_merged,array):
sub_df=pd.DataFrame(columns=['id','num_orders'])
sub_df['id']=test_merged['id'].values
sub_df['num_orders']=array
#sub_df['num_orders_tree']=arrayt
#sub_df['avg']=np.mean(array+arrayt,axis=0)
sub_df.to_csv('sub_tree.csv',index=False)
if __name__=='__main__':
train_merged,test_merged=mergeData(*readData())
catProcess=Categorify(cat_colums,numeric_cols)
catProcess.apply_train(train_merged)
catProcess.apply_test(test_merged)
# print(train_merged.head())
# catmodel=catboost_train(train_merged[numeric_cols+cat_colums],train_merged['num_orders'].astype('float32'))
# pred=catboost_predict(catmodel,test_merged[numeric_cols+cat_colums])
# prepare_submission(test_merged,pred)
rf,xgbr=train_trees(train_merged[numeric_cols+cat_colums],train_merged['num_orders'].astype('float32'))
df=predict_trees([rf,xgbr],test_merged[numeric_cols+cat_colums])
df.to_csv('res.csv')
for res in estimator.predict(input_fn=inp_fn):
predicted.append(res['predictions'])
return np.array(predicted).ravel()
def prepare_submission(test_merged, array, arrayt):
sub_df = pd.DataFrame(columns=['id', 'num_orders'])
sub_df['id'] = test_merged['id'].values
sub_df['num_orders'] = array
sub_df['num_orders_tree'] = arrayt
sub_df['avg'] = np.mean(array + arrayt, axis=0)
sub_df.to_csv('sub_tf4.csv', index=False)
if __name__ == '__main__':
train, test, center_info, meal_info = readData()
train_merged, test_merged = mergeData(train, test, center_info, meal_info)
dense_feat, lin_feat = makeFeatureColum(train_merged)
bucket_colum, indicator_column = makeFeaturesForTrees()
estimator = buildEstimator(dense_feat, lin_feat)
treestimator = buildTreeEstimator(bucket_colum + indicator_column)
train_inp_fn = estInpFunc()
test_inp_fn = estInpFunc(train=False)
#train
estimator = train_estimator(estimator, train_inp_fn)
treestimator = train_estimator(treestimator, train_inp_fn)
#test
arrayt = predict_using_trained_estimator(treestimator, test_inp_fn)
array = predict_using_trained_estimator(estimator, test_inp_fn)
"""
Project 1
At the end you should see something like this
Step Count:1000
Training accuracy: 0.8999999761581421 loss: 0.42281264066696167
Test accuracy: 0.8199999928474426 loss: 0.4739704430103302
play around with your model to try and get an even better score
"""
import tensorflow as tf
import dataUtils
training_data, training_labels = dataUtils.readData("project1trainingdata.csv")
test_data, test_labels = dataUtils.readData("project1testdata.csv")
# Build tensorflow blueprint
# Tensorflow placeholder
input_placeholder = tf.placeholder(tf.float32, shape=[None, 113])
# Neural network hidden layers
w1 = tf.get_variable("w1",
shape=[113, 150],
initializer=tf.contrib.layers.xavier_initializer())
b1 = tf.get_variable("12",
shape=[150],
initializer=tf.contrib.layers.xavier_initializer())
hidden_layer_1 = tf.nn.dropout(tf.layers.batch_normalization(
tf.nn.relu(tf.matmul(input_placeholder, w1) + b1),
axis=1,
center=True,