def test_sparse_adagrad_empty(self, inputs, lr, epsilon, gc, dc):
param, momentum = inputs
grad = np.empty(shape=(0, ) + param.shape[1:], dtype=np.float32)
ref_using_fp16_values = [False]
if gc == hu.gpu_do:
ref_using_fp16_values.append(True)
for ref_using_fp16 in ref_using_fp16_values:
if ref_using_fp16:
print(
"test_sparse_adagrad_empty with half precision embedding")
momentum_i = momentum.astype(np.float16)
param_i = param.astype(np.float16)
else:
print(
"test_sparse_adagrad_empty with full precision embedding")
momentum_i = momentum.astype(np.float32)
param_i = param.astype(np.float32)
adagrad_sparse_test_helper(
self,
[param_i, momentum_i, grad],
lr,
epsilon,
None,
ref_adagrad,
gc,
dc,
)
def test_sparse_adagrad(self, inputs, lr, epsilon, weight_decay, gc, dc):
adagrad_sparse_test_helper(
self,
inputs,
lr,
epsilon,
None,
ref_adagrad,
gc,
dc,
weight_decay=weight_decay,
)
def test_row_wise_sparse_adagrad(self, inputs, lr, epsilon, data_strategy, gc, dc):
adagrad_sparse_test_helper(
self,
inputs,
lr,
epsilon,
None,
functools.partial(ref_adagrad, row_wise=True),
gc,
dc,
row_wise=True,
)
def test_row_wise_sparse_adagrad_empty(self, inputs, lr, epsilon, gc, dc):
param, momentum = inputs
grad = np.empty(shape=(0, ) + param.shape[1:], dtype=np.float32)
adagrad_sparse_test_helper(
self,
[param, momentum, grad],
lr,
epsilon,
None,
ref_adagrad,
gc,
dc,
row_wise=True,
)
def test_row_wise_sparse_adagrad(self, inputs, lr, epsilon, weight_decay,
gc, dc):
adagrad_sparse_test_helper(
self,
inputs,
lr,
epsilon,
None,
functools.partial(ref_adagrad, row_wise=True),
gc,
dc,
row_wise=True,
weight_decay=weight_decay,
)
def test_sparse_adagrad(self, inputs, lr, epsilon, gc, dc):
return adagrad_sparse_test_helper(self, inputs, lr, epsilon, None,
ref_adagrad, gc, dc)
