def test_toplayer_delta():
X = np.random.randn(3, 5).astype(np.float32)
A = 5*np.random.randn(30, 50).astype(np.float32)
D = 5*np.random.randn(30, 50).astype(np.float32)
D_expected = D.copy()
D_expected = A - D_expected
Dd = op.to_gpu(D)
Yh = op.toplayer_delta(A, D, X)
assert_allclose(D_expected, Yh, rtol=1e-5, err_msg="CPU")
Ad = op.to_gpu(A)
Xd = op.to_gpu(X)
Yhd = op.toplayer_delta(Ad, Dd, Xd)
assert_allclose(D_expected, op.to_cpu(Yhd), rtol=1e-5, err_msg="GPU")
def test_toplayer_delta():
X = np.random.randn(3, 5).astype(np.float32)
A = 5 * np.random.randn(30, 50).astype(np.float32)
D = 5 * np.random.randn(30, 50).astype(np.float32)
D_expected = D.copy()
D_expected = A - D_expected
Dd = op.to_gpu(D)
Yh = op.toplayer_delta(A, D, X)
assert_allclose(D_expected, Yh, rtol=1e-5, err_msg="CPU")
Ad = op.to_gpu(A)
Xd = op.to_gpu(X)
Yhd = op.toplayer_delta(Ad, Dd, Xd)
assert_allclose(D_expected, op.to_cpu(Yhd), rtol=1e-5, err_msg="GPU")
def backward_pass(self, out, y, momentum=0.0):
delta = op.toplayer_delta(out, y, self.layers[-1].Z,
stream=op.streams[0])
if self.output_weights is not None:
op.inplace_scale_columns(delta, self.output_weights)
for l in reversed(self.layers):
delta = l.bprop(delta, momentum)
return delta # most of the time this is ignored anyhow
def test_nan_in_toplayer_delta():
size = (200, 10)
X = np.random.normal(size=size).astype(np.float32, order="c")
A = op.sigmoid(X)
Y = np.random.binomial(1.0, p=0.5, size=size).astype(np.float32)
M = np.random.binomial(1.0, p=0.9, size=size).astype(np.float32)
Y[~M.astype(np.bool)] = np.nan
Y_orig = Y.copy()
D = M * (A - Y)
D[~M.astype(np.bool)] = 0.0
Y = op.toplayer_delta(A, Y, X)
assert_allclose(Y, D)
Yd = op.to_gpu(Y_orig)
Ad = op.to_gpu(A)
Xd = op.to_gpu(X)
Yd = op.toplayer_delta(Ad, Yd, Xd)
assert_allclose(Yd.get(), D)
def backward_pass(self, out, y, momentum=0.0):
delta = op.toplayer_delta(out,
y,
self.layers[-1].Z,
stream=op.streams[0])
if self.output_weights is not None:
op.inplace_scale_columns(delta, self.output_weights)
for l in reversed(self.layers):
delta = l.bprop(delta, momentum)
return delta # most of the time this is ignored anyhow
def test_nan_in_toplayer_delta():
size = (200, 10)
X = np.random.normal(size=size).astype(np.float32, order="c")
A = op.sigmoid(X)
Y = np.random.binomial(1.0, p=0.5, size=size).astype(np.float32)
M = np.random.binomial(1.0, p=0.9, size=size).astype(np.float32)
Y[~M.astype(np.bool)] = np.nan
Y_orig = Y.copy()
D = M * (A - Y)
D[~M.astype(np.bool)] = 0.0
Y = op.toplayer_delta(A, Y, X)
assert_allclose(Y, D)
Yd = op.to_gpu(Y_orig)
Ad = op.to_gpu(A)
Xd = op.to_gpu(X)
Yd = op.toplayer_delta(Ad, Yd, Xd)
assert_allclose(Yd.get(), D)
