def get_trained_model(c, corpus, dictionary, non_zero_indices):
model = LogisticRegression(penalty='l1', tol=0.0001, C=c, fit_intercept=True, \
intercept_scaling=1, solver='liblinear', max_iter=args.epochs, multi_class='ovr', verbose=1)
train_sentences1, train_labels = helper.batch_to_one_hot_encoded(
corpus.data, dictionary, non_zero_indices=non_zero_indices)
model.fit(train_sentences1, train_labels)
return model
def eval_routine(corpus, dictionary, model, non_zero_indices=None):
print('one hot encoding...')
eval_sentences, eval_labels = helper.batch_to_one_hot_encoded(corpus.data, dictionary, non_zero_indices = non_zero_indices)
if not non_zero_indices:
print('Creating CSR sparsing...')
eval_sentences = csr_matrix(eval_sentences)
print('Testing...')
acc = model.score(eval_sentences, eval_labels)
print(' Accurcay: ', acc)
return acc
def get_trained_model2(c, corpus, dictionary, non_zero_indices):
model = LogisticRegression(penalty='l1', tol=0.0001, C=c, fit_intercept=True, \
intercept_scaling=1, solver='liblinear', max_iter=args.epochs, multi_class='ovr', verbose=0)
train_batches = helper.batchify(corpus.data, args.batch_size)
print('number of train batches = ', len(train_batches))
num_batches = len(train_batches)
n_correct, n_total = 0, 0
for batch_no in range(1, num_batches + 1):
if batch_no % 500 == 0:
print(' training batch: ', batch_no, ' of ', num_batches,
' percentage: ', batch_no / num_batches)
train_sentences1, train_labels = helper.batch_to_one_hot_encoded(
train_batches[batch_no - 1],
dictionary,
non_zero_indices=non_zero_indices)
model.fit(train_sentences1, train_labels)
return model
def eval_routine(corpus, dictionary, model, non_zero_indices=None):
nexamples = len(corpus.data)
dev_batches = helper.batchify(corpus.data, args.batch_size)
print('number of train batches = ', len(dev_batches))
total_acc = 0.0
correct = 0.0
num_batches = len(dev_batches)
n_correct, n_total = 0, 0
for batch_no in range(1, num_batches + 1):
if batch_no % 500 == 0:
print(' validation batch: ', batch_no, ' of ', num_batches,
' percentage: ', batch_no / num_batches)
eval_sentences, eval_labels = helper.batch_to_one_hot_encoded(
dev_batches[batch_no - 1],
dictionary,
non_zero_indices=non_zero_indices)
acc = model.score(eval_sentences, eval_labels)
correct += acc * len(eval_labels)
total_acc += acc
# if batch_no%500 == 0 :print(' for this minibatch score: ', acc, ' correct: ', acc*len(eval_labels), ' of ', len(eval_labels), 'total accc: ', total_acc, ' total correct: ', correct)
print(' Correct: ', correct, ' acc: ', correct / nexamples,
' sanity check: ', total_acc / num_batches)
return correct / nexamples
# save the model to disk
filename = args.task+'_L1_model.pcl'
best_acc = 0
i=0
bc = -1
# cs = 0.02, and 0.2 for sst for BCN and LSTM (or the otehr sequence)
cs = [] # for IMDB
print('one hot encoding...')
train_sentences1, train_labels = helper.batch_to_one_hot_encoded(train_corpus.data, dictionary, non_zero_indices = [])
print('Training...')
# import pdb
# pdb.set_trace()
# for c in cs:
while (i<1):
# c = numpy.random.uniform(args.c-5, args.c+5)
c = args.c#2.5#numpy.random.uniform(0.7785, 0.7785)
print('training model with c: ', c, ' in iter: ', i+1, "/", len(cs))
i+=1
model = get_trained_model(c, train_sentences1, train_labels, dictionary, non_zero_indices=[])
# print("==="*20, "\nC: ", c, "\n", "=="*20)