def get_numeric_gradient(scope,
op,
inputs,
input_to_check,
output_names,
delta=0.005,
in_place=False):
set_input(scope, op, inputs, core.CPUPlace())
op.infer_shape(scope)
tensor_to_check = scope.find_var(input_to_check).get_tensor()
def product(dim):
return reduce(lambda a, b: a * b, dim, 1)
ctx = core.DeviceContext.create(core.CPUPlace())
def get_output():
sum = 0.0
for output_name in output_names:
op.run(scope, ctx)
sum += np.array(scope.find_var(output_name).get_tensor()).sum()
return sum
tensor_to_check = scope.find_var(input_to_check).get_tensor()
tensor_size = product(tensor_to_check.get_dims())
gradient_flat = np.zeros(shape=(tensor_size, ), dtype='float32')
# we only compute gradient of one element each time.
# we use a for loop to compute the gradient of every element.
for i in xrange(tensor_size):
if in_place:
set_input(scope, op, inputs, core.CPUPlace())
# get one input element throw it's index i.
origin = tensor_to_check.get_float_element(i)
# add delta to it, run op and then get the sum of the result tensor.
x_pos = origin + delta
tensor_to_check.set_float_element(i, x_pos)
y_pos = get_output()
if in_place:
set_input(scope, op, inputs, core.CPUPlace())
x_neg = origin - delta
tensor_to_check.set_float_element(i, x_neg)
y_neg = get_output()
tensor_to_check.set_float_element(i, origin)
gradient_flat[i] = (y_pos - y_neg) / delta / 2
return gradient_flat.reshape(tensor_to_check.get_dims())
def check_grad(self,
inputs_to_check,
output_names,
no_grad_set=None,
in_place=False,
max_relative_error=0.005):
self.scope = core.Scope()
op_inputs = self.inputs if hasattr(self, "inputs") else dict()
op_attrs = self.attrs if hasattr(self, "attrs") else dict()
self.op = create_op(self.scope, self.op_type, op_inputs, self.outputs,
op_attrs)
if no_grad_set is None:
no_grad_set = set()
if not type(output_names) is list:
output_names = [output_names]
numeric_grads = [
get_numeric_gradient(self.scope,
self.op,
self.inputs,
input_to_check,
output_names,
in_place=in_place)
for input_to_check in inputs_to_check
]
grad_names = [
grad_var_name(input_to_check) for input_to_check in inputs_to_check
]
cpu_place = core.CPUPlace()
cpu_analytic_grads = [
get_gradient(self.scope, self.op, self.inputs, self.outputs,
grad_name, cpu_place, no_grad_set)
for grad_name in grad_names
]
self.__assert_is_close(numeric_grads, cpu_analytic_grads, grad_names,
max_relative_error,
"Gradient Check On %s" % str(cpu_place))
if core.is_compile_gpu() and self.op.support_gpu():
gpu_place = core.GPUPlace(0)
gpu_analytic_grads = [
get_gradient(self.scope, self.op, self.inputs, self.outputs,
grad_name, gpu_place, no_grad_set)
for grad_name in grad_names
]
self.__assert_is_close(numeric_grads, gpu_analytic_grads,
grad_names, max_relative_error,
"Gradient Check On %s" % str(gpu_place))
for c_grad, g_grad, name in itertools.izip(cpu_analytic_grads,
gpu_analytic_grads,
grad_names):
self.assertTrue(np.allclose(c_grad, g_grad, atol=1e-4),
"output name: " + name + " has diff")
def forward(self):
self.scope = core.Scope()
self.create_global_variables()
self.create_rnn_op()
self.create_step_net()
ctx = core.DeviceContext.create(core.CPUPlace())
self.rnnop.infer_shape(self.scope)
self.rnnop.run(self.scope, ctx)
return np.array(self.scope.find_var("h").get_tensor())
def test_lod_tensor_init(self):
scope = core.Scope()
place = core.CPUPlace()
lod_py = [[0, 2, 5], [0, 2, 4, 5]]
lod_tensor = core.LoDTensor(lod_py)
lod_tensor.set_dims([5, 2, 3, 4])
lod_tensor.alloc_float(place)
tensor_array = numpy.array(lod_tensor)
tensor_array[0, 0, 0, 0] = 1.0
tensor_array[0, 0, 0, 1] = 2.0
lod_tensor.set(tensor_array, place)
lod_v = numpy.array(lod_tensor)
self.assertAlmostEqual(1.0, lod_v[0, 0, 0, 0])
self.assertAlmostEqual(2.0, lod_v[0, 0, 0, 1])
self.assertListEqual(lod_py, lod_tensor.lod())
def test_all(self):
scope = core.Scope()
kwargs = dict()
places = [core.CPUPlace()]
if core.is_compile_gpu():
places.append(core.GPUPlace(0))
for place in places:
for in_name in Operator.get_op_input_names(self.type):
if hasattr(self, "inputs") and in_name in self.inputs:
kwargs[in_name] = in_name
var = scope.new_var(in_name).get_tensor()
arr = self.inputs[in_name]
var.set_dims(arr.shape)
var.set(arr, place)
else:
kwargs[in_name] = "@EMPTY@"
for out_name in Operator.get_op_output_names(self.type):
if not hasattr(self, "outputs"):
raise ValueError(
"The test op must set self.outputs dict.")
if out_name not in self.outputs:
raise ValueError(
"The %s is not in self.outputs dict." % (out_name))
kwargs[out_name] = out_name
scope.new_var(out_name).get_tensor()
for attr_name in Operator.get_op_attr_names(self.type):
if hasattr(self, "attrs") and attr_name in self.attrs:
kwargs[attr_name] = self.attrs[attr_name]
op = Operator(self.type, **kwargs)
if isinstance(place, core.GPUPlace) and not op.support_gpu():
return
op.infer_shape(scope)
ctx = core.DeviceContext.create(place)
op.run(scope, ctx)
for out_name in Operator.get_op_output_names(self.type):
actual = numpy.array(scope.find_var(out_name).get_tensor())
expect = self.outputs[out_name]
self.assertTrue(numpy.allclose(actual, expect, atol=1e-05),
"output name: " + out_name + "has diff")
def test_int_tensor(self):
scope = core.Scope()
var = scope.new_var("test_tensor")
place = core.CPUPlace()
tensor = var.get_tensor()
tensor.set_dims([1000, 784])
tensor.alloc_int(place)
tensor_array = numpy.array(tensor)
self.assertEqual((1000, 784), tensor_array.shape)
tensor_array[3, 9] = 1
tensor_array[19, 11] = 2
tensor.set(tensor_array, place)
tensor_array_2 = numpy.array(tensor)
self.assertEqual(1, tensor_array_2[3, 9])
self.assertEqual(2, tensor_array_2[19, 11])
def test_int_lod_tensor(self):
place = core.CPUPlace()
scope = core.Scope()
var_lod = scope.new_var("test_lod_tensor")
lod_tensor = var_lod.get_tensor()
lod_tensor.set_dims([4, 4, 6])
lod_tensor.alloc_int(place)
array = numpy.array(lod_tensor)
array[0, 0, 0] = 3
array[3, 3, 5] = 10
lod_tensor.set(array, place)
lod_tensor.set_lod([[0, 2, 4]])
lod_v = numpy.array(lod_tensor)
self.assertTrue(numpy.alltrue(array == lod_v))
lod = lod_tensor.lod()
self.assertEqual(0, lod[0][0])
self.assertEqual(2, lod[0][1])
self.assertEqual(4, lod[0][2])
def test_float_lod_tensor(self):
place = core.CPUPlace()
scope = core.Scope()
var_lod = scope.new_var("test_lod_tensor")
lod_tensor = var_lod.get_tensor()
lod_tensor.set_dims([5, 2, 3, 4])
lod_tensor.alloc_float(place)
tensor_array = numpy.array(lod_tensor)
self.assertEqual((5, 2, 3, 4), tensor_array.shape)
tensor_array[0, 0, 0, 0] = 1.0
tensor_array[0, 0, 0, 1] = 2.0
lod_tensor.set(tensor_array, place)
lod_v = numpy.array(lod_tensor)
self.assertAlmostEqual(1.0, lod_v[0, 0, 0, 0])
self.assertAlmostEqual(2.0, lod_v[0, 0, 0, 1])
self.assertEqual(len(lod_tensor.lod()), 0)
lod_py = [[0, 2, 5], [0, 2, 4, 5]]
lod_tensor.set_lod(lod_py)
lod = lod_tensor.lod()
self.assertListEqual(lod_py, lod)
def test_cpu(self):
self.gaussian_random_test(place=core.CPUPlace())
def create_tensor(scope, name, shape, np_data):
tensor = scope.new_var(name).get_tensor()
tensor.set_dims(shape)
tensor.set(np_data, core.CPUPlace())
return tensor
def check_output(self):
places = [core.CPUPlace()]
if core.is_compile_gpu():
places.append(core.GPUPlace(0))
for place in places:
self.check_output_with_place(place)
import paddle.v2.framework.core as core
from paddle.v2.framework.op import Operator
import numpy
import paddle.v2 as paddle
BATCH_SIZE = 100
scope = core.Scope()
place = core.CPUPlace()
# if you want to test GPU training, you can use gpu place
# place = core.GPUPlace(0)
dev_ctx = core.DeviceContext.create(place)
init_net = core.Net.create()
forward_net = core.Net.create()
backward_net = None
optimize_net = core.Net.create()
def atomic_id():
id = 0
while True:
yield id
id += 1
uniq_id = atomic_id().next
def data_layer(name, dims):
var = scope.new_var(name)
def test_uniform_random_cpu(self):
self.uniform_random_test(place=core.CPUPlace())