def test_softmax_normal(self):
b = Softmax()
d = np.random.random((5, 3))
pred = b.predict(d)
predsums = np.sum(pred, axis=1)
self.assertTrue(np.allclose(predsums, np.ones_like(predsums)))
self.assertEqual(d.shape, pred.shape)
def test_softmax_normal_masked(self):
b = Softmax()
d = np.random.random((5, 3))
m = np.ones_like(d)
m[:, 2] = 0
pred = b.predict(d, m)
self.assertTrue(np.allclose(np.zeros_like(pred[:, 2]), pred[:, 2]))
self.assertEqual(d.shape, pred.shape)
predsums = np.sum(pred, axis=-1)
self.assertTrue(np.allclose(predsums, np.ones_like(predsums)))
def test_masked_3D_gumbel_softmax(self):
d = np.random.random((5, 4, 3))
m = np.ones_like(d)
m[:, :, 2] = 0
sm = Softmax()
gsm = GumbelSoftmax(temperature=0.3)
smpred = sm.predict(d)
gsmpred = gsm.predict(smpred, m)
np.set_printoptions(precision=5, suppress=True)
print gsmpred
predsums = np.sum(gsmpred, axis=-1)
self.assertTrue(np.allclose(predsums, np.ones_like(predsums)))
self.assertEqual(d.shape, gsmpred.shape)
self.assertTrue(
np.allclose(np.zeros_like(gsmpred[:, :, 2]), gsmpred[:, :, 2]))
def test_gumbel_softmax(self):
d = np.random.random((5, 3))
d[0, 0] = 2.
sm = Softmax()
gsm = GumbelSoftmax(temperature=1e-10, _alwaysrandom=True)
smpred = sm.predict(d)
samples = []
gsmpredf = gsm.predict
for i in range(100000):
gsmpred = gsmpredf(smpred)
samples.append(gsmpred[0, :])
samples = np.concatenate([sample[:, np.newaxis] for sample in samples],
axis=1)
sampleavg = np.average(samples, axis=1)
print sampleavg
np.set_printoptions(precision=5, suppress=True)
print smpred
print gsmpred
predsums = np.sum(gsmpred, axis=-1)
self.assertTrue(np.allclose(predsums, np.ones_like(predsums)))
self.assertEqual(d.shape, gsmpred.shape)
self.assertTrue(np.allclose(smpred[0, :], sampleavg, rtol=1e-1))
def test_softmax_normal_with_temperature(self):
b = Softmax(temperature=1e-6)
d = np.random.random((5, 3))
pred = b.predict(d)
print pred