Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

[Bug Report] sigmoid_cross_entropy_with_logits 算子的小算子自动微分与调用反向kernel的计算结果不一致 #64226

Closed
zeroRains opened this issue May 11, 2024 · 3 comments
Closed

[Bug Report] sigmoid_cross_entropy_with_logits 算子的小算子自动微分与调用反向kernel的计算结果不一致 #64226

zeroRains opened this issue May 11, 2024 · 3 comments
Assignees
@cuicheng01
Labels
PFCC Paddle Framework Contributor Club，https://github.com/PaddlePaddle/community/tree/master/pfcc status/close 已关闭 type/bug-report 报bug

Comments

@zeroRains
Copy link
Contributor

zeroRains commented May 11, 2024 •
edited

bug描述 Describe the Bug

在实现sigmoid_cross_entropy_with_logits op的拆解时，用paddle api去实现对应的功能，前向计算得到相同的结果，但是反向计算时产生了精度问题，推测是小算子的自动微分和算子反向计算kernel存在差异。

复现代码如下：

import numpy as np
import paddle

np.random.seed(2023)
paddle.seed(2023)

batch_size = 20
num_classes = 10

x = np.random.uniform(0, 1, (batch_size, num_classes)).astype(
    "float32"
)

lable = np.random.uniform(0, 1, (batch_size, num_classes)).astype(
    "float32"
)

pos_weight = np.random.uniform(0, 1, (batch_size, num_classes)).astype(
    "float32"
)

# pos_weight = np.ones((batch_size, num_classes)).astype("float32")


def fn_ref(x, label, weight):
    out = paddle._C_ops.sigmoid_cross_entropy_with_logits(
        x, label, weight, False, -100)
    loss = out.sum()
    loss.backward()
    return out, x.grad


def fn_comp(x, label, weight):
    zeros = paddle.full((batch_size, num_classes), 0.)
    t1 = paddle.where(x > 0, x, zeros)
    t2 = x * label
    t3 = paddle.log(1 + paddle.exp(-paddle.abs(x)))
    t4 = t1 - t2 + t3 * weight
    t5 = paddle.full((batch_size, num_classes), -100.)
    out = paddle.where(label == t5, zeros, t4)
    loss = out.sum()
    loss.backward()
    return out, x.grad


def cal(fn):
    x1 = paddle.to_tensor(x, stop_gradient=False)
    label1 = paddle.to_tensor(lable)
    pos_weight1 = paddle.to_tensor(pos_weight)
    res = fn(x1, label1, pos_weight1)
    # print(res)
    return res


ref = cal(fn_ref)
actual = cal(fn_comp)


for idx in range(len(ref)):
    np.testing.assert_allclose(ref[idx].numpy(), actual[idx].numpy(
    ), atol=1e-6, rtol=1e-6, err_msg=f"****{idx} index error******")

BUG截图：
image

其他补充信息 Additional Supplementary Information

目前基本可以判断BUG产生的原因在于pos_weight的引入，当不存在可选参数pos_weight时，默认使用全1的 Tensor 代替，这时候自动微分和kernel反向计算的结果一致，但是当他们不是全1时，结果就会产生偏差。

具体分析如下：
kernel中有关pos_weight部分前向计算的代码：

# paddle/phi/kernels/cpu/sigmoid_cross_entropy_with_logits_grad_kernel.cc L48-L52

      T pos_weight_idx = pos_weight_data == nullptr ? 1 : pos_weight_data[idx];
      T term1 = (x > 0) ? x : 0;
      T term2 = x * label;
      T term3 = std::log(static_cast<T>(1) + std::exp(-std::abs(x)));
      out_data[idx] = term1 - term2 + term3 * pos_weight_idx;

可以用公式表示为：

$$ res = x - x*label + In(1+e^{-x})*posWeight $$

对公式求x的偏导如下：

$$ \frac {\partial_{res}}{\partial_x} = 1-label + \frac{-e^{-x} * posWeight}{1+e^{-x}}\\ =\frac{1+e^{-x}}{1+e^{-x}}-label + \frac{-e^{-x} * posWeight}{1+e^{-x}} \\ =\frac{1+e^{-x}-e^{-x} * posWeight}{1+e^{-x}}-label\\ =\frac{1+(1-posWeight)*e^{-x}}{1+e^{-x}}-label $$

但反向计算的代码如下：

# paddle/phi/kernels/cpu/sigmoid_cross_entropy_with_logits_grad_kernel.cc L50-L52

      T simoid_x = static_cast<T>(1) / (static_cast<T>(1) + std::exp(-x));
      T diff = simoid_x * pos_weight_idx - label;
      dx_data[idx] = dout * diff;

对应的公式如下：

$$ \frac {\partial_{res}}{\partial_x} = \frac{posWeight}{1+e^{-x}} - label $$

所以才会在posWeight不为全1Tensor的时候产生差异，不知道我的分析是否正确，希望能够查看一下。
@zeroRains zeroRains changed the title sigmoid_cross_entropy_with_logits 算子的拆解计算与kernel调用计算的方向计算结果不一致 sigmoid_cross_entropy_with_logits 算子的拆解计算与kernel调用计算的反向计算结果不一致 May 11, 2024
@paddle-bot paddle-bot bot added the PFCC Paddle Framework Contributor Club，https://github.com/PaddlePaddle/community/tree/master/pfcc label May 11, 2024
@zeroRains zeroRains mentioned this issue May 13, 2024
Merged
@zeroRains
Copy link
Contributor Author

分析部分有点问题，由于在推导过程中忽略了前向计算中使用的std::abs和 T term1 = (x > 0) ? x : 0;的梯度计算，所以现在修改前向计算公式如下：

$$ res = where(x>0,x,0) - x*label + In(1+e^{-|x|})*posWeight $$

经过推导得到的反向梯度计算为：

$$ \frac {\partial_{res}}{\partial_x} = where(x>0 , 1, 0)-label + \frac{-e^{-|x|} * where(x>=0 ,1, -1)* posWeight}{1+e^{-x}} $$

其中where(x>0, 1, 0)是前向计算中T term1 = (x > 0) ? x : 0;的梯度，where(x>=0,1, -1)是std::abs的梯度

对应的修复PR：

@zeroRains
Copy link
Contributor Author

kernel反向计算的结果，向numpy中采用数值求解的方式(见源码：op_test.py#L148-L323）计算的结果对齐，而拆解算子执行梯度的方式是通过自动微分求解的，其与kernel反向计算结果对齐。推断是Kernel反向实现的计算，存在问题。验证如下：

在执行sigmoid_cross_entropy_with_logits op的TestSigmoidCrossEntropyWithLogitsOp4中，可以观察到相对误差容忍阈值max_relative_error=0.005，设置得比较大，此时当前develop分支对反向kernel的实现可以通过此单测（虽然通过了，但是肉眼可见两个tensor确实有一些不同）

W0515 05:27:49.445072 36810 gpu_resources.cc:119] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 12.4, Runtime API Version: 11.2
W0515 05:27:49.449699 36810 gpu_resources.cc:164] device: 0, cuDNN Version: 8.1.
numeric : 
[array([[ 3.77183495e-04, -2.20797240e-04, -3.14791652e-04, ...,
        -2.02644761e-04,  2.45672779e-04, -6.17090210e-06],
       [ 4.84435251e-04,  2.84202716e-04,  1.83931716e-05, ...,
         6.29521346e-04,  5.20318281e-04,  2.33842612e-05],
       [ 3.12574747e-04, -4.71084098e-04,  1.10182442e-04, ...,
         6.98864401e-04,  2.33956572e-04, -7.56920161e-05],
       ...,
       [ 7.15934865e-04, -3.74937504e-04,  3.26225586e-04, ...,
         3.84216391e-05, -5.20641936e-04, -4.17575856e-04],
       [ 1.96946960e-05,  3.88698082e-04, -2.81023718e-04, ...,
        -5.38852117e-05,  3.67850861e-04, -1.84393860e-04],
       [-9.46350590e-05,  1.44749951e-05, -2.59066396e-04, ...,
         5.43415898e-04,  5.17161748e-05,  5.20940836e-04]])]
analytic_grads : 
[array([[ 1.84299020e-04, -2.20797405e-04, -3.14791844e-04, ...,
        -2.02644764e-04,  1.03724500e-04, -3.28235993e-04],
       [ 1.30288861e-04, -4.92659295e-04, -9.86970736e-05, ...,
         3.65435473e-04,  4.38155145e-04, -7.09606712e-04],
       [-1.62492418e-04, -4.71084140e-04,  1.10182313e-04, ...,
         1.25990168e-04,  1.87285167e-04, -5.22634377e-04],
       ...,
       [-3.11346327e-05, -3.74937645e-04,  3.26225574e-04, ...,
        -3.02651920e-04, -5.20642019e-04, -4.17576055e-04],
       [-3.08952871e-04,  2.82421633e-04, -2.81023766e-04, ...,
        -5.38854093e-05,  6.31943427e-05, -1.84394142e-04],
       [-1.67904546e-04, -1.19940036e-05, -2.59066405e-04, ...,
         3.36552688e-04,  2.25882243e-05, -9.09629301e-05]])]
max_relative_error : 
0.005
.
----------------------------------------------------------------------
Ran 1 test in 2.453s

OK

但是当我把这个容忍阈值改为max_relative_error=0.0005时，则会得到如下结果。

I0515 06:12:58.692179 17707 program_interpreter.cc:221] New Executor is Running.
I0515 06:12:58.693336 17707 interpreter_util.cc:652] Standalone Executor is Used.
numeric : 
[array([[ 3.77183495e-04, -2.20797240e-04, -3.14791652e-04, ...,
        -2.02644761e-04,  2.45672779e-04, -6.17090210e-06],
       [ 4.84435251e-04,  2.84202716e-04,  1.83931716e-05, ...,
         6.29521346e-04,  5.20318281e-04,  2.33842612e-05],
       [ 3.12574747e-04, -4.71084098e-04,  1.10182442e-04, ...,
         6.98864401e-04,  2.33956572e-04, -7.56920161e-05],
       ...,
       [ 7.15934865e-04, -3.74937504e-04,  3.26225586e-04, ...,
         3.84216391e-05, -5.20641936e-04, -4.17575856e-04],
       [ 1.96946960e-05,  3.88698082e-04, -2.81023718e-04, ...,
        -5.38852117e-05,  3.67850861e-04, -1.84393860e-04],
       [-9.46350590e-05,  1.44749951e-05, -2.59066396e-04, ...,
         5.43415898e-04,  5.17161748e-05,  5.20940836e-04]])]
analytic_grads : 
[array([[ 1.84299020e-04, -2.20797405e-04, -3.14791844e-04, ...,
        -2.02644764e-04,  1.03724500e-04, -3.28235993e-04],
       [ 1.30288861e-04, -4.92659295e-04, -9.86970736e-05, ...,
         3.65435473e-04,  4.38155145e-04, -7.09606712e-04],
       [-1.62492418e-04, -4.71084140e-04,  1.10182313e-04, ...,
         1.25990168e-04,  1.87285167e-04, -5.22634377e-04],
       ...,
       [-3.11346327e-05, -3.74937645e-04,  3.26225574e-04, ...,
        -3.02651920e-04, -5.20642019e-04, -4.17576055e-04],
       [-3.08952871e-04,  2.82421633e-04, -2.81023766e-04, ...,
        -5.38854093e-05,  6.31943427e-05, -1.84394142e-04],
       [-1.67904546e-04, -1.19940036e-05, -2.59066405e-04, ...,
         3.36552688e-04,  2.25882243e-05, -9.09629301e-05]])]
max_relative_error : 
0.0005
F
======================================================================
FAIL: test_check_grad (__main__.TestSigmoidCrossEntropyWithLogitsOp4)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/paddle/test/deprecated/legacy_test/test_sigmoid_cross_entropy_with_logits_op.py", line 178, in test_check_grad
    self.check_grad(['X'], 'Out', check_pir=True)
  File "/paddle/build/test/legacy_test/op_test.py", line 2986, in check_grad
    self.check_grad_with_place(
  File "/paddle/build/test/legacy_test/op_test.py", line 3298, in check_grad_with_place
    numeric_grads = self.check_grad_with_place_for_static(
  File "/paddle/build/test/legacy_test/op_test.py", line 3089, in check_grad_with_place_for_static
    self._assert_is_close(
  File "/paddle/build/test/legacy_test/op_test.py", line 2942, in _assert_is_close
    self.assertLessEqual(max_diff, max_relative_error, err_msg())
AssertionError: 0.0007811970012982192 not less than or equal to 0.0005 : Operator sigmoid_cross_entropy_with_logits error, Gradient Check On Place(cpu) variable X (shape: (64, 20), dtype: float64) max gradient diff 7.811970e-04 over limit 5.000000e-04, the first error element is 3, expected 5.481218e-04, but got 2.099690e-05.

----------------------------------------------------------------------
Ran 1 test in 0.521s

FAILED (failures=1)

因此可以推断，是由于容忍阈值比较大，所以使得反向计算错误的问题没有暴露出来。

在修复pr将max_relative_error=0.0005，仍然可以得到相对正确的计算结果，如下图：

W0515 06:13:53.214535 18318 gpu_resources.cc:119] Please NOTE: device: 0, GPU Compute Capability: 7.0, Driver API Version: 12.4, Runtime API Version: 11.2
W0515 06:13:53.220155 18318 gpu_resources.cc:164] device: 0, cuDNN Version: 8.1.
numeric : 
[array([[ 3.77183495e-04, -2.20797240e-04, -3.14791652e-04, ...,
        -2.02644761e-04,  2.45672779e-04, -6.17090210e-06],
       [ 4.84435251e-04,  2.84202716e-04,  1.83931716e-05, ...,
         6.29521346e-04,  5.20318281e-04,  2.33842612e-05],
       [ 3.12574747e-04, -4.71084098e-04,  1.10182442e-04, ...,
         6.98864401e-04,  2.33956572e-04, -7.56920161e-05],
       ...,
       [ 7.15934865e-04, -3.74937504e-04,  3.26225586e-04, ...,
         3.84216391e-05, -5.20641936e-04, -4.17575856e-04],
       [ 1.96946960e-05,  3.88698082e-04, -2.81023718e-04, ...,
        -5.38852117e-05,  3.67850861e-04, -1.84393860e-04],
       [-9.46350590e-05,  1.44749951e-05, -2.59066396e-04, ...,
         5.43415898e-04,  5.17161748e-05,  5.20940836e-04]])]
analytic_grads : 
[array([[ 3.77183699e-04, -2.20797405e-04, -3.14791844e-04, ...,
        -2.02644764e-04,  2.45672821e-04, -6.17087178e-06],
       [ 4.84435362e-04,  2.84202718e-04,  1.83934196e-05, ...,
         6.29521507e-04,  5.20318417e-04,  2.33842613e-05],
       [ 3.12574821e-04, -4.71084140e-04,  1.10182313e-04, ...,
         6.98864469e-04,  2.33956823e-04, -7.56920088e-05],
       ...,
       [ 7.15934877e-04, -3.74937645e-04,  3.26225574e-04, ...,
         3.84217363e-05, -5.20642019e-04, -4.17576055e-04],
       [ 1.96948667e-05,  3.88698345e-04, -2.81023766e-04, ...,
        -5.38854093e-05,  3.67850951e-04, -1.84394142e-04],
       [-9.46347782e-05,  1.44751433e-05, -2.59066405e-04, ...,
         5.43416127e-04,  5.17164533e-05,  5.20940836e-04]])]
max_relative_error : 
0.0005
.
----------------------------------------------------------------------
Ran 1 test in 2.753s

OK

@cyber-pioneer cyber-pioneer changed the title sigmoid_cross_entropy_with_logits 算子的拆解计算与kernel调用计算的反向计算结果不一致 [Bug Report] sigmoid_cross_entropy_with_logits 算子的小算子自动微分与调用反向kernel的计算结果不一致 May 16, 2024
@zeroRains
Copy link
Contributor Author

BUG已修复，详细见PR

@paddle-bot paddle-bot bot added status/close 已关闭 and removed status/new-issue 新建 labels May 16, 2024
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@cuicheng01 cuicheng01

Labels
PFCC Paddle Framework Contributor Club，https://github.com/PaddlePaddle/community/tree/master/pfcc status/close 已关闭 type/bug-report 报bug
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
2 participants
@cuicheng01 @zeroRains