MKL-DNN gives wrong convolution bias gradient if weights gradient is not requested

[matteosal]

Description

When using MKL-DNN and asking the gradient of a convolution with respect to its biases, the result is wrong unless the gradient with respect to the weights is also requested.

Environment info (Required)

----------Python Info----------
Version      : 3.7.2
Compiler     : GCC 7.3.0
Build        : ('default', 'Dec 29 2018 06:19:36')
Arch         : ('64bit', '')
------------Pip Info-----------
Version      : 19.0.1
Directory    : /opt/Anaconda/lib/python3.7/site-packages/pip
----------MXNet Info-----------
Version      : 1.5.0
Directory    : /home/matteo/Git/mxnet/python/mxnet
Hashtag not found. Not installed from pre-built package.
----------System Info----------
Platform     : Linux-4.15.0-54-generic-x86_64-with-debian-buster-sid
system       : Linux
node         : mongolius
release      : 4.15.0-54-generic
version      : #58-Ubuntu SMP Mon Jun 24 10:55:24 UTC 2019
----------Hardware Info----------
machine      : x86_64
processor    : x86_64
Architecture:        x86_64
CPU op-mode(s):      32-bit, 64-bit
Byte Order:          Little Endian
CPU(s):              8
On-line CPU(s) list: 0-7
Thread(s) per core:  2
Core(s) per socket:  4
Socket(s):           1
NUMA node(s):        1
Vendor ID:           GenuineIntel
CPU family:          6
Model:               94
Model name:          Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz
Stepping:            3
CPU MHz:             2700.094
CPU max MHz:         3500,0000
CPU min MHz:         800,0000
BogoMIPS:            5184.00
Virtualization:      VT-x
L1d cache:           32K
L1i cache:           32K
L2 cache:            256K
L3 cache:            6144K
NUMA node0 CPU(s):   0-7
Flags:               fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb invpcid_single pti ssbd ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid fsgsbase tsc_adjust bmi1 hle avx2 smep bmi2 erms invpcid rtm mpx rdseed adx smap clflushopt intel_pt xsaveopt xsavec xgetbv1 xsaves dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp md_clear flush_l1d
----------Network Test----------
Setting timeout: 10
Timing for MXNet: https://github.com/apache/incubator-mxnet, DNS: 0.0010 sec, LOAD: 1.0852 sec.
Timing for Gluon Tutorial(en): http://gluon.mxnet.io, DNS: 0.1153 sec, LOAD: 0.9477 sec.
Timing for Gluon Tutorial(cn): https://zh.gluon.ai, DNS: 0.1108 sec, LOAD: 0.8710 sec.
Timing for FashionMNIST: https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/fashion-mnist/train-labels-idx1-ubyte.gz, DNS: 0.0825 sec, LOAD: 1.2461 sec.
Timing for PYPI: https://pypi.python.org/pypi/pip, DNS: 0.0351 sec, LOAD: 1.1176 sec.
Timing for Conda: https://repo.continuum.io/pkgs/free/, DNS: 0.0400 sec, LOAD: 0.5449 sec.

Using the python interface

Build info (Required if built from source)

Compiler (gcc/clang/mingw/visual studio): gcc

MXNet commit hash: 6a8d9eb

Build config: unchanged config.mk, except for USE_OPENCV = 0

Minimum reproducible example

import mxnet as mx

sym = mx.sym.Convolution(
	mx.sym.Variable('in'), 
	mx.sym.Variable('w'), 
	mx.sym.Variable('b'),
	kernel=(1, 1), 
	num_filter=1
)
args = {
	'in': mx.nd.ones([1, 1, 3, 3]),
	'w': mx.nd.ones([1, 1, 1, 1]),
	'b': mx.nd.ones([1]),
}
grad1 = {
	'in': mx.nd.zeros([1, 1, 3, 3]),
	'w': mx.nd.zeros([1, 1, 1, 1]),
	'b': mx.nd.zeros([1]),
}
grad2 = {
	'in': mx.nd.zeros([1, 1, 3, 3]),
	'w': mx.nd.zeros([1, 1, 1, 1]),
	'b': mx.nd.zeros([1]),
}
req1 = {'in': 'null', 'w': 'write', 'b': 'write'}
req2 = {'in': 'null', 'w': 'null', 'b': 'write'}
outgrad = mx.nd.ones([1, 1, 3, 3])

ex1 = sym.bind(mx.cpu(), args, args_grad=grad1, grad_req=req1)
ex2 = sym.bind(mx.cpu(), args, args_grad=grad2, grad_req=req2)

ex1.forward(True);
ex1.backward(out_grads=outgrad);
ex2.forward(True);
ex2.backward(out_grads=outgrad);

print(grad1['b'])
print(grad2['b'])

The above script prints a wrong value (0) for grad2['b'], while grad1['b'] is correct (9):

[9.]
<NDArray 1 @cpu(0)>

[0.]
<NDArray 1 @cpu(0)>

running with MXNET_MKLDNN_ENABLED=0 produces the correct result (9) for both gradients

[mxnet-label-bot]

Hey, this is the MXNet Label Bot.
Thank you for submitting the issue! I will try and suggest some labels so that the appropriate MXNet community members can help resolve it.
Here are my recommended labels: Bug

[pengzhao-intel]

Thanks for reporting the potential issues.
@zixuanweeei could you help take a look for this?

[zixuanweeei]

@matteosal I think this is perhaps caused by the fact that the forward convolution operator supports calculation either with or without bias. And both cases require weights in the convolutional operator. So in the backward process, the gradient of bias could be either calculated or not. This leads to only two backward APIs supported by MKL-DNN corresponding to the cases of the forward operations. Actually, we can just take the gradients of the weights away. And we have located the problem. Further verification is required before taking a PR.
BTW, could you provide some examples that only requires the gradients of bias? We will very appreciate it. Thanks. 😃

[matteosal]

I don't have a use case for this. I'm reporting the bug because our unittests at Wolfram Research have spotted it.

[pengzhao-intel]

@matteosal no problem. we will create the case and file PR soon.

[matteosal]

Thanks!

[pengzhao-intel]

The fix is merged and close this issue.
Feel free to reopen if anything doesn't work as the expectation.