def check_with_place(self, place):
scope = core.Scope()
# create and initialize Id Variable
ids_tensor = scope.var('Ids').get_tensor()
ids_array = np.array([[0], [4], [3], [5]]).astype("int64")
ids_tensor.set(ids_array, place)
# create and initialize W Variable
rows = [0, 1, 2, 3, 4, 5, 6]
row_numel = 12
w_selected_rows = scope.var('W').get_selected_rows()
w_selected_rows.set_height(len(rows))
w_selected_rows.set_rows(rows)
w_array = np.ones((len(rows), row_numel)).astype("float32")
for i in range(len(rows)):
w_array[i] *= i
w_tensor = w_selected_rows.get_tensor()
w_tensor.set(w_array, place)
# create Out Variable
out_tensor = scope.var('Out').get_tensor()
# create and run lookup_table operator
lookup_table = Operator("lookup_table", W='W', Ids='Ids', Out='Out')
lookup_table.run(scope, place)
# get result from Out
result_array = np.array(out_tensor)
# all(): return True if all elements of the iterable are true (or if the iterable is empty)
for idx, row in enumerate(ids_array):
assert (row[0] == result_array[idx]).all()
def check_with_place(self, place, lazy_mode):
scope = core.Scope()
self.setup(scope, place, lazy_mode)
op_args = dict()
op_args['lazy_mode'] = lazy_mode
for key, np_array in self.dense_inputs.items():
var = scope.var(key).get_tensor()
var.set(np_array, place)
op_args[key] = key
for s in self.sparse_inputs:
op_args[s] = s
for s in self.outputs:
var = scope.var(s).get_tensor()
var.set(self.init_output, place)
op_args[s] = s
for k in self.attrs:
op_args[k] = self.attrs[k]
# create and run sgd operator
adam_op = Operator("adam", **op_args)
adam_op.run(scope, place)
for key, np_array in self.outputs.items():
out_var = scope.var(key).get_tensor()
actual = np.array(out_var)
actual = actual.reshape([actual.size])
np_array = np_array.reshape([np_array.size])
for i in range(np_array.size):
self.assertLess((actual[i] - np_array[i]), 0.00001)
def check_with_place(self, place):
scope = core.Scope()
x_rows = [0, 5, 5, 4, 19]
height = 20
row_numel = 2
np_array = np.ones((len(x_rows), row_numel)).astype("float32")
np_array[1, :] = 2.0
np_array[2, :] = 3.0
np_array[3, :] = 4.0
# initialize input variable X
x = scope.var('X').get_selected_rows()
x.set_rows(x_rows)
x.set_height(height)
x_tensor = x.get_tensor()
x_tensor.set(np_array, place)
# initialize input variable Out
out = scope.var("Out").get_tensor()
op = Operator("get_tensor_from_selected_rows", X="X", Out="Out")
op.run(scope, place)
out_array = np.array(out)
self.assertEqual((5, 2), out_array.shape)
assert (out_array == np_array).all()
def create_op(scope, op_type, inputs, outputs, attrs):
kwargs = dict()
def __create_var__(name, var_name):
scope.var(var_name).get_tensor()
kwargs[name].append(var_name)
for in_name, in_dup in Operator.get_op_inputs(op_type):
if in_name in inputs:
kwargs[in_name] = []
if in_dup:
sub_in = inputs[in_name]
for item in sub_in:
sub_in_name, _ = item[0], item[1]
__create_var__(in_name, sub_in_name)
else:
__create_var__(in_name, in_name)
for out_name, out_dup in Operator.get_op_outputs(op_type):
if out_name in outputs:
kwargs[out_name] = []
if out_dup:
sub_out = outputs[out_name]
for item in sub_out:
sub_out_name, _ = item[0], item[1]
__create_var__(out_name, sub_out_name)
else:
__create_var__(out_name, out_name)
for attr_name in Operator.get_op_attr_names(op_type):
if attr_name in attrs:
kwargs[attr_name] = attrs[attr_name]
return Operator(op_type, **kwargs)
def check_with_place(self, place, inplace):
scope = core.Scope()
if inplace:
self.create_lod_tensor(scope, place, "x1")
self.create_selected_rows(scope, place, "x2", True)
out = scope.var("x1").get_tensor()
out_name = "x1"
else:
self.create_selected_rows(scope, place, "x1", True)
self.create_lod_tensor(scope, place, "x2")
out = scope.var("out").get_tensor()
out_name = "out"
# create and run sum operator
sum_op = Operator("sum", X=["x1", "x2"], Out=out_name)
sum_op.run(scope, place)
result = np.ones((1, self.height)).astype(np.int32).tolist()[0]
for ele in self.rows:
result[ele] += 1
out_t = np.array(out)
self.assertEqual(out_t.shape[0], self.height)
self.assertTrue(
np.array_equal(
out_t,
self._get_array([i
for i in range(self.height)], self.row_numel)
* np.tile(
np.array(result).reshape(self.height, 1), self.row_numel)))
def check_with_place(self, place):
scope = core.Scope()
self.setup(scope, place)
op_args = dict()
for key, np_array in self.dense_inputs.items():
var = scope.var(key).get_tensor()
var.set(np_array, place)
op_args[key] = key
for s in self.sparse_inputs:
op_args[s] = s
for s in self.outputs:
var = scope.var(s).get_tensor()
var.set(self.outputs[s], place)
op_args[s] = s
for k in self.attrs:
op_args[k] = self.attrs[k]
# create and run sgd operator
adam_op = Operator("adam", **op_args)
adam_op.run(scope, place)
for key, np_array in self.outputs.items():
out_var = scope.var(key).get_tensor()
actual = np.array(out_var)
actual = actual.reshape([actual.size])
np_array = np_array.reshape([np_array.size])
for idx, row_id in enumerate(self.rows):
j = 0
while j < self.row_numel:
pos = row_id * self.row_numel + j
self.assertLess(
(actual[pos] - np_array[pos]) / actual[pos], 0.00001)
j += 1
def check_with_place(self, place):
scope = core.Scope()
# create and initialize Variable
feature_len = 12
rows = [0, 4, 4, 7]
np_array = np.ones((len(rows), feature_len)).astype("float32")
in_x = scope.var('X').get_selected_rows()
in_x.set_height(len(rows))
in_x.set_rows(rows)
in_x_tensor = in_x.get_tensor()
in_x_tensor.set(np_array, place)
# create Out Variable
out_tensor = scope.var('Out').get_tensor()
# create and run lookup_table operator
extract_rows_op = Operator("extract_rows", X='X', Out='Out')
extract_rows_op.run(scope, place)
# get result from Out
result_array = np.array(out_tensor)
result_array = [ele[0] for ele in result_array]
assert result_array == rows
def check_with_selected_rows(self, place):
scope = core.Scope()
x_rows = [0, 1, 5, 4, 19]
x_height = 20
row_numel = 2
np_array = np.ones((len(x_rows), row_numel)).astype("float32")
# initialize input variable
x = scope.var('X').get_selected_rows()
x.set_rows(x_rows)
x.set_height(x_height)
x_tensor = x.get_tensor()
x_tensor.set(np_array, place)
# initialize the Out variable
out = scope.var("Out").get_selected_rows()
out_tensor = out.get_tensor()
op = Operator("share_data", X="X", Out="Out")
op.run(scope, place)
out_height = out.height()
out_rows = out.rows()
self.assertTrue(np.allclose(np_array, out_tensor))
self.assertEqual(x_height, out_height)
self.assertEqual(x_rows, out_rows)
def check_with_place(self, place):
scope = core.Scope()
self.setup(scope, place)
op_args = dict()
for key, np_array in self.dense_inputs.iteritems():
var = scope.var(key).get_tensor()
var.set(np_array, place)
op_args[key] = key
for s in self.sparse_inputs:
op_args[s] = s
for s in self.outputs:
var = scope.var(s).get_tensor()
var.set(self.outputs[s], place)
op_args[s] = s
for k in self.attrs:
op_args[k] = self.attrs[k]
# create and run sgd operator
adam_op = Operator("adam", **op_args)
adam_op.run(scope, place)
for key, np_array in self.outputs.iteritems():
out_var = scope.var(key).get_tensor()
actual = np.array(out_var)
actual = actual.reshape([actual.size])
np_array = np_array.reshape([np_array.size])
for idx, row_id in enumerate(self.rows):
j = 0
while j < self.row_numel:
pos = row_id * self.row_numel + j
self.assertLess((actual[pos] - np_array[pos]) / actual[pos],
0.00001)
j += 1
def check_with_place(self, place):
scope = core.Scope()
# create and initialize Grad Variable
height = 10
rows = [0, 4, 7]
self.row_numel = 12
x_selected_rows = scope.var('X1').get_selected_rows()
x_selected_rows.set_height(height)
x_selected_rows.set_rows(rows)
np_array = np.ones((len(rows), self.row_numel)).astype("float32")
np_array[0, 0] = 2.0
np_array[2, 8] = 4.0
x_tensor = x_selected_rows.get_tensor()
x_tensor.set(np_array, place)
out_selected_rows = scope.var('Out1').get_selected_rows()
# create and run sqrt operator
sqrt_op = Operator("sqrt", X='X1', Out='Out1')
sqrt_op.run(scope, place)
# get and compare result
result_array = np.array(out_selected_rows.get_tensor())
self.assertTrue(np.allclose(result_array, np.sqrt(np_array)))
def check_with_place(self, place, in_name, out_name):
scope = core.Scope()
# create and initialize Grad Variable
in_height = 10
in_rows = [0, 4, 7]
in_row_numel = 12
scale = 2.0
in_selected_rows = scope.var(in_name).get_selected_rows()
in_selected_rows.set_height(in_height)
in_selected_rows.set_rows(in_rows)
in_array = np.random.random(
(len(in_rows), in_row_numel)).astype("float32")
in_tensor = in_selected_rows.get_tensor()
in_tensor.set(in_array, place)
# create and initialize Param Variable
out_selected_rows = scope.var(out_name).get_selected_rows()
out_tensor = out_selected_rows.get_tensor()
out_tensor._set_dims(in_tensor._get_dims())
# create and run sgd operator
scale_op = Operator("scale", X=in_name, Out=out_name, scale=scale)
scale_op.run(scope, place)
# get and compare result
out_height = out_selected_rows.height()
out_rows = out_selected_rows.rows()
result_array = np.array(out_tensor)
assert (in_array * scale == result_array).all()
assert in_height == out_height
assert in_rows == out_rows
def test_run(self):
op = Operator(
'beam_search',
pre_ids='pre_ids',
pre_scores='pre_scores',
ids='ids',
scores='scores',
selected_ids='selected_ids',
selected_scores='selected_scores',
parent_idx='parent_idx',
level=0,
beam_size=2,
end_id=0,
)
op.run(self.scope, core.CPUPlace())
selected_ids = self.scope.find_var("selected_ids").get_tensor()
selected_scores = self.scope.find_var("selected_scores").get_tensor()
parent_idx = self.scope.find_var("parent_idx").get_tensor()
self.assertTrue(
np.allclose(np.array(selected_ids),
np.array([4, 2, 3, 8])[:, np.newaxis]))
self.assertTrue(
np.allclose(np.array(selected_scores),
np.array([0.5, 0.6, 0.9, 0.7])[:, np.newaxis]))
self.assertEqual(selected_ids.lod(), [[0, 2, 4], [0, 1, 2, 3, 4]])
self.assertTrue(
np.allclose(np.array(parent_idx), np.array([0, 1, 2, 3])))
def check_with_place(self, place):
scope = core.Scope()
rows = [0, 1, 2, 3, 4, 5, 6]
row_numel = 7
w_selected_rows = scope.var('W').get_selected_rows()
w_selected_rows.set_height(len(rows))
w_selected_rows.set_rows(rows)
w_array = np.ones((len(rows), row_numel)).astype("float32")
for i in range(len(rows)):
w_array[i] *= i
w_tensor = w_selected_rows.get_tensor()
w_tensor.set(w_array, place)
# create and initialize Id Variable
ids = scope.var("Ids").get_tensor()
# create and run lookup_table operator
lookup_table = Operator("lookup_sparse_table_grad_split",
Grad='W',
Row={'Ids'},
Value={'W'},
is_entry=False,
tablename="sparse")
lookup_table.run(scope, place)
# get result from Out
result_array1 = np.array(ids)
print(result_array1)
print("== = = == == = == ==== ==== === ")
value = scope.var("W").get_tensor()
result_array1 = np.array(value)
print(result_array1.shape)
print(result_array1)
def check_with_place(self, place):
scope = core.Scope()
row_width = 12
# create and initialize Grad Variable
grad_height = 10
grad_rows = [0, 4, 7]
grad_selected_rows = scope.var('Grad').get_selected_rows()
grad_selected_rows.set_height(grad_height)
grad_selected_rows.set_rows(grad_rows)
grad_array = np.ones((len(grad_rows), row_width)).astype("float32")
grad_array[0, 0] = 2.0
grad_array[2, 8] = 4.0
grad_tensor = grad_selected_rows.get_tensor()
grad_tensor.set(grad_array, place)
# create and initialize Param Variable
# create and initialize W Variable
param_rows = [0, 1, 2, 3, 4, 5, 6, 7]
# init Param
w_selected_rows = scope.var('Param').get_selected_rows()
w_selected_rows.set_height(len(param_rows))
w_selected_rows.set_rows(param_rows)
w_selected_rows.sync_index()
w_array = np.ones((len(param_rows), row_width)).astype("float32")
for i in range(len(param_rows)):
w_array[i] *= i
w_tensor = w_selected_rows.get_tensor()
w_tensor.set(w_array, place)
w_before_optimize = np.array(w_tensor)
# create and initialize LeraningRate Variable
lr_value = 0.1
lr = scope.var('LearningRate').get_tensor()
lr_array = np.full((1), lr_value).astype("float32")
lr.set(lr_array, place)
# optimize with Python
w_after_optimize = np.copy(w_before_optimize)
for index, id in enumerate(grad_rows):
w_after_optimize[id] = w_before_optimize[
id] - lr_value * grad_array[index]
# create and run sgd operator
sgd_op = Operator(
"sgd",
Param='Param',
Grad='Grad',
ParamOut='Param',
LearningRate='LearningRate')
sgd_op.run(scope, place)
# get and compare result
result_array = np.array(w_tensor)
assert (result_array == w_after_optimize).all()
def check_sgd_step(self, place):
self.setup(place=place, step=15.0)
dgc_momentum_op = Operator(self.op_type, **self.kwargs)
dgc_momentum_op.run(self.scope, self.place)
self.check(np.array(self.param_tensor), self.outputs['SGDOut'],
self.place, self.param_name)
def check_with_place(self, place):
scope = core.Scope()
row_width = 12
# create and initialize Grad Variable
grad_height = 10
grad_rows = [0, 4, 7]
grad_selected_rows = scope.var('Grad').get_selected_rows()
grad_selected_rows.set_height(grad_height)
grad_selected_rows.set_rows(grad_rows)
grad_array = np.ones((len(grad_rows), row_width)).astype("float32")
grad_array[0, 0] = 2.0
grad_array[2, 8] = 4.0
grad_tensor = grad_selected_rows.get_tensor()
grad_tensor.set(grad_array, place)
# create and initialize Param Variable
# create and initialize W Variable
param_rows = [0, 1, 2, 3, 4, 5, 6, 7]
# init Param
w_selected_rows = scope.var('Param').get_selected_rows()
w_selected_rows.set_height(len(param_rows))
w_selected_rows.set_rows(param_rows)
w_array = np.ones((len(param_rows), row_width)).astype("float32")
for i in range(len(param_rows)):
w_array[i] *= i
w_tensor = w_selected_rows.get_tensor()
w_tensor.set(w_array, place)
w_before_optimize = np.array(w_tensor)
# create and initialize LeraningRate Variable
lr_value = 0.1
lr = scope.var('LearningRate').get_tensor()
lr_array = np.full((1), lr_value).astype("float32")
lr.set(lr_array, place)
# optimize with Python
w_after_optimize = np.copy(w_before_optimize)
for index, id in enumerate(grad_rows):
w_after_optimize[id] = w_before_optimize[
id] - lr_value * grad_array[index]
# create and run sgd operator
sgd_op = Operator(
"sgd",
Param='Param',
Grad='Grad',
ParamOut='Param',
LearningRate='LearningRate')
sgd_op.run(scope, place)
# get and compare result
result_array = np.array(w_tensor)
assert (result_array == w_after_optimize).all()
def test_get_set(self):
ids = self.scope.var("ids").get_lod_tensor_array()
scores = self.scope.var("scores").get_lod_tensor_array()
# Construct sample data with 5 steps and 2 source sentences
# beam_size = 2, end_id = 1
# start with start_id
[
self.append_lod_tensor(array, [[0, 1, 2], [0, 1, 2]],
np.array([0, 0], dtype=dtype))
for array, dtype in ((ids, "int64"), (scores, "float32"))
]
[
self.append_lod_tensor(array, [[0, 1, 2], [0, 2, 4]],
np.array([2, 3, 4, 5], dtype=dtype))
for array, dtype in ((ids, "int64"), (scores, "float32"))
]
[
self.append_lod_tensor(array, [[0, 2, 4], [0, 2, 2, 4, 4]],
np.array([3, 1, 5, 4], dtype=dtype))
for array, dtype in ((ids, "int64"), (scores, "float32"))
]
[
self.append_lod_tensor(array, [[0, 2, 4], [0, 1, 2, 3, 4]],
np.array([1, 1, 3, 5], dtype=dtype))
for array, dtype in ((ids, "int64"), (scores, "float32"))
]
[
self.append_lod_tensor(array, [[0, 2, 4], [0, 0, 0, 2, 2]],
np.array([5, 1], dtype=dtype))
for array, dtype in ((ids, "int64"), (scores, "float32"))
]
sentence_ids = self.scope.var("sentence_ids").get_tensor()
sentence_scores = self.scope.var("sentence_scores").get_tensor()
beam_search_decode_op = Operator(
"beam_search_decode",
# inputs
Ids="ids",
Scores="scores",
# outputs
SentenceIds="sentence_ids",
SentenceScores="sentence_scores",
beam_size=2,
end_id=1,
)
beam_search_decode_op.run(self.scope, self.place)
expected_lod = [[0, 2, 4], [0, 4, 7, 12, 17]]
self.assertEqual(sentence_ids.lod(), expected_lod)
self.assertEqual(sentence_scores.lod(), expected_lod)
expected_data = np.array(
[0, 2, 3, 1, 0, 2, 1, 0, 4, 5, 3, 5, 0, 4, 5, 3, 1], "int64")
self.assertTrue(np.array_equal(np.array(sentence_ids), expected_data))
self.assertTrue(
np.array_equal(np.array(sentence_scores), expected_data))
def test_run_and_check(self):
self.setup(place=core.CUDAPlace(0))
kwargs = {
# inputs
'U': self.u_name,
'V': self.v_name,
'Grad': self.grad_name,
'Param': self.param_name,
'current_step': self.current_step_name,
'nranks': self.nranks_name,
# outputs
'U_out': self.u_name,
'V_out': self.v_name,
'EncodeGrad': self.encode_grad_name,
'Grad_out': self.grad_name,
'k': self.k_name,
'GatherBuff': self.gather_buff_name,
# attrs
'm': 0.9,
'sparsity': [0.75, 0.9375, 0.984375, 0.996, 0.999],
'use_nesterov': True,
'rampup_begin_step': float(0.0),
'rampup_step': float(10.0),
'regular_coeff': float(1e-4),
'regular_type': int(2),
}
dgc_op = Operator('dgc', **kwargs)
#atol = 1e-6
dgc_op.run(self.scope, self.place)
u_out = np.array(self.u_tensor)
v_out = np.array(self.v_tensor)
grad_out = np.array(self.grad_tensor)
encode_grad_out = np.array(self.encode_grad_tensor)
k = int(np.array(self.k_tensor)[0])
print("u_out:", u_out[0:20])
print("v_out:", v_out[0:20])
print("encode_grad_out:", encode_grad_out)
print("k_out:", k)
self.assertEqual(k, int(g_array_size * 0.25))
index = encode_grad_out[0:k].view(dtype=np.int32)
value = encode_grad_out[k:2 * k]
acl = 1e-7
for i in range(0, k):
self.assertAlmostEqual(u_out[index[i]], 0.0)
self.assertAlmostEqual(v_out[index[i]], 0.0)
a_min = np.amin(value)
dangling = [x for x in v_out if x > a_min]
def check_momentum_step(self, place):
self.setup(place=place)
dgc_momentum_op = Operator(self.op_type, **self.kwargs)
dgc_momentum_op.run(self.scope, self.place)
self.check(np.array(self.param_tensor), self.outputs['ParamOut'],
self.place, self.param_name)
self.check(np.array(self.velocity_tensor), self.outputs['VelocityOut'],
self.place, self.velocity_name)
def check_with_place(self, place):
scope = core.Scope()
# create and initialize Grad Variable
height = 10
rows = [0, 4, 7]
row_numel = 12
grad_selected_rows = scope.var('Grad').get_selected_rows()
grad_selected_rows.set_height(height)
grad_selected_rows.set_rows(rows)
np_array = np.ones((len(rows), row_numel)).astype("float32")
np_array[0, 0] = 2.0
np_array[2, 8] = 4.0
grad_tensor = grad_selected_rows.get_tensor()
grad_tensor.set(np_array, place)
# create and initialize Param Variable
param = scope.var('Param').get_tensor()
param_array = np.full((height, row_numel), 5.0).astype("float32")
param.set(param_array, place)
# create and initialize LeraningRate Variable
lr = scope.var('LearningRate').get_tensor()
lr_array = np.full((1), 2.0).astype("float32")
lr.set(lr_array, place)
# create and run sgd operator
sgd_op = Operator(
"sgd",
Param='Param',
Grad='Grad',
ParamOut='Param',
LearningRate='LearningRate')
sgd_op.run(scope, place)
# get and compare result
result_array = np.array(param)
# rows[0] = 0, 5.0 - 2.0 * 2.0
self.assertAlmostEqual(1.0, result_array[rows[0], 0])
# rows[0] = 0, 5.0 - 2.0 * 1.0
self.assertAlmostEqual(3.0, result_array[rows[0], 2])
# 5.0 - 2.0 * 0.0
self.assertAlmostEqual(5.0, result_array[1, 0])
# rows[1] = 4, 5.0 - 2.0 * 1.0
self.assertAlmostEqual(3.0, result_array[rows[1], 10])
# 5.0 - 2.0 * 0.0
self.assertAlmostEqual(5.0, result_array[5, 8])
# rows[2] = 7, 5.0 - 2.0 * 1.0
self.assertAlmostEqual(3.0, result_array[rows[2], 1])
# rows[2] = 7, 5.0 - 2.0 * 4.0
self.assertAlmostEqual(-3.0, result_array[rows[2], 8])
def check_with_place(self, place):
scope = core.Scope()
# create and initialize Grad Variable
height = 10
rows = [0, 4, 7]
self.conf()
grad_selected_rows = scope.var('Grad').get_selected_rows()
grad_selected_rows.set_height(height)
grad_selected_rows.set_rows(rows)
np_array = np.ones((len(rows), self.row_numel)).astype("float32")
np_array[0, 0] = 2.0
np_array[2, 8] = 4.0
grad_tensor = grad_selected_rows.get_tensor()
grad_tensor.set(np_array, place)
# create and initialize Param Variable
param = scope.var('Param').get_tensor()
param_array = np.full((height, self.row_numel), 5.0).astype("float32")
param.set(param_array, place)
# create and initialize LeraningRate Variable
lr = scope.var('LearningRate').get_tensor()
lr_array = np.full((1), 2.0).astype("float32")
lr.set(lr_array, place)
# create and run sgd operator
sgd_op = Operator(
"sgd",
Param='Param',
Grad='Grad',
ParamOut='Param',
LearningRate='LearningRate')
sgd_op.run(scope, place)
# get and compare result
result_array = np.array(param)
# rows[0] = 0, 5.0 - 2.0 * 2.0
self.assertAlmostEqual(1.0, result_array[rows[0], 0])
# rows[0] = 0, 5.0 - 2.0 * 1.0
self.assertAlmostEqual(3.0, result_array[rows[0], 2])
# 5.0 - 2.0 * 0.0
self.assertAlmostEqual(5.0, result_array[1, 0])
# rows[1] = 4, 5.0 - 2.0 * 1.0
self.assertAlmostEqual(3.0, result_array[rows[1], 10])
# 5.0 - 2.0 * 0.0
self.assertAlmostEqual(5.0, result_array[5, 8])
# rows[2] = 7, 5.0 - 2.0 * 1.0
self.assertAlmostEqual(3.0, result_array[rows[2], 1])
# rows[2] = 7, 5.0 - 2.0 * 4.0
self.assertAlmostEqual(-3.0, result_array[rows[2], 8])
def check_with_place(self, place):
scope = core.Scope()
self.prepare_input(scope, place)
out_selected_rows = self.create_out_selected_row(scope)
out_selected_rows.set_height(0)
out_selected_rows.set_rows([])
elementwise_mul = Operator("elementwise_mul", X='X', Y='Y', Out='Out')
elementwise_mul.run(scope, place)
self.check_result(out_selected_rows)
def check_with_tensor(self, place):
scope = core.Scope()
np_array = np.random.rand(2, 3, 5).astype("float32")
# initialize input and output variable
x = scope.var('X').get_tensor()
x.set(np_array, place)
out = scope.var("Out").get_tensor()
op = Operator("share_data", X="X", Out="Out")
op.run(scope, place)
self.assertTrue(np.allclose(np_array, out))
def check_with_place(self, place):
scope = core.Scope()
rows = [0, 5, 7, 4, 20]
height = 20
row_numel = 2
# initialize input variable X
x = scope.var('X').get_selected_rows()
x.set_rows(rows)
x.set_height(height)
np_array = np.ones((len(rows), row_numel)).astype("float32")
np_array[0, 0] = 2.0
np_array[2, 1] = 4.0
np_array[4, 1] = 8.0
x_tensor = x.get_tensor()
x_tensor.set(np_array, place)
height_sections = [5, 5, 5, 5, 3]
# initialize output variables [out0, out1]
outs_name = ["out%d" % i for i in range(len(height_sections))]
outs = [
scope.var(var_name).get_selected_rows() for var_name in outs_name
]
# expected output selected rows
expected_out0_rows = [0, 4]
expected_out1_rows = [0, 2]
expected_out2_rows = []
expected_out4_rows = [0]
op = Operator(
"split_selected_rows",
X="X",
Out=outs_name,
height_sections=height_sections)
op.run(scope, place)
self.assertEqual(outs[0].rows(), expected_out0_rows)
self.assertEqual(outs[1].rows(), expected_out1_rows)
self.assertEqual(outs[2].rows(), expected_out2_rows)
self.assertEqual(outs[4].rows(), expected_out4_rows)
self.assertEqual(outs[0].height(), height_sections[0])
self.assertEqual(outs[4].height(), height_sections[4])
self.assertAlmostEqual(2.0, np.array(outs[0].get_tensor())[0, 0])
self.assertAlmostEqual(4.0, np.array(outs[1].get_tensor())[1, 1])
self.assertAlmostEqual(8.0, np.array(outs[4].get_tensor())[0, 1])
self.assertEqual(outs[2].numel(), 0)
self.assertEqual(outs[3].numel(), 0)
def test_get_set(self):
ids = self.scope.var("ids").get_lod_tensor_array()
self.append_lod_tensor(
ids, [[0, 3, 6], [0, 1, 2, 3, 4, 5, 6]],
np.array(
[1, 2, 3, 4, 5, 6], dtype="int64"))
self.append_lod_tensor(
ids, [[0, 3, 6], [0, 1, 1, 3, 5, 5, 6]],
np.array(
[0, 1, 2, 3, 4, 5], dtype="int64"))
self.append_lod_tensor(
ids, [[0, 3, 6], [0, 0, 1, 2, 3, 4, 5]],
np.array(
[0, 1, 2, 3, 4], dtype="int64"))
scores = self.scope.var("scores").get_lod_tensor_array()
self.append_lod_tensor(
scores, [[0, 3, 6], [0, 1, 2, 3, 4, 5, 6]],
np.array(
[1, 2, 3, 4, 5, 6], dtype="float64"))
self.append_lod_tensor(
scores, [[0, 3, 6], [0, 1, 1, 3, 5, 5, 6]],
np.array(
[0, 1, 2, 3, 4, 5], dtype="float64"))
self.append_lod_tensor(
scores, [[0, 3, 6], [0, 0, 1, 2, 3, 4, 5]],
np.array(
[0, 1, 2, 3, 4], dtype="float64"))
sentence_ids = self.scope.var("sentence_ids").get_tensor()
sentence_scores = self.scope.var("sentence_scores").get_tensor()
beam_search_decode_op = Operator(
"beam_search_decode",
# inputs
Ids="ids",
Scores="scores",
# outputs
SentenceIds="sentence_ids",
SentenceScores="sentence_scores")
beam_search_decode_op.run(self.scope, self.place)
expected_lod = [[0, 4, 8], [0, 1, 3, 6, 9, 10, 13, 16, 19]]
self.assertEqual(sentence_ids.lod(), expected_lod)
self.assertEqual(sentence_scores.lod(), expected_lod)
expected_data = np.array(
[2, 1, 0, 3, 1, 0, 3, 2, 1, 5, 4, 3, 2, 4, 4, 3, 6, 5, 4], "int64")
self.assertTrue(np.array_equal(np.array(sentence_ids), expected_data))
self.assertTrue(
np.array_equal(np.array(sentence_scores), expected_data))
def check_with_place(self, place):
scope = core.Scope()
condition = scope.var('Condition').get_tensor()
condition.set(self.cond_data, place)
out = scope.var("Out").get_tensor()
out.set(np.full(self.shape, 0).astype('int64'), place)
op = Operator("where_index", Condition="Condition", Out="Out")
op.run(scope, place)
out_array = np.array(out)
self.assertTrue((out_array == self.out_data).all())
def test_check_output(self):
self.prepare_ids()
self.prepare_w()
out_tensor = self.scope.var('Out').get_tensor()
# create and run lookup_table operator
lookup_table = Operator(self.op_type, W='W', Ids='Ids', Out='Out')
lookup_table.run(self.scope, self.place)
# get result from Out
result_array = np.array(out_tensor)
ref = _lookup(self.w_fp32, self.ids, self.flat_ids, self.op_type)
self._check_output(ref, result_array)
def run_and_check(self):
grad_name = self.grad_sr_name if self.is_sparse else self.grad_name
kwargs = {
'Param': self.param_name,
'Grad': grad_name,
'MeanSquare': self.mean_square_name,
'Moment': self.moment_name,
'LearningRate': self.lr_name,
'ParamOut': self.param_name,
'MeanSquareOut': self.mean_square_name,
'MomentOut': self.moment_name,
'epsilon': self.epsilon,
'decay': self.decay,
'momentum': self.momentum,
'centered': self.centered
}
if self.centered:
kwargs['MeanGrad'] = self.mean_grad_name
kwargs['MeanGradOut'] = self.mean_grad_name
rmsprop_op = Operator('rmsprop', **kwargs)
atol = 1e-6
rmsprop_op.run(self.scope, self.place)
self.check(
np.array(self.mean_square_tensor),
self.ms_out,
self.place,
self.mean_square_name,
atol=atol)
self.check(
np.array(self.moment_tensor),
self.moment_out,
self.place,
self.moment_name,
atol=atol)
self.check(
np.array(self.param_tensor),
self.param_out,
self.place,
self.param_name,
atol=atol)
if self.centered:
self.check(
np.array(self.mean_grad_tensor), self.mg_out, self.place,
self.mean_grad_name)
def check_with_place(self, place):
scope = core.Scope()
out = scope.var("X").get_selected_rows()
paddle.seed(10)
op = Operator("uniform_random",
Out="X",
shape=[1000, 784],
min=-5.0,
max=10.0,
seed=10)
op.run(scope, place)
self.assertEqual(out.get_tensor().shape(), [1000, 784])
hist, prob = output_hist(np.array(out.get_tensor()))
self.assertTrue(np.allclose(hist, prob, rtol=0, atol=0.01),
"hist: " + str(hist))
def test_run(self):
op = Operator(
'beam_search',
pre_ids="pre_ids",
ids='ids',
scores='scores',
selected_ids='selected_ids',
selected_scores='selected_scores',
level=0,
beam_size=2,
end_id=0, )
op.run(self.scope, core.CPUPlace())
selected_ids = self.scope.find_var("selected_ids").get_tensor()
print 'selected_ids', np.array(selected_ids)
print 'lod', selected_ids.lod()
def check_with_place(self, place):
scope = core.Scope()
# create Out Variable
out = scope.var('Out').get_selected_rows()
# create and run fill_constant_op operator
fill_constant_op = Operator(
"fill_constant", shape=[123, 92], value=3.8, Out='Out')
fill_constant_op.run(scope, place)
# get result from Out
result_array = np.array(out.get_tensor())
full_array = np.full((123, 92), 3.8, 'float32')
self.assertTrue(np.array_equal(result_array, full_array))
def check_with_place(self, place):
scope = core.Scope()
rows = [0, 5, 7, 4, 20]
height = 20
row_numel = 2
# initialize input variable X
x = scope.var('X').get_selected_rows()
x.set_rows(rows)
x.set_height(height)
np_array = np.ones((len(rows), row_numel)).astype("float32")
np_array[0, 0] = 2.0
np_array[2, 1] = 4.0
np_array[4, 1] = 8.0
x_tensor = x.get_tensor()
x_tensor.set(np_array, place)
height_sections = [5, 5, 5, 5, 3]
# initialize output variables [out0, out1]
outs_name = ["out%d" % i for i in xrange(len(height_sections))]
outs = [
scope.var(var_name).get_selected_rows() for var_name in outs_name
]
# expected output selected rows
expected_out0_rows = [0, 4]
expected_out1_rows = [0, 2]
expected_out4_rows = [0]
op = Operator(
"split_selected_rows",
X="X",
Out=outs_name,
height_sections=height_sections)
op.run(scope, place)
self.assertEqual(outs[0].rows(), expected_out0_rows)
self.assertEqual(outs[1].rows(), expected_out1_rows)
self.assertEqual(outs[4].rows(), expected_out4_rows)
self.assertEqual(outs[0].height(), height_sections[0])
self.assertEqual(outs[4].height(), height_sections[4])
self.assertAlmostEqual(2.0, np.array(outs[0].get_tensor())[0, 0])
self.assertAlmostEqual(4.0, np.array(outs[1].get_tensor())[1, 1])
self.assertAlmostEqual(8.0, np.array(outs[4].get_tensor())[0, 1])
def set_input(scope, op, inputs, place):
def np_value_to_fluid_value(input):
if input.dtype == np.float16:
input = input.view(np.uint16)
return input
def __set_input__(var_name, var):
if isinstance(var, tuple) or isinstance(var, np.ndarray):
tensor = scope.find_var(var_name).get_tensor()
if isinstance(var, tuple):
tensor.set_recursive_sequence_lengths(var[1])
var = var[0]
tensor._set_dims(var.shape)
tensor.set(np_value_to_fluid_value(var), place)
elif isinstance(var, float):
scope.find_var(var_name).set_float(var)
elif isinstance(var, int):
scope.find_var(var_name).set_int(var)
for in_name, in_dup in Operator.get_op_inputs(op.type()):
if in_name in inputs:
if in_dup:
sub_in = inputs[in_name]
for item in sub_in:
sub_in_name, sub_in_val = item[0], item[1]
__set_input__(sub_in_name, sub_in_val)
else:
__set_input__(in_name, inputs[in_name])
def test_run(self):
op = Operator(
'beam_search',
pre_ids="pre_ids",
ids='ids',
scores='scores',
selected_ids='selected_ids',
selected_scores='selected_scores',
level=0,
beam_size=2,
end_id=0,
)
op.run(self.scope, core.CPUPlace())
selected_ids = self.scope.find_var("selected_ids").get_tensor()
print 'selected_ids', np.array(selected_ids)
print 'lod', selected_ids.lod()
def check_with_place(self, place):
scope = core.Scope()
ids_array = self.prepare_ids(scope, place)
self.prepare_w(scope, place)
out_tensor = self.create_out_tensor(scope, place)
# create and run lookup_table_v2 operator
lookup_table = Operator("lookup_table_v2", W='W', Ids='Ids', Out='Out')
lookup_table.run(scope, place)
# get result from Out
result_array = np.array(out_tensor)
self.check_result(ids_array, result_array)
def check_with_place(self, place):
scope = core.Scope()
# create and initialize Id Variable
ids = scope.var("Ids").get_tensor()
ids_array = np.array([0, 2, 3, 5, 100]).astype("int64")
ids.set(ids_array, place)
# create and initialize W Variable
rows = [0, 1, 2, 3, 4, 5, 6]
row_numel = 10000
w_selected_rows = scope.var('W').get_selected_rows()
w_selected_rows.set_height(len(rows))
w_selected_rows.set_rows(rows)
w_array = np.ones((len(rows), row_numel)).astype("float32")
for i in range(len(rows)):
w_array[i] *= i
w_tensor = w_selected_rows.get_tensor()
w_tensor.set(w_array, place)
# create Out Variable
out_tensor = scope.var('Out').get_tensor()
# create and run lookup_table operator
lookup_table = Operator(
"lookup_sparse_table",
W='W',
Ids='Ids',
Out='Out',
min=-5.0,
max=10.0,
seed=10)
lookup_table.run(scope, place)
# get result from Out
result_array = np.array(out_tensor)
# all(): return True if all elements of the iterable are true (or if the iterable is empty)
for idx, row in enumerate(ids_array[:-2]):
assert (row == result_array[idx]).all()
# check the random value
hist, prob = output_hist(result_array[-1])
self.assertTrue(
np.allclose(
hist, prob, rtol=0, atol=0.01), "hist: " + str(hist))
def check_with_place(self, place):
scope = core.Scope()
out = scope.var("X").get_selected_rows()
op = Operator(
"uniform_random",
Out="X",
shape=[4, 784],
min=-5.0,
max=10.0,
seed=10)
op.run(scope, place)
self.assertEqual(out.get_tensor().shape(), [4, 784])
hist, prob = output_hist(np.array(out.get_tensor()))
self.assertTrue(
np.allclose(
hist, prob, rtol=0, atol=0.01), "hist: " + str(hist))
def check_with_place(self, place):
scope = core.Scope()
out = scope.var("X").get_selected_rows()
shape_tensor = scope.var("Shape").get_tensor()
shape_tensor.set(np.array([4, 784]).astype("int64"), place)
op = Operator("uniform_random",
ShapeTensor="Shape",
Out="X",
min=-5.0,
max=10.0,
seed=10)
op.run(scope, place)
self.assertEqual(out.get_tensor().shape(), [4, 784])
hist, prob = output_hist(np.array(out.get_tensor()))
self.assertTrue(np.allclose(hist, prob, rtol=0, atol=0.01),
"hist: " + str(hist))
def check_grad_with_place(self, place):
scope = core.Scope()
height = 10
row_numel = 2
# attr
height_sections = [5, 5]
# initialize input variable X
out0_grad = scope.var("out0@GRAD").get_selected_rows()
rows0 = [0, 5]
out0_grad.set_rows(rows0)
out0_grad.set_height(height)
out0_grad_tensor = out0_grad.get_tensor()
np_array = np.ones((len(rows0), row_numel)).astype("float32")
np_array[0, 0] = 2.0
out0_grad_tensor.set(np_array, place)
out1_grad = scope.var("out1@GRAD").get_selected_rows()
rows1 = [2, 0]
out1_grad.set_rows(rows1)
out1_grad.set_height(height)
out1_grad_tensor = out1_grad.get_tensor()
np_array = np.ones((len(rows1), row_numel)).astype("float32")
np_array[0, 1] = 4.0
out1_grad_tensor.set(np_array, place)
x_grad = scope.var("X@GRAD").get_selected_rows()
grad_op = Operator(
"sum",
X=["out0@GRAD", "out1@GRAD"],
Out="X@GRAD",
height_sections=height_sections)
grad_op.run(scope, place)
self.assertEqual(x_grad.rows(), rows0 + rows1)
self.assertEqual(x_grad.height(), height)
self.assertAlmostEqual(2.0, np.array(x_grad.get_tensor())[0, 0])
self.assertAlmostEqual(4.0, np.array(x_grad.get_tensor())[2, 1])
def create_op(scope, op_type, inputs, outputs, attrs):
kwargs = dict()
op_maker = core.op_proto_and_checker_maker
op_role_attr_name = op_maker.kOpRoleAttrName()
if op_role_attr_name not in attrs:
attrs[op_role_attr_name] = int(op_maker.OpRole.Forward)
def __create_var__(name, var_name):
scope.var(var_name).get_tensor()
kwargs[name].append(var_name)
for in_name, in_dup in Operator.get_op_inputs(op_type):
if in_name in inputs:
kwargs[in_name] = []
if in_dup:
sub_in = inputs[in_name]
for item in sub_in:
sub_in_name, _ = item[0], item[1]
__create_var__(in_name, sub_in_name)
else:
__create_var__(in_name, in_name)
for out_name, out_dup in Operator.get_op_outputs(op_type):
if out_name in outputs:
kwargs[out_name] = []
if out_dup:
sub_out = outputs[out_name]
for item in sub_out:
sub_out_name, _ = item[0], item[1]
__create_var__(out_name, sub_out_name)
else:
__create_var__(out_name, out_name)
for attr_name in Operator.get_op_attr_names(op_type):
if attr_name in attrs:
kwargs[attr_name] = attrs[attr_name]
return Operator(op_type, **kwargs)
def check_output_with_place(self, place, atol):
outs, fetch_list = self._calc_output(place)
for out_name, out_dup in Operator.get_op_outputs(self.op_type):
if out_name not in self.outputs:
continue
def find_actual(target_name, fetch_list):
found = [
i for i, var in enumerate(fetch_list)
if var.name == target_name
]
self.assertTrue(
len(found) == 1, "Found {} {}".format(
len(found), target_name))
return found[0]
if out_dup:
sub_out = self.outputs[out_name]
if not isinstance(sub_out, list):
raise AssertionError("sub_out type %s is not list",
type(sub_out))
for item in sub_out:
sub_out_name, expect = item[0], item[1]
idx = find_actual(sub_out_name, fetch_list)
actual = outs[idx]
actual_t = np.array(actual)
expect_t = expect[0] \
if isinstance(expect, tuple) else expect
self.assertTrue(
np.allclose(
actual_t, expect_t, atol=atol),
"Output (" + sub_out_name + ") has diff at " +
str(place))
if isinstance(expect, tuple):
self.assertListEqual(
actual.lod(), expect[1], "Output (" + sub_out_name +
") has different lod at " + str(place))
else:
idx = find_actual(out_name, fetch_list)
actual = outs[idx]
actual_t = np.array(actual)
expect = self.outputs[out_name]
expect_t = expect[0] if isinstance(expect, tuple) else expect
self.assertTrue(
np.allclose(
actual_t, expect_t, atol=atol),
"Output (" + out_name + ") has diff at " + str(place) +
str(actual_t) + "\n" + str(expect_t))
if isinstance(expect, tuple):
self.assertListEqual(actual.lod(), expect[1],
"Output (" + out_name +
") has different lod at " + str(place))
def check_with_place(self, place):
scope = core.Scope()
# create and initialize Variable
height = 10
rows = [0, 4, 4, 7]
row_numel = 12
# create and initialize W Variable
W = scope.var('W').get_tensor()
W_array = np.full((height, row_numel), 1.0).astype("float32")
for i in range(height):
W_array[i] *= i
W.set(W_array, place)
# create and initialize Ids Variable
ids_selected_rows = scope.var('Ids').get_selected_rows()
ids_selected_rows.set_height(len(rows))
ids_selected_rows.set_rows(rows)
np_array = np.ones((len(rows), row_numel)).astype("float32")
ids_tensor = ids_selected_rows.get_tensor()
ids_tensor.set(np_array, place)
# create Out Variable
Out = scope.var('Out').get_selected_rows()
# create and run lookup_table operator
concat_rows_op = Operator("lookup_table", W='W', Ids='Ids', Out='Out')
concat_rows_op.run(scope, place)
# get result from Out
Out_tensor = Out.get_tensor()
result_array = np.array(Out_tensor)
# all(): return True if all elements of the iterable are true (or if the iterable is empty)
for idx, row in enumerate(rows):
assert (row == result_array[idx]).all()
def set_input(scope, op, inputs, place):
def __set_input__(var_name, var):
if isinstance(var, tuple) or isinstance(var, np.ndarray):
tensor = scope.find_var(var_name).get_tensor()
if isinstance(var, tuple):
tensor.set_lod(var[1])
var = var[0]
tensor.set_dims(var.shape)
tensor.set(var, place)
elif isinstance(var, float):
scope.find_var(var_name).set_float(var)
elif isinstance(var, int):
scope.find_var(var_name).set_int(var)
for in_name, in_dup in Operator.get_op_inputs(op.type()):
if in_name in inputs:
if in_dup:
sub_in = inputs[in_name]
for item in sub_in:
sub_in_name, sub_in_val = item[0], item[1]
__set_input__(sub_in_name, sub_in_val)
else:
__set_input__(in_name, inputs[in_name])
def check_with_place(self, place, data_layout, dtype, shape):
epsilon = 0.00001
if len(shape) == 2:
x_shape = shape
c = x_shape[1]
else:
n, h, w, c = shape[0], shape[1], shape[2], shape[3]
if data_layout == "NHWC":
x_shape = [n, h, w, c]
elif data_layout == "NCHW":
x_shape = [n, c, h, w]
else:
raise ValueError("Unknown data layout.")
scale_shape = [c]
x_val = np.random.random_sample(x_shape).astype(dtype)
scale_val = np.random.random_sample(scale_shape).astype(np.float32)
bias_val = np.random.random_sample(scale_shape).astype(np.float32)
mean = np.zeros(scale_shape).astype(np.float32)
variance = np.ones(scale_shape).astype(np.float32)
y_out = _reference_testing(x_val, scale_val, bias_val, mean, variance,
epsilon, data_layout).astype(dtype)
scope = core.Scope()
# create input
x_tensor = create_or_get_tensor(scope, "x_val",
OpTest.np_dtype_to_fluid_dtype(x_val),
place)
scale_tensor = create_or_get_tensor(
scope, "scale_val",
OpTest.np_dtype_to_fluid_dtype(scale_val), place)
bias_tensor = create_or_get_tensor(
scope, "bias_val", OpTest.np_dtype_to_fluid_dtype(bias_val), place)
mean_tensor = create_or_get_tensor(scope, "mean",
OpTest.np_dtype_to_fluid_dtype(mean),
place)
variance_tensor = create_or_get_tensor(
scope, "variance", OpTest.np_dtype_to_fluid_dtype(variance), place)
# create output
y_tensor = create_or_get_tensor(scope, "y_out", None, place)
saved_mean_tensor = create_or_get_tensor(scope, "saved_mean", None,
place)
saved_variance_tensor = create_or_get_tensor(scope, "saved_variance",
None, place)
mean_out_tensor = mean_tensor
variance_out_tensor = variance_tensor
batch_norm_op = Operator(
"batch_norm",
# inputs
X="x_val",
Scale="scale_val",
Bias="bias_val",
Mean="mean",
Variance="variance",
# outputs
Y="y_out",
MeanOut="mean",
VarianceOut="variance",
SavedMean="saved_mean",
SavedVariance="saved_variance",
# attrs
is_test=True,
data_layout=data_layout,
use_mkldnn=self.use_mkldnn,
epsilon=epsilon)
batch_norm_op.run(scope, place)
# check inference result
self.__assert_close(
y_tensor,
y_out,
"inference output are different at " + str(place) + ", " +
data_layout + ", " + str(np.dtype(dtype)) +
str(np.array(y_tensor)) + str(y_out),
atol=1e-3)
def check_with_place(self, place):
scope = core.Scope()
# create and initialize Grad Variable
height = 10
rows = [0, 4, 7, 4]
row_numel = 12
grad_selected_rows = scope.var('Grad').get_selected_rows()
grad_selected_rows.set_height(height)
grad_selected_rows.set_rows(rows)
np_array = np.ones((len(rows), row_numel)).astype("float32")
np_array[0, 0] = 2.0
np_array[2, 8] = 4.0
grad_tensor = grad_selected_rows.get_tensor()
grad_tensor.set(np_array, place)
# create and initialize Param Variable
param = scope.var('Param').get_tensor()
param_array = np.full((height, row_numel), 5.0).astype("float32")
param.set(param_array, place)
# create and initialize LeraningRate Variable
lr = scope.var('LearningRate').get_tensor()
lr_array = np.full((1), 2.0).astype("float32")
lr.set(lr_array, place)
# create and initialize moment Variable
moment = scope.var('Moment').get_tensor()
moment_np_array = np.full((height, row_numel), 2.0).astype("float32")
moment.set(moment_np_array, place)
# create and run sgd operator
adagrad_op = Operator(
"adagrad",
Param='Param',
Grad='Grad',
ParamOut='Param',
Moment='Moment',
MomentOut='Moment',
LearningRate='LearningRate',
epsilon=2.0)
adagrad_op.run(scope, place)
# get and compare moment result
moment_result_array = np.array(moment)
self.assertAlmostEqual(6.0, moment_result_array[rows[0], 0])
self.assertAlmostEqual(3.0, moment_result_array[rows[0], 2])
self.assertAlmostEqual(2.0, moment_result_array[1, 0])
# 2.0 + (1.0 + 1.0)^2
self.assertAlmostEqual(6.0, moment_result_array[rows[1], 10])
self.assertAlmostEqual(6.0, moment_result_array[rows[3], 4])
self.assertAlmostEqual(2.0, moment_result_array[5, 8])
self.assertAlmostEqual(3.0, moment_result_array[rows[2], 1])
self.assertAlmostEqual(18.0, moment_result_array[rows[2], 8])
# get and compare param result
result_array = np.array(param)
def get_out(param, lr, grad, m, epsilon):
return param - lr * grad / (math.sqrt(m) + epsilon)
self.assertAlmostEqual(
get_out(5.0, 2.0, 2.0, 6.0, 2.0),
result_array[rows[0], 0],
places=5)
self.assertAlmostEqual(
get_out(5.0, 2.0, 1.0, 3.0, 2.0),
result_array[rows[0], 2],
places=5)
self.assertAlmostEqual(
get_out(5.0, 2.0, 0.0, 2.0, 2.0), result_array[1, 0], places=5)
# grad_merge = 1.0 + 1.0
# m = 6.0
self.assertAlmostEqual(
get_out(5.0, 2.0, 2.0, 6.0, 2.0),
result_array[rows[1], 10],
places=5)
self.assertAlmostEqual(
get_out(5.0, 2.0, 0.0, 2.0, 2.0), result_array[5, 8], places=5)
self.assertAlmostEqual(
get_out(5.0, 2.0, 1.0, 3.0, 2.0),
result_array[rows[2], 1],
places=5)
self.assertAlmostEqual(
get_out(5.0, 2.0, 4.0, 18.0, 2.0),
result_array[rows[2], 8],
places=5)