def test_main(self):
dtypes = [
'float32', 'float64', 'int32', 'int64', 'uint8', 'int8', 'bool'
]
places = [fluid.CPUPlace()]
if fluid.core.is_compiled_with_cuda():
places.append(fluid.CUDAPlace(0))
places.append(fluid.CUDAPinnedPlace())
for p in places:
for dtype in dtypes:
np_arr = np.reshape(
np.array(six.moves.range(np.prod(self.shape))).astype(
dtype), self.shape)
t = fluid.LoDTensor()
t.set(np_arr, p)
ret_np_arr = np.array(t)
self.assertEqual(np_arr.shape, ret_np_arr.shape)
self.assertEqual(np_arr.dtype, ret_np_arr.dtype)
all_equal = np.all(np_arr == ret_np_arr)
self.assertTrue(all_equal)
def test_input_cuda_pinned_var(self):
with fluid.dygraph.guard():
data = np.random.random((2, 80, 16128)).astype('float32')
var = core.VarBase(value=data,
name='',
persistable=False,
place=fluid.CUDAPinnedPlace(),
zero_copy=False)
sliced = var[:, 10:, :var.shape[1]]
self.assertEqual(sliced.shape, [2, 70, 80])
def _set_var(self, var, ndarray):
t = global_scope().find_var(var.name).get_tensor()
p = t._place()
if p.is_cpu_place():
place = fluid.CPUPlace()
elif p.is_cuda_pinned_place():
place = fluid.CUDAPinnedPlace()
else:
p = fluid.core.Place()
p.set_place(t._place())
place = fluid.CUDAPlace(p.gpu_device_id())
t.set(ndarray, place)
def func_tensor_tolist(self):
places = [fluid.CPUPlace()]
if fluid.core.is_compiled_with_cuda():
places.append(fluid.CUDAPlace(0))
places.append(fluid.CUDAPinnedPlace())
for p in places:
np_arr = np.reshape(
np.array(six.moves.range(np.prod(self.shape))), self.shape)
expectlist = np_arr.tolist()
t = paddle.to_tensor(np_arr, place=p)
tensorlist = t.tolist()
self.assertEqual(tensorlist, expectlist)
def test_tensor_fill_true(self):
typelist = ['float32', 'float64', 'int32', 'int64', 'float16']
places = [fluid.CPUPlace()]
if fluid.core.is_compiled_with_cuda():
places.append(fluid.CUDAPlace(0))
places.append(fluid.CUDAPinnedPlace())
for p in places:
np_arr = np.reshape(
np.array(six.moves.range(np.prod(self.shape))), self.shape)
for dtype in typelist:
tensor = paddle.to_tensor(np_arr, place=p, dtype=dtype)
target = tensor.numpy()
target[...] = 0
tensor.zero_()
self.assertEqual((tensor.numpy() == target).all().item(), True)
def _feed_random_data(self, use_gpu, as_lodtensor=False):
print("feed random data")
feed = {}
if use_gpu and as_lodtensor:
#place = fluid.CPUPlace()
place = fluid.CUDAPinnedPlace()
for var in self.feed_vars:
if var.type != fluid.core.VarDesc.VarType.LOD_TENSOR:
raise TypeError("Feed data of non LoDTensor is not supported.")
shape = var.shape
dtype = self.convert_dtype(var.dtype, to_string=True)
data = np.random.random(shape).astype(dtype)
if use_gpu and as_lodtensor:
tensor = fluid.core.LoDTensor()
tensor.set(data, place)
feed[var.name] = tensor
else:
feed[var.name] = data
return feed
def func_test_tensor_fill_true(self):
typelist = ['float32', 'float64', 'int32', 'int64', 'float16']
places = [fluid.CPUPlace()]
if fluid.core.is_compiled_with_cuda():
places.append(fluid.CUDAPlace(0))
places.append(fluid.CUDAPinnedPlace())
for idx, p in enumerate(places):
if idx == 0:
paddle.set_device('cpu')
else:
paddle.set_device('gpu')
np_arr = np.reshape(
np.array(six.moves.range(np.prod(self.shape))), self.shape)
for dtype in typelist:
var = 1.
tensor = paddle.to_tensor(np_arr, place=p, dtype=dtype)
target = tensor.numpy()
target[...] = var
tensor.fill_(var) #var type is basic type in typelist
self.assertEqual((tensor.numpy() == target).all(), True)
def func_test_tensor_fill_backward(self):
typelist = ['float32']
places = [fluid.CPUPlace()]
if fluid.core.is_compiled_with_cuda():
places.append(fluid.CUDAPlace(0))
places.append(fluid.CUDAPinnedPlace())
for idx, p in enumerate(places):
if idx == 0:
paddle.set_device('cpu')
else:
paddle.set_device('gpu')
np_arr = np.reshape(
np.array(six.moves.range(np.prod(self.shape))), self.shape)
for dtype in typelist:
var = int(1)
tensor = paddle.to_tensor(np_arr, place=p, dtype=dtype)
tensor.stop_gradient = False
y = tensor * 2
y.fill_(var)
loss = y.sum()
loss.backward()
self.assertEqual((y.grad.numpy() == 0).all().item(), True)