def test_channels_batch(self):
inpl = theano_var([[
[[0, 0, 0], [0, 0, 0], [0, 0, 0]],
[[0, 0, 0], [0, 0, 0], [3, 0, 0]]
],
[
[[0, 3, 3], [4, 5, 6], [7, 8, 4]],
[[-3, -3, -3], [-3, -3, -3], [3, 3, 3]]
]])
inpu = theano_var([[
[[1, 1, 1], [1, 1, 1], [1, 1, 1]],
[[1, 1, 1], [1, 1, 1], [4, 1, 1]]
],
[
[[2, 4, 4], [9, 9, 9], [9, 9, 9]],
[[2, 2, 2], [2, 2, 2], [5, 5, 5]]
]])
iinp = TheanoInterval(inpl, inpu)
doutl = theano_var([[
[[-1, -2], [-3, -4]],
[[1, 2], [-3, -2]]
],
[
[[1, 2], [-3, -2]],
[[-1, 1], [-1, 1]]
]])
doutu = theano_var([[
[[5, 4], [3, 2]],
[[4, 4], [4, 4]],
],
[
[[4, 5], [0, 1]],
[[0, 2], [0, 2]]
]])
idout = TheanoInterval(doutl, doutu)
shp = (2, 2, 3, 3)
din = d_pool(idout, iinp, shp, poolsize=(2, 2), mode='avg')
l, u = din.eval()
array_almost_equal(l, A([[
[[-1, -3, -2], [-4, -10, -6], [-3, -7, -4]],
[[1, 3, 2], [-2, -2, 0], [-3, -5, -2]]
],
[
[[1, 3, 2], [-2, -2, 0], [-3, -5, -2]],
[[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]]
]]) / 4.0)
array_almost_equal(u, A([[
[[5, 9, 4], [8, 14, 6], [3, 5, 2]],
[[4, 8, 4], [8, 16, 8], [4, 8, 4]]
],
[
[[4, 9, 5], [4, 10, 6], [0, 1, 1]],
[[0, 2, 2], [0, 4, 4], [0, 2, 2]]
]]) / 4.0)
def test_interval_negative(self):
actl = theano_var([-2, -1, -1, 0, 0, 1])
actu = theano_var([-1, 1, 0, 0, 1, 2])
doutl = theano_var([-3, -5, -7, -11, -13, -17])
doutu = theano_var([-2, -3, -5, -7, -11, -13])
iact = TheanoInterval(actl, actu)
idout = TheanoInterval(doutl, doutu)
idin = d_relu(idout, iact)
l, u = idin.eval()
rl = A([0, -5, -7, -11, -13, -17])
ru = A([0, 0, 0, 0, 0, -13])
array_almost_equal(l, rl)
array_almost_equal(u, ru)
def test_interval(self):
actl = theano_var([-2, -1, -1, 0, 0, 1])
actu = theano_var([-1, 1, 0, 0, 1, 2])
doutl = theano_var([2, 3, 4, 7, 11, 13])
doutu = theano_var([3, 5, 7, 11, 13, 17])
iact = TheanoInterval(actl, actu)
idout = TheanoInterval(doutl, doutu)
idin = d_relu(idout, iact)
l, u = idin.eval()
rl = A([0, 0, 0, 0, 0, 13])
ru = A([0, 5, 7, 11, 13, 17])
array_almost_equal(l, rl)
array_almost_equal(u, ru)
def test_simple(self):
inpl = theano_var([[[[0, 0, 0], [0, 0, 0], [0, 0, 0]]]])
inpu = theano_var([[[[1, 1, 1], [1, 1, 1], [1, 1, 1]]]])
iinp = TheanoInterval(inpl, inpu)
doutl = theano_var([[[[-1, -2], [-3, -4]]]])
doutu = theano_var([[[[5, 4], [3, 2]]]])
idout = TheanoInterval(doutl, doutu)
shp = (1, 1, 3, 3)
din = d_pool(idout, iinp, shp, poolsize=(2, 2), mode='avg')
l, u = din.eval()
array_almost_equal(l, A([[[[-1, -3, -2], [-4, -10, -6],
[-3, -7, -4]]]]) / 4.0)
array_almost_equal(u, A([[[[5, 9, 4], [8, 14, 6], [3, 5, 2]]]]) / 4.0)
def test_padding(self):
inpl = theano_var([[[[0, 0, 0], [0, 0, 0], [0, 0, 0]]]])
inpu = theano_var([[[[1, 1, 1], [1, 1, 1], [1, 1, 1]]]])
iinp = TheanoInterval(inpl, inpu)
doutl = theano_var([[[[-1, -2], [-3, -4]]]])
doutu = theano_var([[[[5, 4], [3, 2]]]])
idout = TheanoInterval(doutl, doutu)
shp = (1, 1, 3, 3)
din = d_pool(idout, iinp, shp, poolsize=(2, 2), padding=(1, 1),
stride=(3, 3), mode='max')
l, u = din.eval()
array_almost_equal(l, A([[[[-1, 0, -2], [0, 0, 0], [-3, 0, -4]]]]))
array_almost_equal(u, A([[[[5, 0, 4], [0, 0, 0], [3, 0, 2]]]]))
def test_softmax(self):
dout = TheanoInterval.derest_output(1000)
start_time = timeit.default_timer()
din = d_softmax(dout)
constr_time = timeit.default_timer()
l, u = din.eval()
ex_time = timeit.default_timer()
self.check_time('d_softmax', start_time, constr_time, ex_time)
def test_2D_batches(self):
dout = theano_var([[3, 6], [1, 2]])
idout = TheanoInterval(dout, dout)
w = theano_var([[9, 15], [12, 18]])
shp = (2, 2)
din = d_fully_connected(idout, w, shp)
l, u = din.eval()
array_almost_equal(l, u)
array_almost_equal(l, A([[117, 144], [39, 48]]))
def test_2D_simple_used_1D_of_weights(self):
dout = theano_var([[3, 6]])
idout = TheanoInterval(dout, dout)
w = theano_var([9, 12])
shp = (1, 1)
din = d_fully_connected(idout, w, shp)
l, u = din.eval()
array_almost_equal(l, u)
array_almost_equal(l, A([[99]]))
def test_2D_simple_used_2D_of_weights(self):
dout = theano_var([[3, 0]])
idout = TheanoInterval(dout, dout)
w = theano_var([[6, 0], [9, 0]])
shp = (1, 2)
din = d_fully_connected(idout, w, shp)
l, u = din.eval()
array_almost_equal(l, u)
array_almost_equal(l, A([[18, 27]]))
def test_1D_simple(self):
dout = theano_var([[1]])
idout = TheanoInterval(dout, dout)
w = theano_var([[2]])
shp = (1, 1)
din = d_fully_connected(idout, w, shp)
l, u = din.eval()
array_almost_equal(l, u)
array_almost_equal(l, A([[2]]))
def test_interval_3x3n(self):
doutl = theano_var([[-3, 0, 3], [3, -3, -5], [-3, -2, 1]])
doutu = theano_var([[-3, 2, 3], [5, 3, 2], [-1, 3, 3]])
idout = TheanoInterval(doutl, doutu)
idin = d_dropout(idout, 0.8)
l, u = idin.eval()
rl = A([[-0.6, 0, 0.6], [0.6, -0.6, -1], [-0.6, -0.4, 0.2]])
ru = A([[-0.6, 0.4, 0.6], [1, 0.6, 0.4], [-0.2, 0.6, 0.6]])
array_almost_equal(l, rl)
array_almost_equal(u, ru)
def test_neg_output(self):
inpl = theano_var([[[[1, 1], [1, 1]]]])
inpu = theano_var([[[[2, 2], [2, 2]]]])
iinp = TheanoInterval(inpl, inpu)
idout = theano_interval([[[[-3]]]])
shp = (1, 1, 2, 2)
din = d_pool(idout, iinp, shp, poolsize=(2, 2), mode='max')
l, u = din.eval()
array_almost_equal(l, A([[[[-3, -3], [-3, -3]]]]))
array_almost_equal(u, A([[[[0, 0], [0, 0]]]]))
def test_4x3x2(self):
shp = (4, 3, 2)
act = np.zeros(shp)
ind_n_in, ind_h, ind_w = np.indices(shp)
act = 100 * ind_n_in + 10 * ind_h + ind_w
thact = theano.shared(act)
iact = TheanoInterval(thact, thact)
thdout = theano.shared(np.ones(shp))
idout = TheanoInterval(thdout, thdout)
idin = d_relu(idout, iact)
l, u = idin.eval()
assert_almost_equal(l[0, 0, 0], 0.0)
assert_almost_equal(u[0, 0, 0], 1.0)
assert_almost_equal(l[2, 1, 1], 1.0)
assert_almost_equal(l[2, 2, 1], 1.0)
assert_almost_equal(l[1, 0, 1], 1.0)
assert_almost_equal(l[2, 1, 1], 1.0)
assert_almost_equal(l[2, 2, 0], 1.0)
assert_almost_equal(l[1, 0, 1], 1.0)
def test_2D_Intervals(self):
doutl = theano_var([[-3, -6, 3]])
doutu = theano_var([[9, -3, 6]])
idout = TheanoInterval(doutl, doutu)
w = theano_var([[2, -3, -3], [-3, 1, 2], [5, -4, 3], [-2, -3, -4]])
shp = (1, 2, 2)
din = d_fully_connected(idout, w, shp)
l, u = din.eval()
array_almost_equal(l, A([[[-15, -27], [6, -33]]]))
array_almost_equal(u, A([[[27, 18], [87, 12]]]))
def _get_theano_interval_result(self, lower, upper, function, *args,
**kwargs):
dim = len(lower.shape)
if dim == 1:
t_lower, t_upper = T.dvectors('inpl', 'inpu')
elif dim == 3:
t_lower, t_upper = T.tensor3s('inpl', 'inpu')
else:
raise NotImplementedError
iinp = TheanoInterval(t_lower, t_upper)
res = function(iinp, *args, **kwargs)
d = {t_lower: lower, t_upper: upper}
return res.eval(d)
def test_stride(self):
tinpl = theano.shared(np.arange(25).reshape((1, 1, 5, 5)))
tinpu = theano.shared(np.arange(25).reshape((1, 1, 5, 5)) + 2)
iinp = TheanoInterval(tinpl, tinpu)
idout = theano_interval([[[[-1, 2], [-3, 4]]]])
shp = (1, 1, 5, 5)
din = d_pool(idout, iinp, shp, poolsize=(2, 2), stride=(3, 3),
mode='avg')
l, u = din.eval()
array_almost_equal(l, A([[[[-1, -1, 0, 2, 2],
[-1, -1, 0, 2, 2],
[0, 0, 0, 0, 0],
[-3, -3, 0, 4, 4],
[-3, -3, 0, 4, 4]]]]) / 4.0)
array_almost_equal(u, A([[[[-1, -1, 0, 2, 2],
[-1, -1, 0, 2, 2],
[0, 0, 0, 0, 0],
[-3, -3, 0, 4, 4],
[-3, -3, 0, 4, 4]]]]) / 4.0)
def test_bitonicity_and_extremas_interval(self):
shp = (5, 1, 1)
tinpl, tinpu = T.tensor3s('tinpl', 'tinpu')
iinp = TheanoInterval(tinpl, tinpu)
out = a_norm(iinp, shp)
b = 200.0
a = (2.0 * (50000.0 + b * b))**0.5
inp = A([[[b]], [[0.0]], [[a]], [[0.0]], [[0.0]]])
d = {tinpl: inp, tinpu: inp}
l1, _ = out.eval(d)
inp[2, 0, 0] = a - 20.0
l2, _ = out.eval(d)
inp[2, 0, 0] = a + 20.0
l3, _ = out.eval(d)
inp[2, 0, 0] = a - 19.0
l4, _ = out.eval(d)
inp[2, 0, 0] = a - 18.0
l5, _ = out.eval(d)
inpl = inp.copy()
inpu = inp.copy()
d2 = {tinpl: inpl, tinpu: inpu}
inpl[2, 0, 0] = a - 18.0
inpu[2, 0, 0] = a + 20.0
l6, _ = out.eval(d2)
inpl[2, 0, 0] = a - 19.0
inpu[2, 0, 0] = a + 20.0
l7, _ = out.eval(d2)
inpl[2, 0, 0] = a - 20.0
inpu[2, 0, 0] = a + 20.0
l8, _ = out.eval(d2)
inpl[2, 0, 0] = a - 19.0
inpu[2, 0, 0] = a + 20.0
assert_greater(l1[2, 0, 0], l2[2, 0, 0])
assert_greater(l1[2, 0, 0], l3[2, 0, 0])
assert_greater(l5[2, 0, 0], l3[2, 0, 0])
assert_almost_equal(l3[2, 0, 0], l6[2, 0, 0], places=2)
assert_almost_equal(l3[2, 0, 0], l7[2, 0, 0], places=2)
def theano_interval(x):
v = theano_var(x)
return TheanoInterval(v, v)
def test_1_output(self):
dout = TheanoInterval.derest_output(1)
din = d_softmax(dout)
l, u = din.eval()
array_almost_equal(l, u)
array_almost_equal(l, A([[1]]))
def test_3_outputs(self):
dout = TheanoInterval.derest_output(3)
din = d_softmax(dout)
l, u = din.eval()
array_almost_equal(l, u)
array_almost_equal(l, A([[1, 0, 0], [0, 1, 0], [0, 0, 1]]))