Incubator-mxnet: LSTM with MKL-DNN produces wrong output after weights are changed

[matteosal]

Description

1) Create an RNN op with mode='lstm' and bind it
2) Run a forward pass
3) Change the NDArray holding the RNN parameters
4) Run a forward pass again

The output doesn't change, unless the second forward pass is performed in training mode (is_train=True). Setting MXNET_MKLDNN_ENABLED=0 doesn't fix the issue, but using a build without MKL-DNN does.

This severly impacts training with a validation set, because evaluating the performance on the validation set is typically performed with is_train=False after several updates of the weights. In this case, validation shows no improvement because the output of the layer is stuck at the very first training iteration.

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
Commit hash file "/home/matteo/Git/mxnet/python/mxnet/COMMIT_HASH" not found. Not installed from pre-built package or built from source.
Library      : ['/home/matteo/Git/mxnet/python/mxnet/../../lib/libmxnet.so']
Build features:
✖ CUDA
✖ CUDNN
✖ NCCL
✖ CUDA_RTC
✖ TENSORRT
✔ CPU_SSE
✔ CPU_SSE2
✔ CPU_SSE3
✔ CPU_SSE4_1
✔ CPU_SSE4_2
✖ CPU_SSE4A
✔ CPU_AVX
✖ CPU_AVX2
✖ OPENMP
✖ SSE
✔ F16C
✔ JEMALLOC
✖ BLAS_OPEN
✔ BLAS_ATLAS
✖ BLAS_MKL
✖ BLAS_APPLE
✖ LAPACK
✔ MKLDNN
✖ OPENCV
✖ CAFFE
✖ PROFILER
✖ DIST_KVSTORE
✖ CXX14
✖ INT64_TENSOR_SIZE
✖ SIGNAL_HANDLER
✖ DEBUG
----------System Info----------
Platform     : Linux-4.15.0-55-generic-x86_64-with-debian-buster-sid
system       : Linux
node         : mongolius
release      : 4.15.0-55-generic
version      : #60-Ubuntu SMP Tue Jul 2 18:22:20 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:             2700.253
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: 0.0117 sec, LOAD: 0.8935 sec.
Timing for Gluon Tutorial(en): http://gluon.mxnet.io, DNS: 0.0599 sec, LOAD: 2.1901 sec.
Timing for Gluon Tutorial(cn): https://zh.gluon.ai, DNS: 0.1028 sec, LOAD: 0.9832 sec.
Timing for FashionMNIST: https://apache-mxnet.s3-accelerate.dualstack.amazonaws.com/gluon/dataset/fashion-mnist/train-labels-idx1-ubyte.gz, DNS: 0.0657 sec, LOAD: 1.2597 sec.
Timing for PYPI: https://pypi.python.org/pypi/pip, DNS: 0.0380 sec, LOAD: 0.8543 sec.
Timing for Conda: https://repo.continuum.io/pkgs/free/, DNS: 0.0395 sec, LOAD: 0.4625 sec.

Package used (Python/R/Scala/Julia): python

Build info (Required if built from source)

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

MXNet commit hash: 076b2f330c60f05cb939beea28dd04cd571a34c0

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

Minimum reproducible example

import mxnet as mx

sym = mx.sym.RNN(mode='lstm', num_layers=1, state_outputs=False, state_size=1, name='rnn')

ex = sym.bind(mx.cpu(), 
    {
        'rnn_data': mx.ndarray.random.uniform(low=-1, high=1, shape=(10, 128, 5)),
        'rnn_parameters': mx.ndarray.random.uniform(low=-1, high=1, shape=(32)),
        'rnn_state': mx.ndarray.zeros(shape=(1, 128, 1)),
        'rnn_state_cell': mx.ndarray.zeros(shape=(1, 128, 1)),
    }
)

print('---- Output in training mode:')
ex.forward(is_train=True)
print(ex.output_dict['rnn_output'].sum().asnumpy())

print('\n---- Output in test mode:')
ex.forward(is_train=False)
print(ex.output_dict['rnn_output'].sum().asnumpy())

ex.copy_params_from(    
    {
        'rnn_data': ex.arg_dict['rnn_data'],
        'rnn_parameters': mx.ndarray.random.uniform(low=-1, high=1, shape=(32)),
        'rnn_state': ex.arg_dict['rnn_state'],
        'rnn_state_cell': ex.arg_dict['rnn_state_cell'],
    }
)

print('\n---- Output in training mode after changing weights:')
ex.forward(is_train=True)
print(ex.output_dict['rnn_output'].sum().asnumpy())

print('\n---- Output in test mode after changing weights:')
ex.forward(is_train=False)
print(ex.output_dict['rnn_output'].sum().asnumpy())

When using a build with MKL-DNN, this script print something like this:

---- Output in training mode:
[-112.02175]

---- Output in test mode:
[-112.02175]

---- Output in training mode after changing weights:
[-362.91537]

---- Output in test mode after changing weights:
[-112.02175]

Which shows that the output doesn't change after changing the weights unless the forward pass is performed in training mode. Setting MXNET_MKLDNN_ENABLED=0 doesn't fix the issue, but using a build without MKL-DNN does.

[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 label(s): Bug

[leleamol]

@mxnet-label-bot add [Bug, MKLDNN, RNN]

[ZhennanQin]

Probably it's because the stateful RNN op doesn't check if weight is changed. We will look at this. @pengzhao-intel

[ZhennanQin]

@zixuanweeei Would you please have a look for this?

[zixuanweeei]

@ZhennanQin Sure. Just as you have said, it is definitely caused by that stateful RNN op won't check weights again after it has been initialized with MKL-DNN memory format in inference procedure.

[zixuanweeei]

@matteosal Thanks for you reporting this issue. We are addressing the problem. PR is on the way. Thanks.

[matteosal]

Great!
I also see the same problem with modes rnn_relu and rnn_tanh, while gru is fine.

[zixuanweeei]

@matteosal That's right. The problem won't apear with GRU, because we haven't integrated MKL-DNN GRU into MXNet yet. It will be available in the near future.

[pengzhao-intel]

@matteosal thanks to reporting the issues which are really helpful.
Would you mind introducing some background of how your organizations are using MKL-DNN so that we can have better supports or more cooperations?
My official email: patric.[email protected]

[matteosal]

@matteosal thanks to reporting the issues which are really helpful.
Would you mind introducing some background of how your organizations are using MKL-DNN so that we can have better supports or more cooperations?
My official email: patric.[email protected]

Sure, I'm writing you from my Wolfram work email

[pengzhao-intel]

Fixed and closing. Thanks to reporting the issue :)