def check_gradients_vector(forward_fcn,
param_dictionary,
rtol=1e-5,
atol=1e-8,
eps=1e-3):
params = [
sail.Tensor(param_dictionary[a], requires_grad=True)
for a in param_dictionary
]
output = forward_fcn(*params)
output.backward()
grads_dic = {}
for i, p in enumerate(param_dictionary):
grads_dic[p] = params[i].grad.numpy()
parameters = dictionary_to_vector(param_dictionary)
grads = dictionary_to_vector(grads_dic)
num_params = len(parameters)
j_plus = np.zeros(num_params)
j_minus = np.zeros(num_params)
grad_approx = np.zeros(num_params)
for i in range(len(parameters)):
# pass
params_plus = np.copy(parameters)
params_plus[i] += eps
params_plus_dict = vector_to_dictionary(params_plus, param_dictionary)
params_plus_ = [
sail.Tensor(params_plus_dict[a]) for a in params_plus_dict
]
z = forward_fcn(*params_plus_)
j_plus[i] = z.numpy()
params_minus = np.copy(parameters)
params_minus[i] -= eps
params_minus_dict = vector_to_dictionary(params_minus,
param_dictionary)
params_minus_ = [
sail.Tensor(params_minus_dict[a]) for a in params_minus_dict
]
z = forward_fcn(*params_minus_)
j_minus[i] = z.numpy()
grad_approx[i] = (j_plus[i] - j_minus[i]) / (2 * eps)
return np.allclose(np.array(grad_approx),
np.array(grads),
rtol=rtol,
atol=atol)
def test_base(self):
choices = [(12, 13, 3), (2, 5), (18, 15), (1, 3, 2, 3), (2,)]
times = []
for c in choices:
for i in range(-len(c)-1, len(c)):
c = list(c)
arr1 = np.random.uniform(0, 1, (c))
x1 = sail.Tensor(arr1, requires_grad=False)
arr3 = np.expand_dims(arr1, i)
t = time.time()
x3 = sail.expand_dims(x1, i)
times.append(time.time() - t)
self.assert_eq_np_sail(arr3, x3)
self.assert_eq(arr3.shape, x3.shape)
c_ = list(c)
a = i
if (i < 0):
i += len(c) + 1
if (i == len(c)):
c_.append(1)
else:
c_.insert(i, 1)
self.assert_eq(list(x3.shape), c_)
return
def test_base(self, rq):
choices = elementwise_options
times = []
for c in choices:
arr1 = np.random.uniform(0, 1, (c)).astype(np.float32)
arr2 = np.random.uniform(0, 1, (c)).astype(np.float32)
x1 = sail.Tensor(arr1, requires_grad=rq)
x2 = sail.Tensor(arr2, requires_grad=rq)
t = time.time()
x3 = sail.subtract(x1, x2)
times.append(time.time() - t)
arr3 = arr1 - arr2
self.assert_eq_np_sail(arr3, x3)
self.assert_eq(x3.requires_grad, rq)
return
def test_base(self, rq):
choices = elementwise_options
times = []
for c in choices:
arr1 = np.random.uniform(1, 2, (c))
arr2 = np.random.uniform(1, 2, (c))
x1 = sail.Tensor(arr1, requires_grad=rq)
x2 = sail.Tensor(arr2, requires_grad=rq)
t = time.time()
x3 = sail.power(x1, x2)
times.append(time.time() - t)
arr3 = np.power(arr1, arr2)
self.assert_eq_np_sail(arr3, x3)
self.assert_eq(x3.requires_grad, rq)
return
def test(self, rq):
choices = [(3, 3), (12, 18), (2, 33), (32, 64)]
times = []
for c in choices:
arr1 = np.random.uniform(0, 1, c).astype(np.float32)
targs = np.random.uniform(0, 1, c).astype(np.float32)
x1 = sail.Tensor(arr1, requires_grad=rq)
x2 = sail.Tensor(targs, requires_grad=rq)
lin = sail.losses.MeanSquaredError()
loss_sail = lin(x1, x2)
loss_numpy = np.mean((arr1 - targs) ** 2)
self.assert_eq_np(loss_sail.numpy(), loss_numpy, eps=1e-5)
return
def test_broadcast(self, rq):
choices = broadcasted_options
times = []
for c in choices:
c = list(c)
for i in range(len(c)):
arr1 = np.random.uniform(0, 1, (c))
c[i] = 1
arr2 = np.random.uniform(0, 1, (c))
x1 = sail.Tensor(arr1, requires_grad=rq)
x2 = sail.Tensor(arr2, requires_grad=rq)
t = time.time()
x3 = sail.subtract(x1, x2)
times.append(time.time() - t)
arr3 = arr1 - arr2
self.assert_eq_np_sail(arr3, x3)
self.assert_eq(x3.requires_grad, rq)
return
def test_base(self, rq):
choices = unary_elementwise_options
for c in choices:
arr1 = np.random.normal(5, 2, (c))
x1 = sail.Tensor(arr1, requires_grad=rq)
arr2 = np.clip(arr1, 4, 6)
x2 = sail.clip(x1, 4, 6)
self.assert_eq_np_sail(arr2, x2)
self.assert_eq(x2.requires_grad, rq)
return
def test_integer(self):
choices = unary_elementwise_options
times = []
for c in choices:
arr1 = np.random.uniform(0, 1, (c)).astype(np.int32)
x1 = sail.Tensor(arr1)
x3 = sail.exp(x1)
arr3 = np.exp(arr1)
self.assert_eq_np_sail(arr3, x3, eps=1e-6)
return
def test_integer(self):
choices = unary_elementwise_options
times = []
for c in choices:
arr1 = np.random.uniform(1, 50000, (c)).astype(np.int16)
x1 = sail.Tensor(arr1)
x3 = sail.log(x1)
arr3 = np.log(arr1)
self.assert_eq_np_sail(arr3.astype(x3.numpy().dtype), x3, eps=1e-1)
return
def test_base(self, rq):
choices = [(3, 3), (12, 18), (2, 33), (32, 64)]
choices_2 = [[(3, 3), (3, 1), (3, 10)], [(18, 12), (18, 2)],
[(33, 1), (33, 33)], [(64, 12)]]
times = []
for ca, cbs in zip(choices, choices_2):
for cb in cbs:
verif_shape = (ca[0], cb[1])
arr1 = np.random.uniform(0, 1, (ca))
arr2 = np.random.uniform(0, 1, (cb))
x1 = sail.Tensor(arr1, requires_grad=rq)
x2 = sail.Tensor(arr2, requires_grad=rq)
x3 = sail.matmul(x1, x2)
arr3 = np.matmul(arr1, arr2)
self.assert_eq(x3.shape, verif_shape)
self.assert_eq_np_sail(arr3, x3, eps=1e-7)
self.assert_eq(x3.requires_grad, rq)
return
def test(self):
choices = [(3, 3), (12, 18), (2, 33), (32, 64)]
times = []
for c in choices:
arr1 = np.random.uniform(0, 1, c).astype(np.float32)
targs = np.random.randint(0, c[1], c[0]).astype(np.int64)
x1 = sail.Tensor(arr1, requires_grad=False)
x2 = sail.Tensor(targs, requires_grad=False)
lin = sail.losses.SoftmaxCrossEntropy()
loss_sail = lin(x1, x2)
t_loss_fcn = torch.nn.CrossEntropyLoss()
t1 = torch.from_numpy(arr1)
t2 = torch.from_numpy(targs)
loss_torch = t_loss_fcn(t1, t2)
self.assert_eq_np(loss_sail.numpy(), loss_torch.detach().numpy(), eps=1e-5)
return
def test_base_int(self, rq):
choices = elementwise_options
times = []
for c in choices:
arr1 = np.random.uniform(0, 1, (c)) #.astype(np.float32)
x1 = sail.Tensor(arr1, requires_grad=rq)
x3 = 1 - x1 - 2
arr3 = 1 - arr1 - 2
self.assert_eq_np_sail(arr3, x3, eps=1e-7)
self.assert_eq(x3.requires_grad, rq)
return
def test_min(self, rq):
times = []
for c in choices:
arr1 = np.random.uniform(0, 1, (c["shape"]))
x1 = sail.Tensor(arr1, requires_grad=rq)
x3 = sail.min(x1, c["axis"], keepdims=c["keepdims"])
arr3 = np.min(arr1, c["axis"], keepdims=c["keepdims"])
self.assert_eq(x3.shape, c["result_shape"])
self.assert_eq_np_sail(arr3, x3, 1e-7)
self.assert_eq(x3.requires_grad, rq)
return
def test_base_float(self, rq):
choices = elementwise_options
times = []
for c in choices:
arr1 = np.random.uniform(0, 1, (c)) #.astype(np.float32)
x1 = sail.Tensor(arr1, requires_grad=rq)
x3 = 1.2 / x1 / 1.3
arr3 = 1.2 / arr1 / 1.3
self.assert_eq_np_sail(arr3, x3)
self.assert_eq(x3.requires_grad, rq)
return
def test_base(self, rq):
choices = unary_elementwise_options
times = []
for c in choices:
arr1 = np.random.uniform(1, 2, (c))
x1 = sail.Tensor(arr1, requires_grad=rq)
t = time.time()
x3 = sail.log(x1)
times.append(time.time() - t)
arr3 = np.log(arr1)
self.assert_eq_np_sail(arr3, x3, eps=1e-6)
self.assert_eq(x3.requires_grad, rq)
return
def test(self, rq):
choices = [(3, 3), (12, 18), (2, 33), (32, 64)]
times = []
for c in choices:
arr1 = np.random.uniform(-1, 1, c).astype(np.float32)
x1 = sail.Tensor(arr1, requires_grad=rq)
lin = sail.modules.ReLU()
y = lin(x1)
y2 = np.clip(arr1, 0, 100)
self.assert_eq_np_sail(y2, y, eps=1e-8)
self.assert_eq(y.requires_grad, rq)
return
def test(self, rq):
choices = [(3, 3), (12, 18), (2, 33), (32, 64)]
times = []
for c in choices:
arr1 = np.random.uniform(0, 1, c).astype(np.float32)
x1 = sail.Tensor(arr1, requires_grad=rq)
lin = sail.modules.Softmax()
y = lin(x1)
y2 = np.exp(arr1) / np.sum(np.exp(arr1), 1, keepdims=True)
self.assert_eq_np_sail(y2, y, eps=5e-6)
self.assert_eq(y.requires_grad, rq)
return
def test_bias(self):
choices = [(3, 3), (12, 18), (2, 33), (32, 64)]
choices_2 = [[(3, 3), (3, 1), (3, 10)], [(18, 12), (18, 2)], [(33, 1), (33, 33)], [(64, 12)]]
times = []
for ca, cbs in zip(choices, choices_2):
for cb in cbs:
verif_shape = (ca[0], cb[1])
arr1 = np.random.uniform(0, 1, (ca)).astype(np.float32)
x1 = sail.Tensor(arr1, requires_grad=False)
lin = sail.modules.Linear(cb[0], cb[1], use_bias=True)
y = lin(x1)
y2 = np.matmul(arr1, lin.weights.numpy()) + lin.biases.numpy()
self.assert_eq_np_sail(y2, y, eps=5e-6)
self.assert_eq(y.requires_grad, True)
return
def test_base_reshape(self):
choices = [(12, 13, 3), (2, 5), (18, 15), (1, 3, 2, 3), (2,)]
times = []
for c in choices:
for i in range(len(c)):
c_ = list(c)
random.shuffle(c_)
arr1 = np.random.uniform(0, 1, (c_))
x1 = sail.Tensor(arr1, requires_grad=False)
t = time.time()
x3 = sail.reshape(x1, c)
times.append(time.time() - t)
arr3 = np.reshape(arr1, c)
self.assert_eq_np_sail(arr3, x3)
self.assert_eq(x3.shape, c)
return
def test_catch_error(self):
choices = [(12, 13, 3), (2, 5), (18, 15), (1, 3, 2, 3), (2,)]
times = []
for c in choices:
arr1 = np.random.uniform(0, 1, (c))
x1 = sail.Tensor(arr1, requires_grad=False)
try:
x3 = sail.expand_dims(x1, 10)
except sail.DimensionError as e:
try:
x3 = sail.expand_dims(x1, -10)
except sail.DimensionError as e:
self.assert_eq(True, True)
return
self.assert_eq(False, True)
return
def test_broadcast_float(self, rq):
choices = broadcasted_options
times = []
for c in choices:
for i in range(len(c)):
b = list(c)
b[i] = 1
arr1 = np.random.uniform(0, 1, (b)) #.astype(np.float32)
x1 = sail.Tensor(arr1, requires_grad=rq)
x1 = sail.broadcast_to(x1, c)
arr1 = np.broadcast_to(arr1, c)
x3 = 1.1 - x1 - 2.2
arr3 = 1.1 - arr1 - 2.2
self.assert_eq_np_sail(arr3, x3)
self.assert_eq(x3.requires_grad, rq)
return
def test_catch_error(self):
choices = [(12, 13, 3), (2, 5), (18, 15), (1, 3, 2, 3), (2,)]
times = []
for c in choices:
c_ = list(c)
random.shuffle(c_)
c_[-1] = c_[-1] * 2
arr1 = np.random.uniform(0, 1, (c))
x1 = sail.Tensor(arr1, requires_grad=False)
try:
x3 = sail.reshape(x1, c_)
except sail.DimensionError as e:
self.assert_eq(True, True)
return
self.assert_eq(False, True)
return
def test_broadcast(self):
choices = unary_broadcasted_options
times = []
for c in choices:
for i in range(len(c)):
b = list(c)
b[i] = 1
arr1 = np.random.uniform(1, 2, (b))
x1 = sail.Tensor(arr1, requires_grad=False)
x2 = sail.broadcast_to(x1, c)
arr2 = np.broadcast_to(arr1, c)
t = time.time()
x3 = -x2
times.append(time.time() - t)
arr3 = -arr2
self.assert_eq_np_sail(arr3, x3, eps=1e-6)
return
def test_base(self):
choices = [(12, 13, 3), (2, 5), (18, 15), (1, 3, 2, 3)]
times = []
for c in choices:
for i in range(len(c)):
cr = list(range(len(c)))
random.shuffle(cr)
fixed_shape = [c[r] for r in cr]
arr1 = np.random.uniform(0, 1, (c))
x1 = sail.Tensor(arr1, requires_grad=False)
x3 = sail.transpose(x1, cr)
arr3 = np.transpose(arr1, cr)
self.assert_eq_np_sail(arr3, x3)
self.assert_eq(list(x3.shape), fixed_shape)
return
def test_base(self):
choices = [(12, 13, 3), (2, 5), (18, 15), (1, 3, 2, 3), (2,)]
arrs = []
for c in choices:
temp = []
for i in range(-len(c)-1, len(c)):
x = np.expand_dims(np.random.uniform(0, 1, (c)), i)
temp.append(x)
arrs.append(temp)
times = []
for a, c in zip(arrs, choices):
z = 0
for i in range(-a[0].ndim, a[0].ndim-1):
arr1 = a[z]
x1 = sail.Tensor(arr1, requires_grad=False)
arr3 = np.squeeze(arr1, i)
x3 = sail.squeeze(x1, i)
self.assert_eq_np_sail(arr3, x3)
self.assert_eq(arr3.shape, x3.shape)
z += 1
return
def test_negative_print(self):
x = np.ones((32, 32))
x *= -0.001
x[16:] *= 1.3
comparison = '''
tensor([[-0.00100000 -0.00100000 -0.00100000 ... -0.00100000
-0.00100000 -0.00100000]
[-0.00100000 -0.00100000 -0.00100000 ... -0.00100000
-0.00100000 -0.00100000]
[-0.00100000 -0.00100000 -0.00100000 ... -0.00100000
-0.00100000 -0.00100000]
...
[-0.0013 -0.0013 -0.0013 ... -0.0013
-0.0013 -0.0013 ]
[-0.0013 -0.0013 -0.0013 ... -0.0013
-0.0013 -0.0013 ]
[-0.0013 -0.0013 -0.0013 ... -0.0013
-0.0013 -0.0013 ]], shape=(32, 32))'''
comparison = comparison[1:]
comparison = comparison
x = sail.Tensor(x)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
x = np.ones((32)).astype(np.float64)
x *= -0.00000002
x[16:] *= 1.3
comparison = '''
tensor([-0.00000002 -0.00000002 -0.00000002 -0.00000002 -0.00000002 -0.00000002 -0.00000002 -0.00000002 -0.00000002 -0.00000002
-0.00000002 -0.00000002 -0.00000002 -0.00000002 -0.00000002 -0.00000002 -0.00000003 -0.00000003 -0.00000003 -0.00000003
-0.00000003 -0.00000003 -0.00000003 -0.00000003 -0.00000003 -0.00000003 -0.00000003 -0.00000003 -0.00000003 -0.00000003
-0.00000003 -0.00000003], shape=(32))'''
comparison = comparison[1:]
x = sail.Tensor(x)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
x = np.ones((32)).astype(np.float64)
x *= -100000000000
x[16:] *= 1.3
comparison = '''
tensor([-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11
-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.30000000e+11 -1.30000000e+11 -1.30000000e+11 -1.30000000e+11
-1.30000000e+11 -1.30000000e+11 -1.30000000e+11 -1.30000000e+11 -1.30000000e+11 -1.30000000e+11 -1.30000000e+11 -1.30000000e+11 -1.30000000e+11 -1.30000000e+11
-1.30000000e+11 -1.30000000e+11], shape=(32))'''
comparison = comparison[1:]
x = sail.Tensor(x)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
x = np.ones((32, 12, 4)).astype(np.float64)
x *= -100000000000
x[16:] *= 0.00000002
comparison = '''
tensor([[[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
...
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]]
[[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
...
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]]
[[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
...
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]
[-1.00000000e+11 -1.00000000e+11 -1.00000000e+11 -1.00000000e+11]]
...
[[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
...
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]]
[[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
...
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]]
[[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
...
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]
[-2.00000000e+03 -2.00000000e+03 -2.00000000e+03 -2.00000000e+03]]], shape=(32, 12, 4))'''
comparison = comparison[1:]
x = sail.Tensor(x)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
def test_base(self):
choices = [
{
'a_shape': (4, 3, 2),
'b_shape': (3, 2, 5),
'axes': 2,
'gc_shape': (4, 5)
}, # NOQA
{
'a_shape': (4, 3, 2),
'b_shape': (3, 2, 5),
'axes': ([1, 2], [0, 1]),
'gc_shape': (4, 5)
}, # NOQA
{
'a_shape': (4, 2, 3),
'b_shape': (3, 5, 2),
'axes': ([2, 1], [0, 2]),
'gc_shape': (4, 5)
}, # NOQA
{
'a_shape': (2, 4, 3),
'b_shape': (5, 3, 2),
'axes': ([2, 0], [1, 2]),
'gc_shape': (4, 5)
}, # NOQA
{
'a_shape': (2, 3, 4),
'b_shape': (5, 2, 3),
'axes': ([1, 0], [2, 1]),
'gc_shape': (4, 5)
}, # NOQA
{
'a_shape': (3, 2, 4),
'b_shape': (2, 5, 3),
'axes': ([0, 1], [2, 0]),
'gc_shape': (4, 5)
}, # NOQA
{
'a_shape': (3, 4, 2),
'b_shape': (2, 3, 5),
'axes': ([0, 2], [1, 0]),
'gc_shape': (4, 5)
}, # NOQA
{
'a_shape': (3, 4, 2),
'b_shape': (2, 5, 6),
'axes': 1,
'gc_shape': (3, 4, 5, 6)
}, # NOQA
{
'a_shape': (3, 4, 2),
'b_shape': (2, 5, 6),
'axes': ([2], [0]),
'gc_shape': (3, 4, 5, 6)
}, # NOQA
{
'a_shape': (3, 2, 4),
'b_shape': (5, 2, 6),
'axes': ([1], [1]),
'gc_shape': (3, 4, 5, 6)
}, # NOQA
{
'a_shape': (2, 3, 4),
'b_shape': (5, 6, 2),
'axes': ([0], [2]),
'gc_shape': (3, 4, 5, 6)
}, # NOQA
{
'a_shape': (4, 5, 3, 2),
'b_shape': (3, 2, 6),
'axes': 2,
'gc_shape': (4, 5, 6)
}, # NOQA
{
'a_shape': (4, 5, 3, 2),
'b_shape': (3, 2, 6),
'axes': ([2, 3], [0, 1]),
'gc_shape': (4, 5, 6)
}, # NOQA
{
'a_shape': (4, 5, 2, 3),
'b_shape': (3, 6, 2),
'axes': ([3, 2], [0, 2]),
'gc_shape': (4, 5, 6)
}, # NOQA
{
'a_shape': (4, 2, 5, 3),
'b_shape': (6, 3, 2),
'axes': ([3, 1], [1, 2]),
'gc_shape': (4, 5, 6)
}, # NOQA
{
'a_shape': (2, 4, 5, 3),
'b_shape': (6, 2, 3),
'axes': ([3, 0], [2, 1]),
'gc_shape': (4, 5, 6)
}, # NOQA
{
'a_shape': (2, 4, 3, 5),
'b_shape': (2, 6, 3),
'axes': ([2, 0], [2, 0]),
'gc_shape': (4, 5, 6)
}, # NOQA
{
'a_shape': (2, 3, 4, 5),
'b_shape': (2, 3, 6),
'axes': ([1, 0], [1, 0]),
'gc_shape': (4, 5, 6)
}, # NOQA
{
'a_shape': (3, 2, 4, 5),
'b_shape': (3, 2, 6),
'axes': ([0, 1], [0, 1]),
'gc_shape': (4, 5, 6)
}, # NOQA
{
'a_shape': (3, 2, 5, 4),
'b_shape': (3, 6, 2),
'axes': ([0, 1], [0, 2]),
'gc_shape': (5, 4, 6)
}, # NOQA
{
'a_shape': (3, 5, 2, 4),
'b_shape': (6, 3, 2),
'axes': ([0, 2], [1, 2]),
'gc_shape': (5, 4, 6)
}, # NOQA
{
'a_shape': (5, 3, 2, 4),
'b_shape': (6, 2, 3),
'axes': ([1, 2], [2, 1]),
'gc_shape': (5, 4, 6)
}, # NOQA
{
'a_shape': (5, 4, 3, 2),
'b_shape': (4, 3, 2, 6),
'axes': 3,
'gc_shape': (5, 6)
}, # NOQA
{
'a_shape': (5, 4, 3, 2),
'b_shape': (4, 3, 2, 6),
'axes': ([1, 2, 3], [0, 1, 2]),
'gc_shape': (5, 6)
}, # NOQA
{
'a_shape': (5, 4, 2, 3),
'b_shape': (4, 3, 6, 2),
'axes': ([1, 3, 2], [0, 1, 3]),
'gc_shape': (5, 6)
}, # NOQA
{
'a_shape': (5, 2, 4, 3),
'b_shape': (4, 6, 3, 2),
'axes': ([2, 3, 1], [0, 2, 3]),
'gc_shape': (5, 6)
}, # NOQA
{
'a_shape': (2, 5, 4, 3),
'b_shape': (4, 6, 2, 3),
'axes': ([2, 3, 0], [0, 3, 2]),
'gc_shape': (5, 6)
}, # NOQA
{
'a_shape': (2, 5, 3, 4),
'b_shape': (6, 4, 2, 3),
'axes': ([3, 2, 0], [1, 3, 2]),
'gc_shape': (5, 6)
}, # NOQA
{
'a_shape': (2, 3, 5, 4),
'b_shape': (6, 2, 4, 3),
'axes': ([3, 1, 0], [2, 3, 1]),
'gc_shape': (5, 6)
}, # NOQA
{
'a_shape': (3, 2, 5, 4),
'b_shape': (6, 2, 3, 4),
'axes': ([3, 0, 1], [3, 2, 1]),
'gc_shape': (5, 6)
}, # NOQA
{
'a_shape': (3, 2, 4, 5),
'b_shape': (2, 6, 3, 4),
'axes': ([2, 0, 1], [3, 2, 0]),
'gc_shape': (5, 6)
}, # NOQA
{
'a_shape': (3, 4, 2, 5),
'b_shape': (2, 3, 6, 4),
'axes': ([1, 0, 2], [3, 1, 0]),
'gc_shape': (5, 6)
}, # NOQA
{
'a_shape': (4, 3, 2, 5),
'b_shape': (2, 3, 4, 6),
'axes': ([0, 1, 2], [2, 1, 0]),
'gc_shape': (5, 6)
}, # NOQA
{
'a_shape': (4, 3, 5, 2),
'b_shape': (3, 2, 4, 6),
'axes': ([0, 1, 3], [2, 0, 1]),
'gc_shape': (5, 6)
}, # NOQA
{
'a_shape': (4, 5, 3, 2),
'b_shape': (3, 4, 2, 6),
'axes': ([0, 2, 3], [1, 0, 2]),
'gc_shape': (5, 6)
}, # NOQA
]
for c in choices:
arr1 = np.random.uniform(0, 1, c["a_shape"])
arr2 = np.random.uniform(0, 1, c["b_shape"])
x1 = sail.Tensor(arr1, requires_grad=False)
x2 = sail.Tensor(arr2, requires_grad=False)
arr3 = np.tensordot(arr1, arr2, axes=c["axes"])
x3 = sail.tensordot(x1, x2, axes=c["axes"])
self.assert_eq_np_sail(arr3, x3, eps=1e-7)
def test_general(self):
x = np.ones((32, 32)).astype(np.int32)
x *= 10
x[16:] *= 20
comparison = '''
tensor([[ 10 10 10 ... 10
10 10]
[ 10 10 10 ... 10
10 10]
[ 10 10 10 ... 10
10 10]
...
[200 200 200 ... 200
200 200]
[200 200 200 ... 200
200 200]
[200 200 200 ... 200
200 200]], shape=(32, 32))'''
comparison = comparison[1:]
comparison = comparison
x = sail.Tensor(x)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
x = np.ones((32)).astype(np.float64)
x *= -25.6
x[16:] *= 13.8
comparison = '''
tensor([ -25.6 -25.6 -25.6 -25.6 -25.6 -25.6 -25.6 -25.6 -25.6 -25.6
-25.6 -25.6 -25.6 -25.6 -25.6 -25.6 -353.28 -353.28 -353.28 -353.28
-353.28 -353.28 -353.28 -353.28 -353.28 -353.28 -353.28 -353.28 -353.28 -353.28
-353.28 -353.28], shape=(32))'''
comparison = comparison[1:]
x = sail.Tensor(x)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
x = np.ones((32)).astype(np.int32)
x *= -1000000000
comparison = '''
tensor([-1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000
-1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000
-1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000 -1000000000
-1000000000 -1000000000], shape=(32))'''
comparison = comparison[1:]
x = sail.Tensor(x)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
x = np.ones(1)
comparison = '''
tensor([[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]], shape=(10, 10))'''
comparison = comparison[1:]
x = sail.broadcast_to(sail.Tensor(x), (10, 10))
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
x = np.ones(10)
comparison = '''
tensor([[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]
[1. 1. 1. 1. 1. 1. 1. 1. 1. 1.]], shape=(10, 10))'''
comparison = comparison[1:]
x = sail.broadcast_to(sail.Tensor(x), (10, 10))
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
x = np.array([1.25234312345])
comparison = '''
tensor([[1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312]
[1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312]
[1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312]
[1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312]
[1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312]
[1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312]
[1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312]
[1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312]
[1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312]
[1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312 1.25234312]], shape=(10, 10))'''
comparison = comparison[1:]
x = sail.broadcast_to(sail.Tensor(x), (10, 10))
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
x = np.array([1.25234312345])
comparison = '''tensor([1.25234312], shape=(1))'''
x = sail.Tensor(x)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
comparison = '''
tensor([ True False True True True False False False True True
False False True True False False True True False True
True False False False False False True], shape=(27))'''
comparison = comparison[1:]
x = sail.Tensor(np.array([1, 0, 1, 1, 1, 0, 0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 0, 1, 1, 0, 0,0,0,0,1])).astype(sail.bool_)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
end = batch_size
step = batch_size
total_loss = 0
steps = 0
ar = np.arange(len(x_train))
np.random.shuffle(ar)
x_train = x_train[ar]
y_train = y_train[ar]
u_time = 0
while end < len(x_train):
x_batch = x_train[start:end]
y_batch = y_train[start:end]
x_batch = sail.Tensor(x_batch)
y_batch = sail.Tensor(y_batch)
preds = mnist(x_batch)
loss = loss_fcn(preds, y_batch)
total_loss += loss.numpy()[0]
loss.backward()
opt.update()
start += batch_size
end += batch_size
steps += 1
pred_test = mnist(sail.Tensor(x_test))
def test_print(self):
x = np.ones((32, 32))
x *= 0.001
x[16:] *= 1.3
comparison = '''
tensor([[0.00100000 0.00100000 0.00100000 ... 0.00100000
0.00100000 0.00100000]
[0.00100000 0.00100000 0.00100000 ... 0.00100000
0.00100000 0.00100000]
[0.00100000 0.00100000 0.00100000 ... 0.00100000
0.00100000 0.00100000]
...
[0.0013 0.0013 0.0013 ... 0.0013
0.0013 0.0013 ]
[0.0013 0.0013 0.0013 ... 0.0013
0.0013 0.0013 ]
[0.0013 0.0013 0.0013 ... 0.0013
0.0013 0.0013 ]], shape=(32, 32))'''
comparison = comparison[1:]
comparison = comparison
x = sail.Tensor(x)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
x = np.ones((32)).astype(np.float64)
x *= 0.00000002
x[16:] *= 1.3
comparison = '''
tensor([0.00000002 0.00000002 0.00000002 0.00000002 0.00000002 0.00000002 0.00000002 0.00000002 0.00000002 0.00000002
0.00000002 0.00000002 0.00000002 0.00000002 0.00000002 0.00000002 0.00000003 0.00000003 0.00000003 0.00000003
0.00000003 0.00000003 0.00000003 0.00000003 0.00000003 0.00000003 0.00000003 0.00000003 0.00000003 0.00000003
0.00000003 0.00000003], shape=(32))'''
comparison = comparison[1:]
x = sail.Tensor(x)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
x = np.ones((32)).astype(np.float64)
x *= 100000000000
x[16:] *= 1.3
comparison = '''
tensor([1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11
1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11 1.30000000e+11 1.30000000e+11 1.30000000e+11 1.30000000e+11
1.30000000e+11 1.30000000e+11 1.30000000e+11 1.30000000e+11 1.30000000e+11 1.30000000e+11 1.30000000e+11 1.30000000e+11 1.30000000e+11 1.30000000e+11
1.30000000e+11 1.30000000e+11], shape=(32))'''
comparison = comparison[1:]
x = sail.Tensor(x)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
x = np.ones((32, 12, 4)).astype(np.float64)
x *= 100000000000
x[16:] *= 0.00000002
comparison = '''
tensor([[[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
...
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]]
[[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
...
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]]
[[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
...
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]
[1.00000000e+11 1.00000000e+11 1.00000000e+11 1.00000000e+11]]
...
[[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
...
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]]
[[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
...
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]]
[[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
...
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]
[2.00000000e+03 2.00000000e+03 2.00000000e+03 2.00000000e+03]]], shape=(32, 12, 4))'''
comparison = comparison[1:]
x = sail.Tensor(x)
self.assert_eq(x.__repr__().replace(" ", ""), comparison.replace(" ", ""))
