def run(sentset, labelset, postagset, all_feats, info, weights, testdata, ad):
tsents, tgoldtagseqs, tpostagseqs, tinfo = testdata
weights_avg = init_weights(all_feats)
order = [i for i in range(len(sentset))]
shuffle(order)
k = 0
for j in order:
sys.stderr.write(str(k)+"\r")
sent = sentset[j]
labelseq = labelset[j]
postagseq = postagset[j]
predseq = execute(sent, all_labels, postagseq, weights, labelseq, info)
if labelseq != predseq:
update(weights, predseq, labelseq, sent, postagseq, info, ad)
add_weights(weights_avg, weights)
k += 1
if k % 10000 == 0:
framework.write_weights(weights, k)
decode(tsents, tgoldtagseqs, tpostagseqs, tinfo, weights)
decode(tsents, tgoldtagseqs, tpostagseqs, tinfo, weights)
return weights_avg, weights
def learn_and_decode(trainfile, featlistfile, gazfile, brownfile, num_iter, testfile):
sentset, labelset, postagset, all_feats, info = framework.get_all(trainfile, gazfile, featlistfile, brownfile)
sys.stderr.write("\n" + str(len(all_feats)) + " features in all\n")
sys.stderr.write("\nreading test data \n")
tsents, tgoldtagseqs, tpostagseqs = framework.read_data(testfile)
tinfo = framework.get_maps(tsents, tpostagseqs, gazfile, brownfile)
testdata = (tsents, tgoldtagseqs, tpostagseqs, tinfo)
weights = init_weights(all_feats)
tot_weights = init_weights(all_feats)
#ADAGRAD
ad = init_weights(all_feats)
for ite in range(num_iter):
sys.stderr.write("Iteration " + str(ite) + "\n---------------------------\ntotal train sentences = "+ str(len(sentset)) + "\n")
weights_a, weights = run(sentset, labelset, postagset, all_feats, info, weights, testdata, ad) #ADAGRAD
framework.write_weights(weights, ite)
add_weights(tot_weights, weights_a)
for key in tot_weights.iterkeys():
tot_weights[key] /= num_iter*len(sentset)
sys.stderr.write("\n\nfinal performance on test\n")
decode(tsents, tgoldtagseqs, tpostagseqs, tinfo, tot_weights)
