def get_pc(data, We, weight4ind, params):
"Comput the principal component"
def get_weighted_average(We, x, w):
"Compute the weighted average vectors"
n_samples = x.shape[0]
emb = np.zeros((n_samples, We.shape[1]))
for i in range(n_samples):
emb[i, :] = w[i, :].dot(We[x[i, :], :]) / np.count_nonzero(w[i, :])
return emb
for i in data:
i[0].populate_embeddings(words)
if not params.task == "sentiment":
i[1].populate_embeddings(words)
if params.task == "ent":
(scores, g1x, g1mask, g2x, g2mask) = data_io.getDataEntailment(data)
if params.weightfile:
g1mask = data_io.seq2weight(g1x, g1mask, weight4ind)
elif params.task == "sim":
(scores, g1x, g1mask, g2x, g2mask) = data_io.getDataSim(data, -1)
if params.weightfile:
g1mask = data_io.seq2weight(g1x, g1mask, weight4ind)
elif params.task == "sentiment":
(scores, g1x, g1mask) = data_io.getDataSentiment(data)
if params.weightfile:
g1mask = data_io.seq2weight(g1x, g1mask, weight4ind)
emb = get_weighted_average(We, g1x, g1mask)
svd = TruncatedSVD(n_components=params.npc, n_iter=7, random_state=0)
svd.fit(emb)
return svd.components_
def train_util(model, train_data, dev, test, train, words, params):
"utility function for training the model"
start_time = time()
try:
for eidx in range(params.epochs):
kf = data_io.get_minibatches_idx(len(train_data),
params.batchsize,
shuffle=True)
uidx = 0
for _, train_index in kf:
uidx += 1
batch = [train_data[t] for t in train_index]
# load the word ids
for i in batch:
i[0].populate_embeddings(words)
if not params.task == "sentiment":
i[1].populate_embeddings(words)
# load the data
if params.task == "ent":
(scores, g1x, g1mask, g2x,
g2mask) = data_io.getDataEntailment(batch)
elif params.task == "sim":
(scores, g1x, g1mask, g2x,
g2mask) = data_io.getDataSim(batch, model.nout)
elif params.task == "sentiment":
(scores, g1x, g1mask) = data_io.getDataSentiment(batch)
else:
raise ValueError('Task should be ent or sim.')
# train
if not params.task == "sentiment":
if params.weightfile:
g1mask = data_io.seq2weight(g1x, g1mask,
params.weight4ind)
g2mask = data_io.seq2weight(g2x, g2mask,
params.weight4ind)
cost = model.train_function(scores, g1x, g2x, g1mask,
g2mask)
else:
if params.weightfile:
g1mask = data_io.seq2weight(g1x, g1mask,
params.weight4ind)
cost = model.train_function(scores, g1x, g1mask)
if np.isnan(cost) or np.isinf(cost):
print('NaN detected')
# undo batch to save RAM
for i in batch:
i[0].representation = None
i[0].unpopulate_embeddings()
if not params.task == "sentiment":
i[1].representation = None
i[1].unpopulate_embeddings()
# evaluate
if params.task == "sim":
dp, ds = eval.supervised_evaluate(model, words, dev, params)
tp, ts = eval.supervised_evaluate(model, words, test, params)
rp, rs = eval.supervised_evaluate(model, words, train, params)
print(("evaluation: ", dp, ds, tp, ts, rp, rs))
elif params.task == "ent" or params.task == "sentiment":
ds = eval.supervised_evaluate(model, words, dev, params)
ts = eval.supervised_evaluate(model, words, test, params)
rs = eval.supervised_evaluate(model, words, train, params)
print(("evaluation: ", ds, ts, rs))
else:
raise ValueError('Task should be ent or sim.')
print(('Epoch ', (eidx + 1), 'Cost ', cost))
sys.stdout.flush()
except KeyboardInterrupt:
print("Training interupted")
end_time = time()
print(("total time:", (end_time - start_time)))