代码拉取完成，页面将自动刷新
class SoftMarginLossMock(OpsFactory):
def __init__(self, attributes={'reduction': 'mean'}, inputs=None, grads=None):
self.ms_type = inputs[0].dtype
super().__init__(dtype=dtype_to_nptype(self.ms_type))
self.reduction = attributes.get('reduction', 'mean')
self.logits = inputs[0]
self.logits_np = inputs[0].asnumpy()
self.labels = inputs[1]
self.labels_np = inputs[1].asnumpy()
if grads is None:
self.out_grad_np = None
elif self.reduction == "sum" or self.reduction == "mean":
self.out_grad_np = np.array(grads)
elif self.reduction == "none":
self.out_grad_np = grads
def forward_mindspore_impl(self):
net = SoftMarginLoss(reduction=self.reduction)
out = net(self.logits, self.labels)
return out.asnumpy()
def forward_pytorch_impl(self):
net = torch.nn.SoftMarginLoss(reduction=self.reduction)
logits = torch.from_numpy(self.logits_np.astype(np.float32))
labels = torch.from_numpy(self.labels_np.astype(np.float32))
output = net(logits, labels)
return output.detach().numpy().astype(self.dtype)
def forward_cmp(self):
out_pytorch = self.forward_pytorch_impl()
out_mindspore = self.forward_mindspore_impl()
allclose_nparray(out_pytorch, out_mindspore, self.loss, self.loss)
def grad_mindspore_impl(self):
if self.out_grad_np is None:
self.out_grad_np = self.forward_pytorch_impl()
net = SoftMarginLoss(reduction=self.reduction)
grad_net = GradOfFirstInput(net)
grad_net.set_train()
out_grad = grad_net(self.logits, self.labels, Tensor(self.out_grad_np))
return out_grad.asnumpy()
def grad_pytorch_impl(self):
if self.out_grad_np is None:
self.out_grad_np = self.forward_pytorch_impl()
net = torch.nn.SoftMarginLoss(reduction=self.reduction)
logits = torch.from_numpy(self.logits_np.astype(np.float32))
labels = torch.from_numpy(self.labels_np.astype(np.float32))
logits.requires_grad = True
output = net(logits, labels)
output_grad = torch.from_numpy(self.out_grad_np.copy().astype(np.float32))
output.backward(gradient=output_grad)
return logits.grad.numpy().astype(self.dtype)
def grad_cmp(self):
input_grad_mindspore = self.grad_mindspore_impl()
input_grad_pytorch = self.grad_pytorch_impl()
allclose_nparray(input_grad_pytorch, input_grad_mindspore, self.loss, self.loss)
1，
GPU环境两种模式下，当输入数据类型为float16时，偶现精度问题
(用例执行十次，失败三次，通过七次)
collected 10 items
test_softmarginloss.py::test_softmarginloss_input_2d_dtype_float16[1-10] PASSED [ 10%]
test_softmarginloss.py::test_softmarginloss_input_2d_dtype_float16[2-10] PASSED [ 20%]
test_softmarginloss.py::test_softmarginloss_input_2d_dtype_float16[3-10] PASSED [ 30%]
test_softmarginloss.py::test_softmarginloss_input_2d_dtype_float16[4-10] PASSED [ 40%]
test_softmarginloss.py::test_softmarginloss_input_2d_dtype_float16[5-10] FAILED [ 50%]
test_softmarginloss.py::test_softmarginloss_input_2d_dtype_float16[6-10] PASSED [ 60%]
test_softmarginloss.py::test_softmarginloss_input_2d_dtype_float16[7-10] FAILED [ 70%]
test_softmarginloss.py::test_softmarginloss_input_2d_dtype_float16[8-10] FAILED [ 80%]
test_softmarginloss.py::test_softmarginloss_input_2d_dtype_float16[9-10] PASSED [ 90%]
test_softmarginloss.py::test_softmarginloss_input_2d_dtype_float16[10-10] PASSED [100%]
======================================================================== FAILURES =========================================================================
____________________________________________________ test_softmarginloss_input_2d_dtype_float16[5-10] _____________________________________________________
def test_softmarginloss_input_2d_dtype_float16():
input_list = []
logits = Tensor(np.random.randn(6, 8), dtype=mstype.float16)
labels = Tensor(np.random.randn(6, 8), dtype=mstype.float16)
input_list.append(logits)
input_list.append(labels)
fact = SoftMarginLossMock(attributes={'reduction': 'sum'}, inputs=input_list)
fact.loss = 1e-3
fact.forward_cmp()
> fact.grad_cmp()
test_softmarginloss.py:342:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
../share/ops/nn/softmarginloss_ops.py:78: in grad_cmp
allclose_nparray(input_grad_pytorch, input_grad_mindspore, self.loss, self.loss)
../share/utils.py:31: in allclose_nparray
_count_unequal_element(data_expected, data_me, rtol, atol)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
data_expected = array([[ -0.8887, 6.652 , -15.14 , -20.62 , 3.377 , 13.02 ,
40.5 , -26.78 ],
[ 32.34 , 1...996 ],
[ 8.805 , -11.79 , 46.84 , 4.184 , 3.857 , -10.2 ,
-8.9 , -55.94 ]], dtype=float16)
data_me = array([[ -0.889 , 6.652 , -15.15 , -20.64 , 3.377 , 13.016 ,
40.5 , -26.75 ],
[ 32.34 , 1... ],
[ 8.805 , -11.79 , 46.84 , 4.184 , 3.854 , -10.21 ,
-8.9 , -55.9 ]], dtype=float16)
rtol = 0.001, atol = 0.001
def _count_unequal_element(data_expected, data_me, rtol, atol):
assert data_expected.shape == data_me.shape
total_count = len(data_expected.flatten())
error = np.abs(data_expected - data_me)
greater = np.greater(error, atol + np.abs(data_me) * rtol)
loss_count = np.count_nonzero(greater)
assert (loss_count / total_count) < rtol, \
"\ndata_expected_std:{0}\ndata_me_error:{1}\nloss:{2}". \
> format(data_expected[greater], data_me[greater], error[greater])
E AssertionError:
E data_expected_std:[-26.78]
E data_me_error:[-26.75]
E loss:[0.03125]
2，
test_softmarginloss_input_5d_dtype_float16
def test_softmarginloss_input_5d_dtype_float16():
input_list = []
logits = Tensor(np.random.randn(64, 128, 5, 4, 2), dtype=mstype.float16)
labels = Tensor(np.random.randn(64, 128, 5, 4, 2), dtype=mstype.float16)
input_list.append(logits)
input_list.append(labels)
fact = SoftMarginLossMock(inputs=input_list)
fact.loss = 1e-3
fact.forward_cmp()
> fact.grad_cmp()
test_softmarginloss.py:421:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
../share/ops/nn/softmarginloss_ops.py:78: in grad_cmp
allclose_nparray(input_grad_pytorch, input_grad_mindspore, self.loss, self.loss)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
data_expected = array([[[[[-2.38e-07, 4.77e-07],
[ 5.96e-07, 1.67e-06],
[ 1.19e-06, -4.17e-07],
[-5.96... [ 1.97e-06, 7.75e-07],
[-1.37e-06, 5.96e-08],
[ 2.38e-07, -5.36e-07]]]]], dtype=float16)
data_me = array([[[[[-0., 0.],
[ 0., 0.],
[ 0., -0.],
[-0., -0.]],
[[-0., -0.],
... 0., 0.]],
[[-0., -0.],
[ 0., 0.],
[-0., 0.],
[ 0., -0.]]]]], dtype=float16)
rtol = 0.001, atol = 0.001, equal_nan = True
def allclose_nparray(data_expected, data_me, rtol, atol, equal_nan=True):
if np.any(np.isnan(data_expected)) or np.any(np.isnan(data_me)):
> assert np.allclose(data_expected, data_me, rtol, atol, equal_nan=equal_nan)
E AssertionError
Ascend/GPU/CPU) / 硬件环境:Please delete the backend not involved / 请删除不涉及的后端:
/device ascend/GPU/CPU/kirin/等其他芯片
Software Environment / 软件环境 (Mandatory / 必填):
-- MindSpore version (e.g., 1.7.0.Bxxx) :
-- Python version (e.g., Python 3.7.5) :
-- OS platform and distribution (e.g., Linux Ubuntu 16.04):
-- GCC/Compiler version (if compiled from source):
Excute Mode / 执行模式 (Mandatory / 必填)(PyNative/Graph):
Please delete the mode not involved / 请删除不涉及的模式:
/mode pynative
/mode graph
Please assign maintainer to check this issue.
请为此issue分配处理人。
@xuebao_zhang
此处可能存在不合适展示的内容，页面不予展示。您可通过相关编辑功能自查并修改。
如您确认内容无涉及 不当用语 / 纯广告导流 / 暴力 / 低俗色情 / 侵权 / 盗版 / 虚假 / 无价值内容或违法国家有关法律法规的内容，可点击提交进行申诉，我们将尽快为您处理。
Please add labels (comp or sig), also you can visit https://gitee.com/mindspore/community/blob/master/sigs/dx/docs/labels.md to find more.
为了让代码尽快被审核，请您为Pull Request打上 组件(comp)或兴趣组(sig) 标签，打上标签的PR可直接推送给责任人进行审核。
更多的标签可以查看https://gitee.com/mindspore/community/blob/master/sigs/dx/docs/labels.md
以组件相关代码提交为例，如果你提交的是data组件代码，你可以这样评论：
//comp/data
当然你也可以邀请data SIG组来审核代码，可以这样写：
//sig/data
另外你还可以给这个PR标记类型，例如是bugfix或者是特性需求：
//kind/bug or //kind/feature
恭喜你，你已经学会了使用命令来打标签，接下来就在下面的评论里打上标签吧！
用例问题，修改ops文件为以下方式：
if self.out_grad_np is None:
self.out_grad_np = np.array(
np.random.randn(*list(self.forward_pytorch_impl().shape))).astype(self.dtype)
登录 后才可以发表评论