This repository was archived by the owner on Nov 17, 2023. It is now read-only.
This repository was archived by the owner on Nov 17, 2023. It is now read-only.
Log_Sigmoid bugs #20371
Description
Description
- When the input's
shape=(), the backward log_sigmoid activation is incorrect - There is an error when run the log_sigmoid in gpu.
Error Message And To Reproduce
git clone --recursive https://github.com/apache/incubator-mxnet
rm -rf build
mkdir -p build && cd build
cmake -GNinja \
-DUSE_CUDA=OFF \
-DUSE_MKL_IF_AVAILABLE=ON \
-DCMAKE_CUDA_COMPILER_LAUNCHER=ccache \
-DCMAKE_C_COMPILER_LAUNCHER=ccache \
-DCMAKE_CXX_COMPILER_LAUNCHER=ccache \
-DCMAKE_BUILD_TYPE=Release \
..
ninja
Use the following code to test. Reference: https://github.com/apache/incubator-mxnet/blob/master/tests/python/unittest/test_numpy_op.py
import numpy as _np
import mxnet as mx
from mxnet import np, npx
from mxnet.gluon import HybridBlock
from mxnet.base import MXNetError
from mxnet.test_utils import same, assert_almost_equal, rand_shape_nd, rand_ndarray
from mxnet.test_utils import check_numeric_gradient, use_np, collapse_sum_like, effective_dtype
@use_np
def test_activation():
def np_log_sigmoid(a):
return _np.log(_np.divide(1.0, _np.add(1.0, _np.exp(-a))))
def np_log_sigmoid_grad(a):
return _np.divide(1.0, _np.add(1.0, _np.exp(a)))
class TestLogSigmoid(HybridBlock):
def __init__(self):
super(TestLogSigmoid, self).__init__()
def hybrid_forward(self, F, a):
return F.npx.activation(a, act_type='log_sigmoid')
shape = ()
# shape = (1,)
test_log_sigmoid = TestLogSigmoid()
x = mx.np.random.uniform(low=-1.0, high=1.0, size=shape)
x.attach_grad()
np_out = np_log_sigmoid(x.asnumpy())
with mx.autograd.record():
mx_out = test_log_sigmoid(x)
assert mx_out.shape == np_out.shape
assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
mx_out.backward()
np_backward = np_log_sigmoid_grad(x.asnumpy())
print(np_backward)
print(x.grad.asnumpy())
assert_almost_equal(x.grad.asnumpy(), np_backward, rtol=1e-3, atol=1e-5)
mx_out = npx.activation(x, act_type='log_sigmoid')
np_out = np_log_sigmoid(x.asnumpy())
assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
@use_np
def test_activation_gpu():
def np_log_sigmoid(a):
return _np.log(_np.divide(1.0, _np.add(1.0, _np.exp(-a))))
def np_log_sigmoid_grad(a):
return _np.divide(1.0, _np.add(1.0, _np.exp(a)))
class TestLogSigmoid(HybridBlock):
def __init__(self):
super(TestLogSigmoid, self).__init__()
def hybrid_forward(self, F, a):
return F.npx.activation(a, act_type='log_sigmoid')
# shape = ()
shape = (1,)
test_log_sigmoid = TestLogSigmoid()
x = mx.np.random.uniform(low=-1.0, high=1.0, size=shape, ctx=mx.gpu())
x.attach_grad()
np_out = np_log_sigmoid(x.asnumpy())
with mx.autograd.record():
mx_out = test_log_sigmoid(x)
assert mx_out.shape == np_out.shape
assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
mx_out.backward()
np_backward = np_log_sigmoid_grad(x.asnumpy())
print(np_backward)
print(x.grad.asnumpy())
assert_almost_equal(x.grad.asnumpy(), np_backward, rtol=1e-3, atol=1e-5)
mx_out = npx.activation(x, act_type='log_sigmoid')
np_out = np_log_sigmoid(x.asnumpy())
assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
When run test_activation(), the error message is:
Traceback (most recent call last):
File "log_sigmoid_backward_bug.py", line 191, in <module>
test_activation()
File "/home/lthpc/anaconda3/lib/python3.6/site-packages/mxnet/util.py", line 299, in _with_np_shape
return func(*args, **kwargs)
File "/home/lthpc/anaconda3/lib/python3.6/site-packages/mxnet/util.py", line 480, in _with_np_array
return func(*args, **kwargs)
File "log_sigmoid_backward_bug.py", line 62, in test_activation
assert_almost_equal(x.grad.asnumpy(), np_backward, rtol=1e-3, atol=1e-5)
File "/home/lthpc/anaconda3/lib/python3.6/site-packages/mxnet/test_utils.py", line 740, in assert_almost_equal
raise AssertionError(msg)
AssertionError:
Items are not equal:
Error 371.466237 exceeds tolerance rtol=1.000000e-03, atol=1.000000e-05.
ACTUAL: array(0.65600127, dtype=float32)
DESIRED: 0.47561258889520563
But it can success when the shape=(1,) running on CPU.
When run test_activation_gpu(), the error message is:
[19:24:54] ../src/base.cc:80: cuDNN lib mismatch: linked-against version 8005 != compiled-against version 8004. Set MXNET_CUDNN_LIB_CHECKING=0 to quiet this warning.
[19:24:57] ../src/storage/storage.cc:199: Using Pooled (Naive) StorageManager for GPU
Traceback (most recent call last):
File "log_sigmoid_backward_bug.py", line 153, in <module>
test_activation_gpu()
File "/home/lthpc/anaconda3/lib/python3.6/site-packages/mxnet/util.py", line 299, in _with_np_shape
return func(*args, **kwargs)
File "/home/lthpc/anaconda3/lib/python3.6/site-packages/mxnet/util.py", line 480, in _with_np_array
return func(*args, **kwargs)
File "log_sigmoid_backward_bug.py", line 65, in test_activation_gpu
assert_almost_equal(mx_out.asnumpy(), np_out, rtol=1e-3, atol=1e-5)
File "/home/lthpc/anaconda3/lib/python3.6/site-packages/mxnet/ndarray/ndarray.py", line 2626, in asnumpy
ctypes.c_size_t(data.size)))
File "/home/lthpc/anaconda3/lib/python3.6/site-packages/mxnet/base.py", line 246, in check_call
raise get_last_ffi_error()
mxnet.base.MXNetError: Traceback (most recent call last):
File "../src/operator/nn/./cudnn/cudnn_activation-inl.h", line 61
MXNetError: Not implmented