def test_case():
w_val = np_utils.random_vec((in_features, out_features))
w = torch.tensor(w_val, requires_grad=True).double()
b_val = np_utils.random_vec((out_features, )).reshape(out_features)
b = torch.tensor(b_val, requires_grad=True).double()
x_val = np_utils.random_vec((batch_size, in_features))
x = torch.tensor(x_val, requires_grad=True).double()
torch.autograd.gradcheck(linear, (x, w, b))
def sample():
a = torch.Tensor(np_utils.random_vec((batch_size, emb_size)))
b = torch.Tensor(np_utils.random_vec((batch_size, emb_size)))
c = np_utils.random_vec([], low=0.0, high=1.)
c = float(c)
logger.debug("c: {}".format(c))
res = mobius.add(-a, mobius.add(a, b, c), c).data.numpy()
expected = b.data.numpy()
assert res.shape == expected.shape
assert np.allclose(res, expected)
def test_case():
p_val = np_utils.random_vec((K, emb_size))
a_val = np_utils.random_vec((K, emb_size))
cval = 1.
#c = torch.Tensor(cval).double()
c = cval
p = torch.tensor(p_val, requires_grad=True).double()
a = torch.tensor(a_val, requires_grad=True).double()
x_val = np_utils.random_vec((batch_size, emb_size))
x = torch.tensor(x_val, requires_grad=True).double()
torch.autograd.gradcheck(mobius.logits, (x, p, a, c))
def sample():
aval = np_utils.random_vec((batch_size, emb_size), low=-0.01)
bval = np_utils.random_vec((batch_size, emb_size), low=-0.01)
a = torch.Tensor(aval)
b = torch.Tensor(bval)
#cval = np_utils.random_vec((1, ), low=0.0, high=1.)
# TODO: Why does this fail with random c?
cval = 1.
c = cval
logger.info("c: {}".format(c))
res = mobius.squared_distance(a, b, c).data.numpy()
expected = np_utils.squared_distance(aval, bval, cval)
assert res.shape == expected.shape
assert np.allclose(res, expected)
def test_logits():
batch_size = 4
emb_size = 5
K = 3
p_val = np_utils.random_vec((K, emb_size))
a_val = np_utils.random_vec((K, emb_size))
cval = 1.
#c = torch.Tensor(cval).double()
c = cval
p = torch.tensor(p_val).double()
a = torch.tensor(a_val).double()
x_val = np_utils.random_vec((batch_size, emb_size))
x = torch.tensor(x_val).double()
logits = mobius.logits(x, p, a, c)
assert tuple(logits.shape) == (batch_size, K)
def test_case():
#c_val = np_utils.random_vec((1, ), low=0.5, high=1.)
c_val = 1.
logger.debug("c={}".format(c_val))
c = c_val
x = torch.Tensor(np_utils.random_vec((batch_size, in_features)))
hnn_linear = hnn.Linear(in_features, out_features, c=c)
M = hnn_linear.weight.data.numpy()
b = hnn_linear.bias.data.numpy()
np_res = np_utils.Linear(x.data.numpy(), M, b, c_val)
torch_res = hnn_linear(x)
assert np.allclose(np_res, torch_res.data.numpy(), atol=1e-7)
def test_case():
x = torch.tensor(
np_utils.random_vec((batch_size, in_features)),
requires_grad=True).double()
torch.autograd.gradcheck(hnn_dense, x)
def test_case():
x = torch.tensor(np_utils.random_vec((batch_size,
in_features)), ).double()
res_dense = hnn_dense(x).data.numpy()
res_act_linear = torch.tanh(hnn_linear(x)).data.numpy()
assert np.allclose(res_dense, res_act_linear)
def test_case():
x = torch.tensor(
np_utils.random_vec((batch_size, timesteps, emb_size)),
requires_grad=True).double()
torch.autograd.gradcheck(hnn_rnn, x)