def cmd_hack(args):
info('hack called with args %s' % str(args))
warn('this is a warning')
err('this is an error')
info('this is merely info')
print getTags()
a, b = u.split([i for i in range(20)], 17)
print a
print b
fail('now you\'ve done it')
def cmd_train(args):
if len(args) == 0:
err('error: not enough arguments.')
err('usage: run <classify-tag>')
sys.exit(1)
tag = args[0]
tags = u.getTags()
if tag not in tags:
fail('error: tag "%s" not found.\nKnown tags are: %s' %
(tag, str(tags)))
info('classifying for tag "%s"' % tag)
imageFiles, imageData, labels, imageJson = u.loadInputs()
# see multi-label learning comment below. for now, slice() to one label
labels = u.slice(labels, tags.index(tag))
NO_YES = ['NO ' + tag, 'YES ' + tag]
verbose('data shape: %s labels shape: %s' %
(imageData.shape, labels.shape))
trainedFiles, testFiles = u.split(imageFiles, 326)
trainedImages, testImages = u.split(imageData, 326)
trainedLabels, testLabels = u.split(labels, 326)
# some classifiers can only handle a single target
# http://stackoverflow.com/questions/31881138/predicting-outcome-of-multiple-targets-in-scikit-learn
# try this? http://scikit-learn.org/stable/modules/generated/sklearn.multioutput.MultiOutputClassifier.html
verbose('trained data shape: %s trained labels shape: %s' %
(trainedImages.shape, trainedLabels.shape))
classifier = GaussianNB(
) # 92.6% correct on training data set, wrong between white <-> gray
#classifier = LogisticRegression() # 100% correct on training, empirically worse on test data tho
#classifier = KNeighborsClassifier() # not as good as GaussianNB, particularly false positives/negatives on blue
#classifier = ElasticNet()
classifier.fit(trainedImages, trainedLabels)
predicts = classifier.predict(testImages)
verbose('predicts.shape: %s' % str(predicts.shape))
#print predicts
probs = classifier.predict_proba(testImages)
verbose('probs shape: %s' % str(probs.shape))
verbose('testLabels: %s' % str(testLabels))
htmlFile = tag + '-gaussian-unclassified.html'
u.outputHtml(htmlFile, testFiles, [NO_YES[int(p)] for p in predicts],
[NO_YES[int(i)] for i in testLabels], None)
info('saved test results: %s' % htmlFile)
def cmd_hack(args):
tset = u.loadInputs2(True, '28x28')
print len(tset.images), 'image(s)'
print tset.images[0]
print tset.images[0].np_data.dtype
sys.exit(1)
info('hack called with args %s' % str(args))
warn('this is a warning')
err('this is an error')
info('this is merely info')
a, b = u.split([i for i in range(20)], 17)
print a
print b
nda = np.array([0, 0, 1])
print(nda)
print(n2c(nda))
fail('now you\'ve done it')
def cmd_run(args):
if len(args) == 0:
err('error: not enough arguments.')
err('usage: run <classify-tag>')
sys.exit(1)
tag = args[0]
tags = u.getTags()
if tag not in tags:
fail('error: tag "%s" not found.\nKnown tags are: %s' %
(tag, str(tags)))
info('classifying for tag "%s"' % tag)
imageFiles, imageData, labels, imgJson = u.loadInputs()
# see multi-label learning comment below. for now, slice() to one label
labels = u.slice(labels, tags.index(tag))
NO_YES = ['NO ' + tag, 'YES ' + tag]
verbose('data shape: %s labels shape: %s' %
(imageData.shape, labels.shape))
trainedFiles, testFiles = u.split(imageFiles, 326)
trainedImages, testImages = u.split(imageData, 326)
trainedLabels, testLabels = u.split(labels, 326)
# some classifiers can only handle a single target
# http://stackoverflow.com/questions/31881138/predicting-outcome-of-multiple-targets-in-scikit-learn
# try this? http://scikit-learn.org/stable/modules/generated/sklearn.multioutput.MultiOutputClassifier.html
verbose('trained data shape: %s trained labels shape: %s' %
(trainedImages.shape, trainedLabels.shape))
model = Sequential()
model.add(Dense(None, input_dim=14400))
model.add(Activation('relu'))
verbose('model summary:')
model.summary()
info('model built, compiling')
model.compile(optimizer='rmsprop',
loss='binary_crossentropy',
metrics=['accuracy'])
# Convert labels to categorical one-hot encoding
# think I've done the equivalent of this in the slice() above (?)
#binaryTrainedLabels = keras.utils.to_categorical(labels, num_classes=10)
model.fit(trainedImages, trainedLabels, epochs=10, batch_size=326)
classifier = GaussianNB(
) # 92.6% correct on training data set, wrong between white <-> gray
#classifier = LogisticRegression() # 100% correct on training, empirically worse on test data tho
#classifier = ElasticNet()
classifier.fit(trainedImages, trainedLabels)
predicts = classifier.predict(testImages)
verbose('predicts.shape: %s' % str(predicts.shape))
#print predicts
probs = classifier.predict_proba(testImages)
verbose('probs shape: %s' % str(probs.shape))
verbose('testLabels: %s' % str(testLabels))
htmlFile = tag + '-unclassified.html'
u.outputHtml(htmlFile, testFiles, [NO_YES[int(p)] for p in predicts],
[NO_YES[int(i)] for i in testLabels], None)
info('saved test results: %s' % htmlFile)
predicts = classifier.predict(trainedImages)
htmlFile = tag + '-classified.html'
u.outputHtml(htmlFile, trainedFiles, [NO_YES[int(p)] for p in predicts],
[NO_YES[int(i)] for i in trainedLabels], None)
info('saved test results: %s' % htmlFile)