def doc_vect(alldocs):
print 'Doc2Vec Each Tag is ID'
train_docs = [doc for doc in alldocs if doc.split == 'train']
test_docs = [doc for doc in alldocs if doc.split == 'test']
print('%d docs: %d train-sentiment, %d test-sentiment' % (len(alldocs), len(train_docs), len(test_docs)))
documents = []
for doc in train_docs:
sentence = TaggedDocument(doc.words, doc.tags)
documents.append(sentence)
print len(documents)
cores = multiprocessing.cpu_count()
simple_models = [
# PV-DM w/concatenation - window=5 (both sides) approximates paper's 10-word total window size
Doc2Vec(documents, dm=1, dm_concat=1, size=400, window=5, negative=5, hs=1, sample=1e-3, iter=20, min_count=1, workers=cores),
# PV-DBOW
Doc2Vec(documents, dm=0, size=400, window=5, negative=5, hs=1, sample=1e-3, iter=20, min_count=1, workers=cores),
# PV-DM w/average
Doc2Vec(documents, dm=1, dm_mean=1, size=400, window=5, negative=5, hs=1, sample=1e-3, iter=20, min_count=1, workers=cores),
]
models_by_name = OrderedDict((str(model), model) for model in simple_models)
models_by_name['dbow+dmm'] = ConcatenatedDoc2Vec([simple_models[1], simple_models[2]])
models_by_name['dbow+dmc'] = ConcatenatedDoc2Vec([simple_models[1], simple_models[0]])
for name, model in models_by_name.items():
print name
train_targets, train_regressors = zip(*[(doc.sentiment, model.docvecs[doc.tags[0]]) for doc in train_docs])
test_targets, test_regressors = zip(*[(doc.sentiment, model.infer_vector(doc.words)) for doc in test_docs])
util.logit(train_regressors, train_targets, test_regressors, test_targets)
util.svm(train_regressors, train_targets, test_regressors, test_targets)
def get_mu_sigma(self, X_noisy, t):
"""
Generate mu and sigma for one step in the reverse trajectory,
starting from a minibatch of images X_noisy, and at timestep t.
"""
Z = self.mlp.apply(X_noisy)
mu_coeff, beta_coeff = self.temporal_readout(Z, t)
# reverse variance is perturbation around forward variance
beta_forward = self.get_beta_forward(t)
# make impact of beta_coeff scaled appropriately with mu_coeff
beta_coeff_scaled = beta_coeff / np.sqrt(self.trajectory_length).astype(theano.config.floatX)
beta_reverse = T.nnet.sigmoid(beta_coeff_scaled + util.logit(beta_forward))
# # reverse mean is decay towards mu_coeff
# mu = (X_noisy - mu_coeff)*T.sqrt(1. - beta_reverse) + mu_coeff
# reverse mean is a perturbation around the mean under forward
# process
# # DEBUG -- use these lines to test objective is 0 for isotropic Gaussian model
# beta_reverse = beta_forward
# mu_coeff = mu_coeff*0
mu = X_noisy*T.sqrt(1. - beta_forward) + mu_coeff*T.sqrt(beta_forward)
sigma = T.sqrt(beta_reverse)
mu.name = 'mu p'
sigma.name = 'sigma p'
return mu, sigma
def classify(self, doc):
model_file_path = self.path + "svm_model"
model_file = '%ssvm_model' % self.path
if not self.feats: util.die('Incomplete model')
if not os.path.isfile(model_file):
util.die('no model [%s]' % model_file)
## testing data file
sys.stderr.write('SVM classifying... ')
lines = []
frag = doc.frag
while frag:
if frag.label == None: svm_label = '0'
elif frag.label: svm_label = '+1'
else: svm_label = '-1'
line = '%s ' % svm_label
feats = [f + '_' + v for f, v in frag.features.items()]
svm_feats = [self.feats[f] for f in feats if f in self.feats]
svm_feats.sort(lambda x, y: x - y)
line += ' '.join(['%d:1' % x for x in svm_feats])
lines.append(line)
frag = frag.next
unused, test_file = tempfile.mkstemp()
fh = open(test_file, 'w')
fh.write('\n'.join(lines) + '\n')
fh.close()
## classify test data
unused, pred_file = tempfile.mkstemp()
options = '-v 0'
#cmd = '%s %s %s %s %s' %(SVM_CLASSIFY, options, test_file, model_file, pred_file)
#cmd = SVM_CLASSIFY + " " +options + " " +test_file + " " + model_file + " " + pred_file
cmd = "\"" + SVM_CLASSIFY + "\" " + options + " \"" + test_file + "\" \"" + model_file_path + "\" \"" + pred_file + "\""
print cmd
os.system(cmd)
## get predictions
total = 0
preds = map(float, open(pred_file).read().splitlines())
frag = doc.frag
while frag:
frag.pred = util.logit(preds[total])
frag = frag.next
total += 1
## clean up
#os.remove(test_file)
#os.remove(pred_file)
sys.stderr.write('done!\n')
def classify(self, doc):
model_file_path = self.path + "svm_model"
model_file = '%ssvm_model' %self.path
if not self.feats: util.die('Incomplete model')
if not os.path.isfile(model_file): util.die('no model [%s]' %model_file)
## testing data file
sys.stderr.write('SVM classifying... ')
lines = []
frag = doc.frag
while frag:
if frag.label == None: svm_label = '0'
elif frag.label: svm_label = '+1'
else: svm_label = '-1'
line = '%s ' %svm_label
feats = [f+'_'+v for f,v in frag.features.items()]
svm_feats = [self.feats[f] for f in feats if f in self.feats]
svm_feats.sort(lambda x,y: x-y)
line += ' '.join(['%d:1' %x for x in svm_feats])
lines.append(line)
frag = frag.next
unused, test_file = tempfile.mkstemp()
fh = open(test_file, 'w')
fh.write('\n'.join(lines) + '\n')
fh.close()
## classify test data
unused, pred_file = tempfile.mkstemp()
options = '-v 0'
#cmd = '%s %s %s %s %s' %(SVM_CLASSIFY, options, test_file, model_file, pred_file)
#cmd = SVM_CLASSIFY + " " +options + " " +test_file + " " + model_file + " " + pred_file
cmd = "\""+SVM_CLASSIFY + "\" " +options + " \"" +test_file + "\" \"" + model_file_path + "\" \"" + pred_file+"\""
print cmd
os.system(cmd)
## get predictions
total = 0
preds = map(float, open(pred_file).read().splitlines())
frag = doc.frag
while frag:
frag.pred = util.logit(preds[total])
frag = frag.next
total += 1
## clean up
#os.remove(test_file)
#os.remove(pred_file)
sys.stderr.write('done!\n')
def spec_loss(self, y_hat, y, mask, priority_bin=None, priority_w=0):
masked_l1 = MaskedL1Loss()
l1 = nn.L1Loss()
w = self.hparams.masked_loss_weight
# L1 loss
if w > 0:
assert mask is not None
l1_loss = w * masked_l1(y_hat, y, mask=mask) + (1 - w) * l1(
y_hat, y)
else:
assert mask is None
l1_loss = l1(y_hat, y)
# Priority L1 loss
if priority_bin is not None and priority_w > 0:
if w > 0:
priority_loss = w * masked_l1(
y_hat[:, :, :priority_bin], y[:, :, :priority_bin], mask=mask) \
+ (1 - w) * l1(y_hat[:, :, :priority_bin], y[:, :, :priority_bin])
else:
priority_loss = l1(y_hat[:, :, :priority_bin],
y[:, :, :priority_bin])
l1_loss = (1 - priority_w) * l1_loss + priority_w * priority_loss
# Binary divergence loss
if self.w <= 0:
binary_div = y.data.new(1).zero_()
else:
y_hat_logits = logit(y_hat)
z = -y * y_hat_logits + torch.log1p(torch.exp(y_hat_logits))
if w > 0:
binary_div = w * masked_mean(z, mask) + (1 - w) * z.mean()
else:
binary_div = z.mean()
return l1_loss, binary_div
def cross_validation(X_train, y_train, params, X_test=None, verbose_eval=False):
NUM_BOOST_ROUND = 1000
best_iterations = []
train_scores = []
valid_scores = []
y_preds = []
kf = KFold(y_train.shape[0], n_folds=5, shuffle=True, random_state=12345)
for train_index, valid_index in kf:
_X_train, _X_valid = X_train.ix[train_index], X_train.ix[valid_index]
_y_train, _y_valid = y_train[train_index], y_train[valid_index]
dtrain = xgb.DMatrix(_X_train, _y_train)
dvalid = xgb.DMatrix(_X_valid, _y_valid)
watchlist = [(dtrain, 'train'), (dvalid, 'eval')]
bst = xgb.train(params, dtrain, NUM_BOOST_ROUND, evals=watchlist,
early_stopping_rounds=200, verbose_eval=verbose_eval)
# best iterations and valid score
best_iterations.append(bst.best_iteration + 1)
valid_scores.append(bst.best_score)
if X_test is not None:
dtest = xgb.DMatrix(X_test)
y_pred = bst.predict(dtest, ntree_limit=bst.best_iteration)
y_preds.append(y_pred)
y_pred = util.sigmoid(np.mean(util.logit(np.array(y_preds)), axis=0))
result = {"best-iterations": best_iterations,
"best-iteration": np.mean(best_iterations),
"valid-score": np.mean(valid_scores),
"valid-scores": valid_scores,
"y_pred": y_pred,
"y_preds": y_preds}
return result
def label_vect(alldocs):
print 'Label2Vec with pre-classification'
train_docs = [doc for doc in alldocs if doc.split == 'train']
test_docs = [doc for doc in alldocs if doc.split == 'test']
non_docs = [doc for doc in alldocs if doc.split == 'extra']
print('%d docs: %d train-sentiment, %d test-sentiment' %
(len(alldocs), len(train_docs), len(test_docs)))
ylin = pre_class(train_docs, test_docs, non_docs)
documents = []
for doc in train_docs:
sentence = TaggedDocument(doc.words, [str(doc.sentiment)])
documents.append(sentence)
i = 0
for doc in test_docs + non_docs:
sentence = TaggedDocument(doc.words, [str(ylin[i])])
documents.append(sentence)
i += 1
print len(documents)
cores = multiprocessing.cpu_count()
simple_models = [
# PV-DM w/concatenation - window=5 (both sides) approximates paper's 10-word total window size
Doc2Vec(documents,
dm=1,
dm_concat=1,
size=100,
window=10,
negative=5,
hs=1,
sample=1e-3,
min_count=1,
workers=cores),
# PV-DBOW
Doc2Vec(documents,
dm=0,
size=100,
window=10,
negative=5,
hs=1,
sample=1e-3,
min_count=1,
workers=cores),
# PV-DM w/average
Doc2Vec(documents,
dm=1,
dm_mean=1,
size=100,
window=10,
negative=5,
hs=1,
sample=1e-3,
min_count=1,
workers=cores),
]
models_by_name = OrderedDict(
(str(model), model) for model in simple_models)
models_by_name['dbow+dmm'] = ConcatenatedDoc2Vec(
[simple_models[1], simple_models[2]])
models_by_name['dbow+dmc'] = ConcatenatedDoc2Vec(
[simple_models[1], simple_models[0]])
for name, model in models_by_name.items():
print name
train_targets, train_regressors = zip(*[(doc.sentiment,
model.infer_vector(doc.words))
for doc in train_docs])
test_targets, test_regressors = zip(*[(doc.sentiment,
model.infer_vector(doc.words))
for doc in test_docs])
util.logit(train_regressors, train_targets, test_regressors,
test_targets)
def _logit_transform(x):
"""
Transforms pixel values with logit to be unconstrained.
"""
return util.logit(CIFAR10.alpha + (1 - 2 * CIFAR10.alpha) * x)
def inv(i, max):
sc = (i / max) * 0.999 + 0.0005
return logit(sc)
