for imb, ymb in tqdm(iter_data(trX, trY, size=nbatch), total=ntrain / nbatch):
imb = transform(imb)
ymb = floatX(OneHot(ymb, ny))
zmb = floatX(np_rng.uniform(-1.0, 1.0, size=(len(imb), nz)))
if n_updates % (k + 1) == 0:
cost = _train_g(imb, zmb, ymb)
else:
cost = _train_d(imb, zmb, ymb)
n_updates += 1
n_examples += len(imb)
if (epoch - 1) % 5 == 0:
g_cost = float(cost[0])
d_cost = float(cost[1])
gX, gY = gen_samples(100000)
gX = gX.reshape(len(gX), -1)
va_nnc_acc_1k = nnc_score(gX[:1000], gY[:1000], vaX, vaY, metric="euclidean")
va_nnc_acc_10k = nnc_score(gX[:10000], gY[:10000], vaX, vaY, metric="euclidean")
va_nnc_acc_100k = nnc_score(gX[:100000], gY[:100000], vaX, vaY, metric="euclidean")
va_nnd_1k = nnd_score(gX[:1000], vaX, metric="euclidean")
va_nnd_10k = nnd_score(gX[:10000], vaX, metric="euclidean")
va_nnd_100k = nnd_score(gX[:100000], vaX, metric="euclidean")
log = [
n_epochs,
n_updates,
n_examples,
time() - t,
va_nnc_acc_1k,
va_nnc_acc_10k,
va_nnc_acc_100k,
va_nnd_1k,
va_nnd_10k,
def _nnc(inputs, labels, f=None):
assert len(inputs) == len(labels) == 2
if f is not None:
inputs = (_get_feats(f, x) for x in inputs)
(vaX, trX), (vaY, trY) = inputs, labels
return nnc_score(flat(trX), trY, flat(vaX), vaY, **kwargs)
ymb = floatX(OneHot(ymb, ny))
zmb = floatX(np_rng.uniform(-1., 1., size=(len(imb), nz)))
if n_updates % (k + 1) == 0:
cost = _train_g(imb, zmb, ymb)
else:
cost = _train_d(imb, zmb, ymb)
n_updates += 1
n_examples += len(imb)
if (epoch - 1) % 5 == 0:
g_cost = float(cost[0])
d_cost = float(cost[1])
gX, gY = gen_samples(100000)
gX = gX.reshape(len(gX), -1)
va_nnc_acc_1k = nnc_score(gX[:1000],
gY[:1000],
vaX,
vaY,
metric='euclidean')
va_nnc_acc_10k = nnc_score(gX[:10000],
gY[:10000],
vaX,
vaY,
metric='euclidean')
va_nnc_acc_100k = nnc_score(gX[:100000],
gY[:100000],
vaX,
vaY,
metric='euclidean')
va_nnd_1k = nnd_score(gX[:1000], vaX, metric='euclidean')
va_nnd_10k = nnd_score(gX[:10000], vaX, metric='euclidean')
va_nnd_100k = nnd_score(gX[:100000], vaX, metric='euclidean')