def grid_search():
"""
Do a grid search over lambda for the optimal learning rate.
"""
lambdas = [1e2, 1e3, 1e4, 1e5]
# load data
train_labels_og = "./dataset/trainingLabels.txt"
train_sentences_og = "./dataset/trainingSentences.txt"
train_labels, train_sentences, validation_labels, validation_sentences = sr.shuffle_examples(
train_labels_og, train_sentences_og, pct_train=0.5
)
gridfile = open("gridsearch_results.txt", "w")
# run over values of lambda
for lr in lambdas:
print lr
weights, scores, epoch, ept_avg = SGD_train(
train_labels[:100],
train_sentences[:100],
20,
validation_labels[:100],
validation_sentences[:100],
"word",
lr,
)
print scores
gridfile.write(str(scores[-2]) + "\n")
gridfile.close()
import SGD_CRF
import subroutines as sr
import dataproc as dp
import numpy as np
# first load and shuffle the data
train_labels_og = './dataset/trainingLabels.txt'
train_sentences_og = './dataset/trainingSentences.txt'
train_labels, train_sentences, validation_labels, validation_sentences = sr.shuffle_examples(train_labels_og, train_sentences_og, pct_train=0.5)
weights, scores, epoch, ept_avg = SGD_CRF.SGD_train(train_labels, train_sentences, 20,
validation_labels, validation_sentences, 'word')
np.save('sgd_w_0p5.npy', np.array(weights))
np.save('sgd_s_0p5.npy', np.array(scores))
np.save('sgd_ep_0p5.npy', np.array([epoch,ept_avg]))