def test_softmax_gradient(samples, features, classes, check_count=20):
dummy_weights = 0.1 * np.random.randn(features, classes) + 1.0
dummy_features = np.random.randn(samples, features)
dummy_labels = np.argmax(np.random.randn(samples, classes), axis=1)
#rough estimate of how many should be correct (assuming random: -np.log(p), where p is the probability of guessing correctly)
#assert np.abs(-np.log(1./classes) - (softmaxRegression(dummy_features, dummy_labels, dummy_weights, 0.0)[0])) / -np.log(1./classes) <= 0.15
f = lambda w: softmaxRegression(dummy_features,
dummy_labels,
w,
regularization=0.0,
nopredictions=False)[0]
g = lambda w: softmaxRegression(dummy_features,
dummy_labels,
w,
regularization=0.0,
nopredictions=False)[1]
W = dummy_weights
grad_analytic = g(W)
for i in range(check_count):
ix = tuple([random.randrange(m) for m in W.shape])
shift = np.zeros(W.shape)
shift[ix] = 1e-7
grad_numerical = (f(W + shift) - f(W - shift)) / (2 * 1e-7)
assert (abs(grad_numerical - grad_analytic[ix]) /
(abs(grad_numerical) + abs(grad_analytic[ix])) < 0.0002)
def test_softmax_regression(samples, features, classes):
dummy_weights = 0.1 * np.random.randn(features, classes) + 1.0
dummy_features = np.random.randn(samples, features)
dummy_labels = np.argmax(np.random.randn(samples, classes), axis=1)
#np.sqrt(p * (1 - p) / N )
#n/N
#rough estimate of how many should be correct (assuming random: -np.log(p), where p is the probability of guessing correctly)
assert np.abs(-np.log(1./classes) - (softmaxRegression(dummy_features, dummy_labels, dummy_weights, 0.0)[0])) / -np.log(1./classes) <= 0.15
def test_softmax_gradient(samples, features, classes, check_count=20):
dummy_weights = 0.1 * np.random.randn(features, classes) + 1.0
dummy_features = np.random.randn(samples, features)
dummy_labels = np.argmax(np.random.randn(samples, classes), axis=1)
#rough estimate of how many should be correct (assuming random: -np.log(p), where p is the probability of guessing correctly)
#assert np.abs(-np.log(1./classes) - (softmaxRegression(dummy_features, dummy_labels, dummy_weights, 0.0)[0])) / -np.log(1./classes) <= 0.15
f = lambda w: softmaxRegression(dummy_features, dummy_labels, w, regularization=0.0, nopredictions=False)[0]
g = lambda w: softmaxRegression(dummy_features, dummy_labels, w, regularization=0.0, nopredictions=False)[1]
W = dummy_weights
grad_analytic = g(W)
for i in range(check_count):
ix = tuple([random.randrange(m) for m in W.shape])
shift = np.zeros(W.shape)
shift[ix] = 1e-7
grad_numerical = (f(W + shift) - f(W - shift)) / (2 * 1e-7)
assert( abs(grad_numerical - grad_analytic[ix]) / (abs(grad_numerical) + abs(grad_analytic[ix])) < 0.0002)
def test_softmax_regression(samples, features, classes):
dummy_weights = 0.1 * np.random.randn(features, classes) + 1.0
dummy_features = np.random.randn(samples, features)
dummy_labels = np.argmax(np.random.randn(samples, classes), axis=1)
#np.sqrt(p * (1 - p) / N )
#n/N
#rough estimate of how many should be correct (assuming random: -np.log(p), where p is the probability of guessing correctly)
assert np.abs(-np.log(1. / classes) - (
softmaxRegression(dummy_features, dummy_labels, dummy_weights, 0.0)[0])
) / -np.log(1. / classes) <= 0.15
def getSentiment():
dataSet = StanfordSentiment()
tokens = dataSet.tokens()
# Load the word vectors we trained earlier
_, wordVectors, _ = load_saved_params()
dimVectors = wordVectors.shape[1]
# Load the test sentences
sentences = dataSet.getTestSentences()
print sentences
nSentences = len(sentences)
sentenceFeatures = np.zeros((nSentences, dimVectors))
sentenceLabels = np.zeros((nSentences, ), dtype=np.int32)
for i in xrange(nSentences):
words, sentenceLabels[i] = sentences[i]
#print sentences[i]
#print words
#print sentenceLabels[i]
sentenceFeatures[i, :] = getSentenceFeature(tokens, wordVectors, words)
#train weights
random.seed(3141)
np.random.seed(59265)
weights = np.random.randn(dimVectors, nSentences)
regularization = 0.00001
weights = sgd(lambda weights: softmax_wrapper(
sentenceFeatures, sentenceLabels, weights, regularization),
weights,
3.0,
10,
PRINT_EVERY=10)
#pred = np.sum((weights * sentenceFeatures.T).T, axis = 1)
#pred = softmax(pred)
#testAccuracy = accuracy(sentenceLabels, pred)
prob = softmax(sentenceFeatures.dot(weights))
#prob = normalizeRows(prob)
_, _, pred = softmaxRegression(sentenceFeatures, sentenceLabels, weights)
pred = 1 / (1 + np.exp(-pred))
#for polarity in pred:
#print categorify(polarity)
#_, _, pred = softmaxRegression(sentenceFeatures, sentenceLabels, weights)
for label in pred:
print categorify(label)
#pred = categorify(pred)
print prob
print pred
for regularization in REGULARIZATION:
random.seed(3141)
np.random.seed(59265)
weights = np.random.randn(dimVectors, 5)
print "Training for reg=%f" % regularization
# We will do batch optimization
weights = sgd(lambda weights: softmax_wrapper(trainFeatures, trainLabels,
weights, regularization),
weights,
3.0,
10000,
PRINT_EVERY=100)
# Test on train set
_, _, pred = softmaxRegression(trainFeatures, trainLabels, weights)
trainAccuracy = accuracy(trainLabels, pred)
print "Train accuracy (%%): %f" % trainAccuracy
# Test on dev set
_, _, pred = softmaxRegression(devFeatures, devLabels, weights)
devAccuracy = accuracy(devLabels, pred)
print "Dev accuracy (%%): %f" % devAccuracy
# Save the results and weights
results.append({
"reg": regularization,
"weights": weights,
"train": trainAccuracy,
"dev": devAccuracy
})
devFeatures[i, :] = getSentenceFeature(tokens, wordVectors, words)
# Try our regularization parameters
results = []
for regularization in REGULARIZATION:
random.seed(3141)
np.random.seed(59265)
weights = np.random.randn(dimVectors, 5)
print "Training for reg=%f" % regularization
# We will do batch optimization
weights = sgd(lambda weights: softmax_wrapper(trainFeatures, trainLabels,
weights, regularization), weights, 3.0, 10000, PRINT_EVERY=100)
# Test on train set
_, _, pred = softmaxRegression(trainFeatures, trainLabels, weights)
trainAccuracy = accuracy(trainLabels, pred)
print "Train accuracy (%%): %f" % trainAccuracy
# Test on dev set
_, _, pred = softmaxRegression(devFeatures, devLabels, weights)
devAccuracy = accuracy(devLabels, pred)
print "Dev accuracy (%%): %f" % devAccuracy
# Save the results and weights
results.append({
"reg" : regularization,
"weights" : weights,
"train" : trainAccuracy,
"dev" : devAccuracy})
devFeatures[i, :] = getSentenceFeature(tokens, wordVectors, words)
# Try our regularization parameters
results = []
for regularization in REGULARIZATION:
random.seed(3141)
np.random.seed(59265)
weights = np.random.randn(dimVectors, 5)
print "Training for reg=%f" % regularization
# We will do batch optimization
weights = sgd(lambda weights: softmax_wrapper(trainFeatures, trainLabels,
weights, regularization), weights, 3.0, 10000, PRINT_EVERY=100)
# Test on train set
_, _, pred = softmaxRegression(trainFeatures, trainLabels, weights)
trainAccuracy = accuracy(trainLabels, pred)
print "Train accuracy (%%): %f" % trainAccuracy
# Test on dev set
_, _, pred = softmaxRegression(devFeatures, devLabels, weights)
devAccuracy = accuracy(devLabels, pred)
print "Dev accuracy (%%): %f" % devAccuracy
# Save the results and weights
results.append({
"reg" : regularization,
"weights" : weights,
"train" : trainAccuracy,
"dev" : devAccuracy})
for regularization in REGULARIZATION:
random.seed(3141)
np.random.seed(59265)
weights = np.random.randn(dimVectors, 5)
print "Training for reg=%f" % regularization
# We will do batch optimization
weights = sgd(lambda weights: softmax_wrapper(trainFeatures, trainLabels,
weights, regularization),
weights,
3.0,
10000,
PRINT_EVERY=100)
# Test on train set
_, _, pred = softmaxRegression(trainFeatures, trainLabels, weights,
regularization)
trainAccuracy = accuracy(trainLabels, pred)
print "Train accuracy (%%): %f" % trainAccuracy
# Test on dev set
_, _, pred = softmaxRegression(devFeatures, devLabels, weights,
regularization)
devAccuracy = accuracy(devLabels, pred)
print "Dev accuracy (%%): %f" % devAccuracy
# Save the results and weights
results.append({
"reg": regularization,
"weights": weights,
"train": trainAccuracy,
"dev": devAccuracy
devFeatures = np.zeros((nDev, dimVectors))
devLabels = np.zeros((nDev,), dtype=np.int32)
for i in xrange(nDev):
words, devLabels[i] = devset[i]
devFeatures[i, :] = getSentenceFeature(tokens, wordVectors, words)
# Try our regularization parameters
results = []
for regularization in REGULARIZATION:
random.seed(3141)
np.random.seed(59265)
weights = np.random.randn(dimVectors, 5)
print "Training for reg=%f" % regularization
# We will do batch optimization
weights = sgd(lambda weights: softmaxRegression(trainFeatures, trainLabels,
weights, regularization, nopredictions=True), weights, 0.3, 200000, PRINT_EVERY=20000)
# Test on train set
_, _, pred = softmaxRegression(trainFeatures, trainLabels, weights)
trainAccuracy = accuracy(trainLabels, pred)
print "Train accuracy (%%): %f" % trainAccuracy
# Test on dev set
_, _, pred = softmaxRegression(devFeatures, devLabels, weights)
devAccuracy = accuracy(devLabels, pred)
print "Dev accuracy (%%): %f" % devAccuracy
print "Prediction VS ||Weights||^2: %f - %f" % \
(predictionLoss(trainFeatures, trainLabels, weights),
regularizationLoss(weights, 1.0))
# Save the results and weights
