def test_Slice2D(self):
len = np.sqrt(self.len).astype(int)
np_a = np.reshape(self.np_a, (len, len))
# dc_a = dc.reshape(self.dc_a, (len, len)) $ BUG (reported bug.py) - for now using a WA
dc_a = dc.reshape(self.dc_a, (64, 64))
np_start = np.random.randint(len / 2)
np_end = np.random.randint(len / 2 + 1, len)
np_step = np.random.randint(1, len - 1)
np_axes = 1
np_start2 = np.random.randint(len / 2)
np_end2 = np.random.randint(len / 2 + 1, len)
np_step2 = np.random.randint(1, len - 1)
np_axes2 = 2
dc_start = dc.array(2).asTypeULong()
dc_end = dc.array(2).asTypeULong()
dc_axes = dc.array(2).asTypeInt()
dc_step = dc.array(2).asTypeULong()
dc_start[0] = np_start
dc_end[0] = np_end
dc_axes[0] = np_axes
dc_step[0] = np_step
dc_start[1] = np_start2
dc_end[1] = np_end2
dc_axes[1] = np_axes2
dc_step[1] = np_step2
npr = np_a[np_start:np_end:np_step, np_start2:np_end2:np_step2]
dcr = dc.slice(dc_a, dc_start, dc_end, dc_axes, dc_step)
np.testing.assert_allclose(npr.flatten(),
np.array(dcr.data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def setUp(self):
self.len = 24
self.np_a_float = np.random.randn(self.len).astype(np.int32)
self.np_b_float = np.random.randn(self.len).astype(np.int32)
self.dc_a_float = dc.array(list(self.np_a_float));
self.dc_b_float = dc.array(list(self.np_b_float));
self.fmod_flag = 1
def test_binary():
t1 = dc.array(2, 3).asTypeFloat()
t2 = dc.array(2, 3).asTypeInt()
add = dc.add(t1, t1)
add = t1 + t1
def setUp(self):
self.len = 24
self.np_a = np.random.randn(self.len).astype(np.float32)
self.np_b = np.random.randn(self.len).astype(np.float32)
self.dc_a = dc.array(list(self.np_a))
self.dc_b = dc.array(list(self.np_b))
def setUp(self):
self.len = 12
self.np_a = np.random.randint(12, size=self.len)
self.np_b = np.random.randint(12, size=self.len)
self.dc_a = dc.array(list(self.np_a)).asTypeInt()
self.dc_b = dc.array(list(self.np_b)).asTypeInt()
self.a_zero_point = dc.array(0).asTypeInt()
self.b_zero_point = dc.array(0).asTypeInt()
def setUp(self):
self.len = 24
self.np_x = np.random.randn(self.len).astype(np.int)
self.np_x_scale = np.random.randn(1).astype(np.float32)
self.np_x_zero_point = np.random.randn(1).astype(np.int)
self.dc_x = dc.array(list(self.np_x)).asTypeInt()
self.dc_x_scale = dc.array(list(self.np_x_scale))
self.dc_x_zero_point = dc.array(list(self.np_x_zero_point)).asTypeInt()
def setUp(self):
self.len = 48
self.np_float_a = np.random.randn(self.len).astype(np.float32)
self.dc_float_a = dc.array(list(self.np_float_a))
self.np_double_a = np.random.randn(self.len).astype(np.float64)
self.dc_double_a = dc.array(list(self.np_double_a))
def setUp(self):
self.len1 = 20
self.len2 = 40
self.np_a = np.random.randn(self.len1).astype(np.float32)
self.np_b = np.random.randn(self.len2).astype(np.float32)
self.np_c = np.random.randn(self.len1).astype(np.float32)
self.np_d = np.random.randn(self.len2).astype(np.float32)
self.dc_a = dc.array(list(self.np_a))
self.dc_b = dc.array(list(self.np_b))
self.dc_c = dc.array(list(self.np_c))
self.dc_d = dc.array(list(self.np_d))
self.err = "operands could not be broadcast together with shapes (2,5,2) (4,5,2) "
def setUp(self):
self.len = 24
self.np_a = np.random.randn(self.len).astype(np.float32)
self.dc_a = dc.array(list(self.np_a))
self.np_slope_1 = np.random.randn(1).astype(np.float32)
self.dc_slope_1 = dc.array(list(self.np_slope_1))
self.np_slope = np.random.randn(self.len).astype(np.float32)
self.dc_slope = dc.array(list(self.np_slope))
self.prelu_true = prelu_util(self.np_a, self.np_slope)
self.prelu_true_1 = prelu_util(self.np_a, self.np_slope_1)
def test_LSTM_1(self):
"""
input_shape: [7, 6, 8]
weight_shape: [1, 72, 8]
recurrence_weight_shape: [1, 72, 18]
bias_shape: [1, 144]
output_shape: [7, 1, 6, 18]
"""
np_X = np.load('swig/result/LSTM/test_LSTM_1/test_LSTM_1_X.npy')
np_W = np.load('swig/result/LSTM/test_LSTM_1/test_LSTM_1_W.npy')
np_R = np.load('swig/result/LSTM/test_LSTM_1/test_LSTM_1_R.npy')
np_B = np.load('swig/result/LSTM/test_LSTM_1/test_LSTM_1_B.npy')
np_sequence_lens = np.load('swig/result/LSTM/test_LSTM_1/test_LSTM_1_sequence_lens.npy')
np_initial_h = np.load('swig/result/LSTM/test_LSTM_1/test_LSTM_1_initial_h.npy')
np_initial_c = np.load('swig/result/LSTM/test_LSTM_1/test_LSTM_1_initial_c.npy')
np_P = np.load('swig/result/LSTM/test_LSTM_1/test_LSTM_1_P.npy')
dc_X = dc.array(np_X.flatten().tolist()).reshape(np_X.shape)
dc_W = dc.array(np_W.flatten().tolist()).reshape(np_W.shape)
dc_R = dc.array(np_R.flatten().tolist()).reshape(np_R.shape)
dc_B = dc.array(np_B.flatten().tolist()).reshape(np_B.shape)
dc_sequence_lens = dc.array(np_sequence_lens.flatten().tolist()).reshape(np_sequence_lens.shape)
# print(dc_sequence_lens)
dc_initial_h = dc.array(np_initial_h.flatten().tolist()).reshape(np_initial_h.shape)
dc_initial_c = dc.array(np_initial_c.flatten().tolist()).reshape(np_initial_c.shape)
dc_P = dc.array(np_P.flatten().tolist()).reshape(np_P.shape)
activation_alpha = [0.4966638953530237, 0.43607014563539637, 0.8097313919008828]
activation_beta = [0.12651506658849576, 0.1647539653231257, 0.04623650102301935]
activations = ['tanh', 'relu', 'sigmoid']
clip = 2.135794928171123
direction = "forward"
hidden_size = 18
input_forget = 1
rtr = np.load('swig/result/LSTM/test_LSTM_1/test_LSTM_1_Y.npy')
dcr = dc.lstm(dc_X, dc_W, dc_R, dc_B, dc_sequence_lens, dc_initial_h, dc_initial_c, dc_P)
def test_InstanceNormalization4D_1(self):
np_a = np.reshape(self.np_a, (2, 2, 2, 3))
dc_a = dc.reshape(self.dc_a, (2, 2, 2, 3))
scale = np.reshape(self.scale[0:2], (1, 2, 1, 1))
B = np.reshape(self.B[0:2], (1, 2, 1, 1))
np_a_mean = np_a.mean(axis=(0, 2, 3), keepdims=True)
np_a_var = np_a.var(axis=(0, 2, 3), keepdims=True)
npr = scale * (np_a - np_a_mean) / np.sqrt(np_a_var + self.epsilon) + B
dcr = dc.instancenormalization(
dc_a, dc.reshape(dc.array(list(self.scale[0:2])), (2)),
dc.reshape(dc.array(list(self.B[0:2])), (2)), self.epsilon)
np.testing.assert_allclose(npr.flatten(),
np.array(dcr.data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def test_Slice4D(self):
len = np.sqrt(np.sqrt(self.len)).astype(int)
np_a = np.reshape(self.np_a, (len, len, len, len))
# dc_a = dc.reshape(self.dc_a, (len, len, len))
dc_a = dc.reshape(self.dc_a, (8, 8, 8, 8))
np_start = np.random.randint(len / 2)
np_end = np.random.randint(len / 2 + 1, len)
np_step = np.random.randint(1, len - 1)
np_axes = 1
np_start2 = np.random.randint(len / 2)
np_end2 = np.random.randint(len / 2 + 1, len)
np_step2 = np.random.randint(1, len - 1)
np_axes2 = 2
np_start3 = np.random.randint(len / 2)
np_end3 = np.random.randint(len / 2 + 1, len)
np_step3 = np.random.randint(1, len - 1)
np_axes3 = 3
np_start4 = np.random.randint(len / 2)
np_end4 = np.random.randint(len / 2 + 1, len)
np_step4 = np.random.randint(1, len - 1)
np_axes4 = 4
dc_start = dc.array(4).asTypeULong()
dc_end = dc.array(4).asTypeULong()
dc_axes = dc.array(4).asTypeInt()
dc_step = dc.array(4).asTypeULong()
dc_start[0] = np_start
dc_end[0] = np_end
dc_axes[0] = np_axes
dc_step[0] = np_step
dc_start[1] = np_start2
dc_end[1] = np_end2
dc_axes[1] = np_axes2
dc_step[1] = np_step2
dc_start[2] = np_start3
dc_end[2] = np_end3
dc_axes[2] = np_axes3
dc_step[2] = np_step3
dc_start[3] = np_start4
dc_end[3] = np_end4
dc_axes[3] = np_axes4
dc_step[3] = np_step4
npr = np_a[np_start:np_end:np_step, np_start2:np_end2:np_step2,
np_start3:np_end3:np_step3, np_start4:np_end4:np_step4]
dcr = dc.slice(dc_a, dc_start, dc_end, dc_axes, dc_step)
np.testing.assert_allclose(npr.flatten(),
np.array(dcr.data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def test_Expand1D(self):
npr = np.broadcast_to(self.np_a, (3, 3))
dcr = dc.expand(self.dc_a, dc.array([3, 3]).asTypeInt())
np.testing.assert_allclose(npr,
np.array(dcr.data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def setUp(self):
self.len = 24
self.np_a = np.random.randn(self.len).astype(np.float32)
self.dc_a = dc.array(list(self.np_a))
self.scale = np.random.randn(4).astype(np.float32)
self.B = np.random.randn(4).astype(np.float32)
self.epsilon = 1e-05
def setUp(self):
self.len = 24
self.np_a = np.random.randn(self.len).astype(np.float32)
self.dc_a = dc.array(list(self.np_a))
self.alpha = 0.0001
self.beta = 0.75
self.bias = 1.0
self.size = 2
def setUp(self):
self.len = 24
self.np_a = np.random.randn(self.len)
self.np_a.ravel()[np.random.choice(self.np_a.size, 5,
replace=False)] = np.inf
self.np_a.ravel()[np.random.choice(self.np_a.size, 5,
replace=False)] = -np.inf
self.dc_a = dc.array(list(self.np_a))
self.detect_positive = 0
self.detect_negative = 1
def test_Slice1D (self):
len = self.len
np_a = self.np_a
dc_a = self.dc_a
np_start = np.random.randint(len/2)
np_end = np.random.randint(len/2 + 1, len)
np_step = np.random.randint(1, len-1)
np_axes = 0
dc_start = dc.array(1).asTypeInt()
dc_end = dc.array(1).asTypeInt()
dc_axes = dc.array(1).asTypeInt()
dc_step = dc.array(1).asTypeInt()
dc_start[0] = np_start
dc_end[0] = np_end
dc_axes[0] = np_axes
dc_step[0] = np_step
npr = np_a[np_start:np_end:np_step]
dcr = dc.slice(dc_a, dc_start, dc_end, dc_axes, dc_step)
np.testing.assert_allclose(npr, np.array(dcr.data()).astype(np.float32),
rtol=1e-3, atol=1e-3)
def test_LSTM_8(self):
"""
input_shape: [5, 1, 9]
weight_shape: [2, 4, 9]
recurrence_weight_shape: [2, 4, 1]
bias_shape: [2, 8]
output_shape: [5, 2, 1, 1]
"""
np_X = np.load('swig/result/LSTM/test_LSTM_8/test_LSTM_8_X.npy')
np_W = np.load('swig/result/LSTM/test_LSTM_8/test_LSTM_8_W.npy')
np_R = np.load('swig/result/LSTM/test_LSTM_8/test_LSTM_8_R.npy')
np_B = np.load('swig/result/LSTM/test_LSTM_8/test_LSTM_8_B.npy')
np_sequence_lens = np.load(
'swig/result/LSTM/test_LSTM_8/test_LSTM_8_sequence_lens.npy')
np_initial_h = np.load(
'swig/result/LSTM/test_LSTM_8/test_LSTM_8_initial_h.npy')
np_initial_c = np.load(
'swig/result/LSTM/test_LSTM_8/test_LSTM_8_initial_c.npy')
np_P = np.load('swig/result/LSTM/test_LSTM_8/test_LSTM_8_P.npy')
dc_X = dc.array(np_X.flatten().tolist()).reshape(np_X.shape)
dc_W = dc.array(np_W.flatten().tolist()).reshape(np_W.shape)
dc_R = dc.array(np_R.flatten().tolist()).reshape(np_R.shape)
dc_B = dc.array(np_B.flatten().tolist()).reshape(np_B.shape)
dc_sequence_lens = dc.array(
np_sequence_lens.flatten().tolist()).reshape(
np_sequence_lens.shape)
dc_initial_h = dc.array(np_initial_h.flatten().tolist()).reshape(
np_initial_h.shape)
dc_initial_c = dc.array(np_initial_c.flatten().tolist()).reshape(
np_initial_c.shape)
dc_P = dc.array(np_P.flatten().tolist()).reshape(np_P.shape)
activation_alpha = [
0.7746672952847123, 0.036382870533804956, 0.4848161740062119,
0.9830896771807061, 0.017064708201858125, 0.6242851269185792
]
activation_beta = [
0.2517994027716025, 0.28976631245816886, 0.38611683342345127,
0.13080875018242, 0.40170849770653727, 0.956570288835856
]
activations = ['sigmoid', 'relu', 'sigmoid', 'relu', 'tanh', 'tanh']
clip = 2.72219901402834
direction = "bidirectional"
hidden_size = 1
input_forget = 20
rtr = np.load('swig/result/LSTM/test_LSTM_8/test_LSTM_8_Y.npy')
dcr = dc.lstm(dc_X, dc_W, dc_R, dc_B, dc_sequence_lens, dc_initial_h,
dc_initial_c, dc_P)
np.testing.assert_allclose(rtr.flatten(),
np.array(dcr[0].data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def test_LSTM_2(self):
"""
input_shape: [8, 4, 1]
weight_shape: [2, 64, 1]
recurrence_weight_shape: [2, 64, 16]
bias_shape: [2, 128]
output_shape: [8, 2, 4, 16]
"""
np_X = np.load('swig/result/LSTM/test_LSTM_2/test_LSTM_2_X.npy')
np_W = np.load('swig/result/LSTM/test_LSTM_2/test_LSTM_2_W.npy')
np_R = np.load('swig/result/LSTM/test_LSTM_2/test_LSTM_2_R.npy')
np_B = np.load('swig/result/LSTM/test_LSTM_2/test_LSTM_2_B.npy')
np_sequence_lens = np.load(
'swig/result/LSTM/test_LSTM_2/test_LSTM_2_sequence_lens.npy')
np_initial_h = np.load(
'swig/result/LSTM/test_LSTM_2/test_LSTM_2_initial_h.npy')
np_initial_c = np.load(
'swig/result/LSTM/test_LSTM_2/test_LSTM_2_initial_c.npy')
np_P = np.load('swig/result/LSTM/test_LSTM_2/test_LSTM_2_P.npy')
dc_X = dc.array(np_X.flatten().tolist()).reshape(np_X.shape)
dc_W = dc.array(np_W.flatten().tolist()).reshape(np_W.shape)
dc_R = dc.array(np_R.flatten().tolist()).reshape(np_R.shape)
dc_B = dc.array(np_B.flatten().tolist()).reshape(np_B.shape)
dc_sequence_lens = dc.array(
np_sequence_lens.flatten().tolist()).reshape(
np_sequence_lens.shape)
dc_initial_h = dc.array(np_initial_h.flatten().tolist()).reshape(
np_initial_h.shape)
dc_initial_c = dc.array(np_initial_c.flatten().tolist()).reshape(
np_initial_c.shape)
dc_P = dc.array(np_P.flatten().tolist()).reshape(np_P.shape)
activation_alpha = [
0.20332784907676504, 0.22637955219185357, 0.6021193542725863,
0.6168572580474495, 0.40207405192136414, 0.036317260701121845
]
activation_beta = [
0.7717703726511062, 0.027305984207814826, 0.8047659241021807,
0.6452577518231254, 0.7319012533727602, 0.25505174775324035
]
activations = ['tanh', 'tanh', 'sigmoid', 'relu', 'sigmoid', 'relu']
clip = 2.907158875085247
direction = "bidirectional"
hidden_size = 16
input_forget = 10
rtr = np.load('swig/result/LSTM/test_LSTM_2/test_LSTM_2_Y.npy')
dcr = dc.lstm(dc_X, dc_W, dc_R, dc_B, dc_sequence_lens, dc_initial_h,
dc_initial_c, dc_P)
np.testing.assert_allclose(rtr.flatten(),
np.array(dcr[0].data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def test_LSTM_10(self):
"""
input_shape: [9, 6, 2]
weight_shape: [2, 8, 2]
recurrence_weight_shape: [2, 8, 2]
bias_shape: [2, 16]
output_shape: [9, 2, 6, 2]
"""
np_X = np.load('swig/result/LSTM/test_LSTM_10/test_LSTM_10_X.npy')
np_W = np.load('swig/result/LSTM/test_LSTM_10/test_LSTM_10_W.npy')
np_R = np.load('swig/result/LSTM/test_LSTM_10/test_LSTM_10_R.npy')
np_B = np.load('swig/result/LSTM/test_LSTM_10/test_LSTM_10_B.npy')
np_sequence_lens = np.load(
'swig/result/LSTM/test_LSTM_10/test_LSTM_10_sequence_lens.npy')
np_initial_h = np.load(
'swig/result/LSTM/test_LSTM_10/test_LSTM_10_initial_h.npy')
np_initial_c = np.load(
'swig/result/LSTM/test_LSTM_10/test_LSTM_10_initial_c.npy')
np_P = np.load('swig/result/LSTM/test_LSTM_10/test_LSTM_10_P.npy')
dc_X = dc.array(np_X.flatten().tolist()).reshape(np_X.shape)
dc_W = dc.array(np_W.flatten().tolist()).reshape(np_W.shape)
dc_R = dc.array(np_R.flatten().tolist()).reshape(np_R.shape)
dc_B = dc.array(np_B.flatten().tolist()).reshape(np_B.shape)
dc_sequence_lens = dc.array(
np_sequence_lens.flatten().tolist()).reshape(
np_sequence_lens.shape)
dc_initial_h = dc.array(np_initial_h.flatten().tolist()).reshape(
np_initial_h.shape)
dc_initial_c = dc.array(np_initial_c.flatten().tolist()).reshape(
np_initial_c.shape)
dc_P = dc.array(np_P.flatten().tolist()).reshape(np_P.shape)
activation_alpha = [
0.5494076090797351, 0.4486022544214028, 0.8555569145519173,
0.36385914141140563, 0.2786060330869964, 0.3709594247211093
]
activation_beta = [
0.6841038069275263, 0.12454085979724905, 0.16010194778825715,
0.43645368358634684, 0.2006827543226236, 0.025382308479808713
]
activations = ['relu', 'tanh', 'relu', 'sigmoid', 'sigmoid', 'tanh']
clip = 7.52494780016543
direction = "bidirectional"
hidden_size = 2
input_forget = 19
rtr = np.load('swig/result/LSTM/test_LSTM_10/test_LSTM_10_Y.npy')
dcr = dc.lstm(dc_X, dc_W, dc_R, dc_B, dc_sequence_lens, dc_initial_h,
dc_initial_c, dc_P)
np.testing.assert_allclose(rtr.flatten(),
np.array(dcr[0].data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def test_LSTM_7(self):
"""
input_shape: [6, 8, 6]
weight_shape: [2, 40, 6]
recurrence_weight_shape: [2, 40, 10]
bias_shape: [2, 80]
output_shape: [6, 2, 8, 10]
"""
np_X = np.load('swig/result/LSTM/test_LSTM_7/test_LSTM_7_X.npy')
np_W = np.load('swig/result/LSTM/test_LSTM_7/test_LSTM_7_W.npy')
np_R = np.load('swig/result/LSTM/test_LSTM_7/test_LSTM_7_R.npy')
np_B = np.load('swig/result/LSTM/test_LSTM_7/test_LSTM_7_B.npy')
np_sequence_lens = np.load(
'swig/result/LSTM/test_LSTM_7/test_LSTM_7_sequence_lens.npy')
np_initial_h = np.load(
'swig/result/LSTM/test_LSTM_7/test_LSTM_7_initial_h.npy')
np_initial_c = np.load(
'swig/result/LSTM/test_LSTM_7/test_LSTM_7_initial_c.npy')
np_P = np.load('swig/result/LSTM/test_LSTM_7/test_LSTM_7_P.npy')
dc_X = dc.array(np_X.flatten().tolist()).reshape(np_X.shape)
dc_W = dc.array(np_W.flatten().tolist()).reshape(np_W.shape)
dc_R = dc.array(np_R.flatten().tolist()).reshape(np_R.shape)
dc_B = dc.array(np_B.flatten().tolist()).reshape(np_B.shape)
dc_sequence_lens = dc.array(
np_sequence_lens.flatten().tolist()).reshape(
np_sequence_lens.shape)
dc_initial_h = dc.array(np_initial_h.flatten().tolist()).reshape(
np_initial_h.shape)
dc_initial_c = dc.array(np_initial_c.flatten().tolist()).reshape(
np_initial_c.shape)
dc_P = dc.array(np_P.flatten().tolist()).reshape(np_P.shape)
activation_alpha = [
0.28920619362824995, 0.747465052565989, 0.661162342694396,
0.8477376049646675, 0.07881817761441567, 0.16208001287665696
]
activation_beta = [
0.7627506699799991, 0.6606114297796492, 0.9585330972395699,
0.5549681443136113, 0.059042596260018065, 0.04648254501072813
]
activations = ['sigmoid', 'sigmoid', 'tanh', 'relu', 'relu', 'tanh']
clip = 3.879685115272961
direction = "bidirectional"
hidden_size = 10
input_forget = 11
rtr = np.load('swig/result/LSTM/test_LSTM_7/test_LSTM_7_Y.npy')
dcr = dc.lstm(dc_X, dc_W, dc_R, dc_B, dc_sequence_lens, dc_initial_h,
dc_initial_c, dc_P)
np.testing.assert_allclose(rtr.flatten(),
np.array(dcr[0].data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def test_LSTM_4(self):
"""
input_shape: [2, 1, 1]
weight_shape: [2, 72, 1]
recurrence_weight_shape: [2, 72, 18]
bias_shape: [2, 144]
output_shape: [2, 2, 1, 18]
"""
np_X = np.load('swig/result/LSTM/test_LSTM_4/test_LSTM_4_X.npy')
np_W = np.load('swig/result/LSTM/test_LSTM_4/test_LSTM_4_W.npy')
np_R = np.load('swig/result/LSTM/test_LSTM_4/test_LSTM_4_R.npy')
np_B = np.load('swig/result/LSTM/test_LSTM_4/test_LSTM_4_B.npy')
np_sequence_lens = np.load(
'swig/result/LSTM/test_LSTM_4/test_LSTM_4_sequence_lens.npy')
np_initial_h = np.load(
'swig/result/LSTM/test_LSTM_4/test_LSTM_4_initial_h.npy')
np_initial_c = np.load(
'swig/result/LSTM/test_LSTM_4/test_LSTM_4_initial_c.npy')
np_P = np.load('swig/result/LSTM/test_LSTM_4/test_LSTM_4_P.npy')
dc_X = dc.array(np_X.flatten().tolist()).reshape(np_X.shape)
dc_W = dc.array(np_W.flatten().tolist()).reshape(np_W.shape)
dc_R = dc.array(np_R.flatten().tolist()).reshape(np_R.shape)
dc_B = dc.array(np_B.flatten().tolist()).reshape(np_B.shape)
dc_sequence_lens = dc.array(
np_sequence_lens.flatten().tolist()).reshape(
np_sequence_lens.shape)
dc_initial_h = dc.array(np_initial_h.flatten().tolist()).reshape(
np_initial_h.shape)
dc_initial_c = dc.array(np_initial_c.flatten().tolist()).reshape(
np_initial_c.shape)
dc_P = dc.array(np_P.flatten().tolist()).reshape(np_P.shape)
activation_alpha = [
0.9860778314893995, 0.12417696210947016, 0.0006744261981547206,
0.24339585920465567, 0.7498252461249489, 0.30754908604622977
]
activation_beta = [
0.1603792258866038, 0.1880417110347281, 0.6952466604231525,
0.11767276043277997, 0.61860245840078, 0.6615465711832315
]
activations = ['sigmoid', 'relu', 'sigmoid', 'tanh', 'relu', 'tanh']
clip = 3.7019881776389996
direction = "bidirectional"
hidden_size = 18
input_forget = 8
rtr = np.load('swig/result/LSTM/test_LSTM_4/test_LSTM_4_Y.npy')
dcr = dc.lstm(dc_X, dc_W, dc_R, dc_B, dc_sequence_lens, dc_initial_h,
dc_initial_c, dc_P)
np.testing.assert_allclose(rtr.flatten(),
np.array(dcr[0].data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def test_LSTM_5(self):
"""
input_shape: [2, 3, 10]
weight_shape: [2, 20, 10]
recurrence_weight_shape: [2, 20, 5]
bias_shape: [2, 40]
output_shape: [2, 2, 3, 5]
"""
np_X = np.load('swig/result/LSTM/test_LSTM_5/test_LSTM_5_X.npy')
np_W = np.load('swig/result/LSTM/test_LSTM_5/test_LSTM_5_W.npy')
np_R = np.load('swig/result/LSTM/test_LSTM_5/test_LSTM_5_R.npy')
np_B = np.load('swig/result/LSTM/test_LSTM_5/test_LSTM_5_B.npy')
np_sequence_lens = np.load(
'swig/result/LSTM/test_LSTM_5/test_LSTM_5_sequence_lens.npy')
np_initial_h = np.load(
'swig/result/LSTM/test_LSTM_5/test_LSTM_5_initial_h.npy')
np_initial_c = np.load(
'swig/result/LSTM/test_LSTM_5/test_LSTM_5_initial_c.npy')
np_P = np.load('swig/result/LSTM/test_LSTM_5/test_LSTM_5_P.npy')
dc_X = dc.array(np_X.flatten().tolist()).reshape(np_X.shape)
dc_W = dc.array(np_W.flatten().tolist()).reshape(np_W.shape)
dc_R = dc.array(np_R.flatten().tolist()).reshape(np_R.shape)
dc_B = dc.array(np_B.flatten().tolist()).reshape(np_B.shape)
dc_sequence_lens = dc.array(
np_sequence_lens.flatten().tolist()).reshape(
np_sequence_lens.shape)
dc_initial_h = dc.array(np_initial_h.flatten().tolist()).reshape(
np_initial_h.shape)
dc_initial_c = dc.array(np_initial_c.flatten().tolist()).reshape(
np_initial_c.shape)
dc_P = dc.array(np_P.flatten().tolist()).reshape(np_P.shape)
activation_alpha = [
0.9958868560901981, 0.5615704868314114, 0.5054884381550756,
0.5125119319409338, 0.18310275479264726, 0.4990119412451889
]
activation_beta = [
0.2876466600692591, 0.560778821439632, 0.2632346842213401,
0.13121922832510213, 0.8822817678248556, 0.9880592276419286
]
activations = ['tanh', 'relu', 'tanh', 'sigmoid', 'sigmoid', 'relu']
clip = 6.117108798702516
direction = "bidirectional"
hidden_size = 5
input_forget = 17
rtr = np.load('swig/result/LSTM/test_LSTM_5/test_LSTM_5_Y.npy')
dcr = dc.lstm(dc_X, dc_W, dc_R, dc_B, dc_sequence_lens, dc_initial_h,
dc_initial_c, dc_P)
np.testing.assert_allclose(rtr.flatten(),
np.array(dcr[0].data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def test_LSTM_9(self):
"""
input_shape: [1, 2, 9]
weight_shape: [1, 52, 9]
recurrence_weight_shape: [1, 52, 13]
bias_shape: [1, 104]
output_shape: [1, 1, 2, 13]
"""
np_X = np.load('swig/result/LSTM/test_LSTM_9/test_LSTM_9_X.npy')
np_W = np.load('swig/result/LSTM/test_LSTM_9/test_LSTM_9_W.npy')
np_R = np.load('swig/result/LSTM/test_LSTM_9/test_LSTM_9_R.npy')
np_B = np.load('swig/result/LSTM/test_LSTM_9/test_LSTM_9_B.npy')
np_sequence_lens = np.load(
'swig/result/LSTM/test_LSTM_9/test_LSTM_9_sequence_lens.npy')
np_initial_h = np.load(
'swig/result/LSTM/test_LSTM_9/test_LSTM_9_initial_h.npy')
np_initial_c = np.load(
'swig/result/LSTM/test_LSTM_9/test_LSTM_9_initial_c.npy')
np_P = np.load('swig/result/LSTM/test_LSTM_9/test_LSTM_9_P.npy')
dc_X = dc.array(np_X.flatten().tolist()).reshape(np_X.shape)
dc_W = dc.array(np_W.flatten().tolist()).reshape(np_W.shape)
dc_R = dc.array(np_R.flatten().tolist()).reshape(np_R.shape)
dc_B = dc.array(np_B.flatten().tolist()).reshape(np_B.shape)
dc_sequence_lens = dc.array(
np_sequence_lens.flatten().tolist()).reshape(
np_sequence_lens.shape)
dc_initial_h = dc.array(np_initial_h.flatten().tolist()).reshape(
np_initial_h.shape)
dc_initial_c = dc.array(np_initial_c.flatten().tolist()).reshape(
np_initial_c.shape)
dc_P = dc.array(np_P.flatten().tolist()).reshape(np_P.shape)
activation_alpha = [
0.08447232888329703, 0.6786879671317316, 0.6558691737892577
]
activation_beta = [
0.7615097936520958, 0.5651098460911419, 0.2265325436094976
]
activations = ['sigmoid', 'relu', 'tanh']
clip = 6.4355391083683635
direction = "forward"
hidden_size = 13
input_forget = 14
rtr = np.load('swig/result/LSTM/test_LSTM_9/test_LSTM_9_Y.npy')
dcr = dc.lstm(dc_X, dc_W, dc_R, dc_B, dc_sequence_lens, dc_initial_h,
dc_initial_c, dc_P)
np.testing.assert_allclose(rtr.flatten(),
np.array(dcr[0].data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def test_LSTM_3(self):
"""
input_shape: [8, 1, 4]
weight_shape: [1, 56, 4]
recurrence_weight_shape: [1, 56, 14]
bias_shape: [1, 112]
output_shape: [8, 1, 1, 14]
"""
np_X = np.load('swig/result/LSTM/test_LSTM_3/test_LSTM_3_X.npy')
np_W = np.load('swig/result/LSTM/test_LSTM_3/test_LSTM_3_W.npy')
np_R = np.load('swig/result/LSTM/test_LSTM_3/test_LSTM_3_R.npy')
np_B = np.load('swig/result/LSTM/test_LSTM_3/test_LSTM_3_B.npy')
np_sequence_lens = np.load(
'swig/result/LSTM/test_LSTM_3/test_LSTM_3_sequence_lens.npy')
np_initial_h = np.load(
'swig/result/LSTM/test_LSTM_3/test_LSTM_3_initial_h.npy')
np_initial_c = np.load(
'swig/result/LSTM/test_LSTM_3/test_LSTM_3_initial_c.npy')
np_P = np.load('swig/result/LSTM/test_LSTM_3/test_LSTM_3_P.npy')
dc_X = dc.array(np_X.flatten().tolist()).reshape(np_X.shape)
dc_W = dc.array(np_W.flatten().tolist()).reshape(np_W.shape)
dc_R = dc.array(np_R.flatten().tolist()).reshape(np_R.shape)
dc_B = dc.array(np_B.flatten().tolist()).reshape(np_B.shape)
dc_sequence_lens = dc.array(
np_sequence_lens.flatten().tolist()).reshape(
np_sequence_lens.shape)
dc_initial_h = dc.array(np_initial_h.flatten().tolist()).reshape(
np_initial_h.shape)
dc_initial_c = dc.array(np_initial_c.flatten().tolist()).reshape(
np_initial_c.shape)
dc_P = dc.array(np_P.flatten().tolist()).reshape(np_P.shape)
activation_alpha = [
0.5353786525215217, 0.0047814145847226985, 0.17116077889292602
]
activation_beta = [
0.8724323449420001, 0.9207316192126214, 0.7391156087035118
]
activations = ['relu', 'sigmoid', 'tanh']
clip = 7.5397611403351
direction = "reverse"
hidden_size = 14
input_forget = 14
rtr = np.load('swig/result/LSTM/test_LSTM_3/test_LSTM_3_Y.npy')
dcr = dc.lstm(dc_X, dc_W, dc_R, dc_B, dc_sequence_lens, dc_initial_h,
dc_initial_c, dc_P)
np.testing.assert_allclose(rtr.flatten(),
np.array(dcr[0].data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def test_LSTM_6(self):
"""
input_shape: [7, 5, 9]
weight_shape: [1, 64, 9]
recurrence_weight_shape: [1, 64, 16]
bias_shape: [1, 128]
output_shape: [7, 1, 5, 16]
"""
np_X = np.load('swig/result/LSTM/test_LSTM_6/test_LSTM_6_X.npy')
np_W = np.load('swig/result/LSTM/test_LSTM_6/test_LSTM_6_W.npy')
np_R = np.load('swig/result/LSTM/test_LSTM_6/test_LSTM_6_R.npy')
np_B = np.load('swig/result/LSTM/test_LSTM_6/test_LSTM_6_B.npy')
np_sequence_lens = np.load(
'swig/result/LSTM/test_LSTM_6/test_LSTM_6_sequence_lens.npy')
np_initial_h = np.load(
'swig/result/LSTM/test_LSTM_6/test_LSTM_6_initial_h.npy')
np_initial_c = np.load(
'swig/result/LSTM/test_LSTM_6/test_LSTM_6_initial_c.npy')
np_P = np.load('swig/result/LSTM/test_LSTM_6/test_LSTM_6_P.npy')
dc_X = dc.array(np_X.flatten().tolist()).reshape(np_X.shape)
dc_W = dc.array(np_W.flatten().tolist()).reshape(np_W.shape)
dc_R = dc.array(np_R.flatten().tolist()).reshape(np_R.shape)
dc_B = dc.array(np_B.flatten().tolist()).reshape(np_B.shape)
dc_sequence_lens = dc.array(
np_sequence_lens.flatten().tolist()).reshape(
np_sequence_lens.shape)
dc_initial_h = dc.array(np_initial_h.flatten().tolist()).reshape(
np_initial_h.shape)
dc_initial_c = dc.array(np_initial_c.flatten().tolist()).reshape(
np_initial_c.shape)
dc_P = dc.array(np_P.flatten().tolist()).reshape(np_P.shape)
activation_alpha = [
0.1508855746391079, 0.4507448733258578, 0.41656131175216204
]
activation_beta = [
0.5657658415464043, 0.21611300965755376, 0.15922967506138452
]
activations = ['tanh', 'relu', 'sigmoid']
clip = 3.1767036746309287
direction = "forward"
hidden_size = 16
input_forget = 14
rtr = np.load('swig/result/LSTM/test_LSTM_6/test_LSTM_6_Y.npy')
dcr = dc.lstm(dc_X, dc_W, dc_R, dc_B, dc_sequence_lens, dc_initial_h,
dc_initial_c, dc_P)
np.testing.assert_allclose(rtr.flatten(),
np.array(dcr[0].data()).astype(np.float32),
rtol=1e-3,
atol=1e-3)
def setUp(self):
self.nullT = dc.array(0)
self.zeros = dc.zeros(2, 3).asTypeInt()
self.ones = dc.ones(2, 3).asTypeInt()
self.f0_4 = dc.arange(5)
self.f5_9 = dc.arange(10, 5)
self.np_f0_4 = np.arange(5)
self.np_f5_9 = np.arange(10, 5)
self.i0_4 = self.f0_4.asTypeInt()
self.i5_9 = self.f5_9.asTypeInt()
self.np_i0_4 = self.np_f0_4.astype(np.int)
self.np_i5_9 = self.np_f5_9.astype(np.int)
self.b0_4 = self.f0_4.asTypeBool()
self.b5_9 = self.f5_9.asTypeBool()
def setUp(self):
self.num_directions = 1
self.seq_length = 1
self.input_size = 4
self.batch_size = 3
self.hidden_size = 5
self.np_x = np.random.randn(self.seq_length * self.batch_size * self.input_size).astype(np.float32)
self.dc_x = dc.array(list(self.np_x))
self.np_w = np.random.randn(self.num_directions * 4 * self.hidden_size * self.input_size).astype(np.float32)
self.dc_w = dc.array(list(self.np_w))
self.np_r = np.random.randn(self.num_directions * 4 * self.hidden_size * self.hidden_size).astype(np.float32)
self.dc_r = dc.array(list(self.np_r))
self.np_b = np.random.randn(self.num_directions * 8 * self.hidden_size).astype(np.float32)
self.dc_b = dc.array(list(self.np_b))
self.np_h = np.random.randn(self.num_directions * self.batch_size * self.hidden_size).astype(np.float32)
self.dc_h = dc.array(list(self.np_h))
self.np_c = np.random.randn(self.num_directions * self.batch_size * self.hidden_size).astype(np.float32)
self.dc_c = dc.array(list(self.np_c))
self.np_p = np.random.randn(self.num_directions * 3 * self.hidden_size).astype(np.float32)
self.dc_p = dc.array(list(self.np_p))
self.np_x = np.reshape(self.np_x, (self.seq_length, self.batch_size, self.input_size))
self.dc_x = dc.reshape(self.dc_x, (self.seq_length, self.batch_size, self.input_size))
self.np_w = np.reshape(self.np_w, (self.num_directions, 4 * self.hidden_size, self.input_size))
self.dc_w = dc.reshape(self.dc_w, (self.num_directions, 4 * self.hidden_size, self.input_size))
self.np_r = np.reshape(self.np_r, (self.num_directions, 4 * self.hidden_size, self.hidden_size))
self.dc_r = dc.reshape(self.dc_r, (self.num_directions, 4 * self.hidden_size, self.hidden_size))
self.np_b = np.reshape(self.np_b, (self.num_directions, 8 * self.hidden_size))
self.dc_b = dc.reshape(self.dc_b, (self.num_directions, 8 * self.hidden_size))
self.np_h = np.reshape(self.np_h, (self.num_directions, self.batch_size, self.hidden_size))
self.dc_h = dc.reshape(self.dc_h, (self.num_directions, self.batch_size, self.hidden_size))
self.np_c = np.reshape(self.np_c, (self.num_directions, self.batch_size, self.hidden_size))
self.dc_c = dc.reshape(self.dc_c, (self.num_directions, self.batch_size, self.hidden_size))
self.np_p = np.reshape(self.np_p, (self.num_directions, 3 * self.hidden_size))
self.dc_p = dc.reshape(self.dc_p, (self.num_directions, 3 * self.hidden_size))
def test_Slice3D (self):
len = np.cbrt(self.len).astype(int)
np_a = np.reshape(self.np_a, (len, len, len))
# dc_a = dc.reshape(self.dc_a, (len, len, len))
dc_a = dc.reshape(self.dc_a, (16, 16, 16))
np_start = np.random.randint(len/2)
np_end = np.random.randint(len/2 + 1, len)
np_step = np.random.randint(1, len-1)
np_axes = 0
np_start2 = np.random.randint(len/2)
np_end2 = np.random.randint(len/2 + 1, len)
np_step2 = np.random.randint(1, len-1)
np_axes2 = 1
np_start3 = np.random.randint(len/2)
np_end3 = np.random.randint(len/2 + 1, len)
np_step3 = np.random.randint(1, len-1)
np_axes3 = 2
dc_start = dc.array(3).asTypeInt()
dc_end = dc.array(3).asTypeInt()
dc_axes = dc.array(3).asTypeInt()
dc_step = dc.array(3).asTypeInt()
dc_start[0] = np_start
dc_end[0] = np_end
dc_axes[0] = np_axes
dc_step[0] = np_step
dc_start[1] = np_start2
dc_end[1] = np_end2
dc_axes[1] = np_axes2
dc_step[1] = np_step2
dc_start[2] = np_start3
dc_end[2] = np_end3
dc_axes[2] = np_axes3
dc_step[2] = np_step3
npr = np_a[np_start:np_end:np_step, np_start2:np_end2:np_step2, np_start3:np_end3:np_step3]
dcr = dc.slice(dc_a, dc_start, dc_end, dc_axes, dc_step)
np.testing.assert_allclose(npr.flatten(), np.array(dcr.data()).astype(np.float32),
rtol=1e-3, atol=1e-3)
def test_detailed():
t_dc = dc.array([[0, 1, 2], [3, 4, 5], [6, 7, 8], [9, 10, 11]])
t_dc1 = dc.array([[21, 22], [23, 24], [26, 27]])
t_dc[2]
t_dc[2, 1]
int(t_dc[2, 1])
t_dc[2:3, :]
t_dc[2]
t_dc[2, :]
t_dc[2:3, 1:2]
t_dc[1, ::2]
t_dc[1:2:1, 1:2]
t_dc[1:2:1, ...]
t_dc[..., 1]
t_dc[..., ::-2]
t_dc[1:2:-1, ::-2]
t_dc[2:3, 1:2] = 30
t_dc[2, 1] = 1
t_dc[0:3, 0:2] = t_dc1
t1 = dc.array(2, 3).asTypeFloat()
t2 = dc.array(2, 3).asTypeInt()
add = dc.add(t1, t1)
add = t1 + t1
