Bug when iterating over HybridSequential elements

[feevos]

Description

Dear all, there is a bug when iterating over a HybridSequential treated as a container. This bug depends on the length of the container. If the length is small, the error does not appear. See minimal example below.

Error Message

---

MXNetError Traceback (most recent call last)

in 6 net.hybridize() 7 xx = nd.random.uniform(shape=[7,nfilters,F,F]) ----> 8 out = net(xx) ~/.local/lib/python3.6/site-packages/mxnet/gluon/block.py in __call__(self, *args) 691 hook(self, args) 692 --> 693 out = self.forward(*args) 694 695 for hook in self._forward_hooks.values(): ~/.local/lib/python3.6/site-packages/mxnet/gluon/block.py in forward(self, x, *args) 1146 'Find all contexts = {}'.format(ctx_set)) 1147 with ctx: -> 1148 return self._call_cached_op(x, *args) 1149 with ctx: 1150 try: ~/.local/lib/python3.6/site-packages/mxnet/gluon/block.py in _call_cached_op(self, *args) 979 def _call_cached_op(self, *args): 980 if self._cached_op is None: --> 981 self._build_cache(*args) 982 assert self._cached_op, "cached op is not None" 983 if self._callback: ~/.local/lib/python3.6/site-packages/mxnet/gluon/block.py in _build_cache(self, *args) 967 flags = [('data_indices', data_indices), ('param_indices', param_indices)] + \ 968 self._flags --> 969 self._cached_op = ndarray.CachedOp(out, flags) 970 971 def _deferred_infer_shape(self, *args): ~/.local/lib/python3.6/site-packages/mxnet/_ctypes/ndarray.py in __init__(self, sym, flags) 134 c_str_array([key for key, _ in flags]), 135 c_str_array([str(val) for _, val in flags]), --> 136 ctypes.byref(self.handle))) 137 138 def __del__(self): ~/.local/lib/python3.6/site-packages/mxnet/base.py in check_call(ret) 253 """ 254 if ret != 0: --> 255 raise MXNetError(py_str(_LIB.MXGetLastError())) 256 257 MXNetError: [14:27:32] src/core/graph.cc:110: Check failed: it != node2index_.end(): control dep not found in graph Stack trace: [bt] (0) /home/dia021/.local/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x65928b) [0x7f6e5b82f28b] [bt] (1) /home/dia021/.local/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x8385e68) [0x7f6e6355be68] [bt] (2) /home/dia021/.local/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x8386fe8) [0x7f6e6355cfe8] [bt] (3) /home/dia021/.local/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x8387b70) [0x7f6e6355db70] [bt] (4) /home/dia021/.local/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x389a709) [0x7f6e5ea70709] [bt] (5) /home/dia021/.local/lib/python3.6/site-packages/mxnet/libmxnet.so(MXCreateCachedOpEx+0x2cf) [0x7f6e5e97d72f] [bt] (6) /usr/lib/x86_64-linux-gnu/libffi.so.6(ffi_call_unix64+0x4c) [0x7f6edc06fdae] [bt] (7) /usr/lib/x86_64-linux-gnu/libffi.so.6(ffi_call+0x22f) [0x7f6edc06f71f] [bt] (8) /usr/lib/python3.6/lib-dynload/_ctypes.cpython-36m-x86_64-linux-gnu.so(_ctypes_callproc+0x2b4) [0x7f6edc2835c4] ## To Reproduce A simple example to reproduce the error, in this particular example if the length ```len(kernel_sizes)``` is greater or equal to 7 it produces an error. If it is smaller than 7 the error does not appear. ```Python class Demo(HybridBlock): def __init__(self, kernel_sizes = [3]*7,**kwards): super().__init__(**kwards) with self.name_scope(): self.net = gluon.nn.HybridSequential() for k in kernel_sizes: self.net.add(gluon.nn.Conv2D(32,kernel_size=k,padding=1)) def hybrid_forward(self, F, input): x = input for conv in self.net: x = x + conv(input) return x # This reproduces the error. nfilters=32 F = 256 net = Demo(kernel_sizes=[3]*7) # <=== CHANGE HERE, for length of list < 7 this script runs fine. net.initialize() net.hybridize() xx = nd.random.uniform(shape=[7,nfilters,F,F]) out = net(xx) ``` ### Steps to reproduce Run the script above ## What have you tried to solve it? I tried adding different inputs to the container, this affected the tolerance in the length in the for loop. I cannot explain it. ## Environment We recommend using our script for collecting the diagnositc information. Run the following command and paste the outputs below: ``` curl --retry 10 -s https://raw.githubusercontent.com/dmlc/gluon-nlp/master/tools/diagnose.py | python # paste outputs here ----------Python Info---------- Version : 3.6.8 Compiler : GCC 8.3.0 Build : ('default', 'Oct 7 2019 12:59:55') Arch : ('64bit', 'ELF') ------------Pip Info----------- Version : 9.0.1 Directory : /usr/lib/python3/dist-packages/pip ----------MXNet Info----------- Version : 1.6.0 Directory : /home/dia021/.local/lib/python3.6/site-packages/mxnet Num GPUs : 1 Commit Hash : 77e8f516e7f3d7d2a6c40387028a82ffd761909c ----------System Info---------- Platform : Linux-4.15.0-66-generic-x86_64-with-Ubuntu-18.04-bionic system : Linux node : vpro-fl release : 4.15.0-66-generic version : #75-Ubuntu SMP Tue Oct 1 05:24:09 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) i7-6920HQ CPU @ 2.90GHz Stepping: 3 CPU MHz: 2866.413 CPU max MHz: 3800.0000 CPU min MHz: 800.0000 BogoMIPS: 5808.00 Virtualisation: VT-x L1d cache: 32K L1i cache: 32K L2 cache: 256K L3 cache: 8192K 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 smx 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: 0.0004 sec, LOAD: 0.8517 sec. Timing for GluonNLP GitHub: https://github.com/dmlc/gluon-nlp, DNS: 0.0001 sec, LOAD: 0.6894 sec. Timing for GluonNLP: http://gluon-nlp.mxnet.io, DNS: 0.1959 sec, LOAD: 0.4118 sec. Timing for D2L: http://d2l.ai, DNS: 0.0546 sec, LOAD: 0.4319 sec. Timing for D2L (zh-cn): http://zh.d2l.ai, DNS: 0.0407 sec, LOAD: 0.3036 sec. Timing for FashionMNIST: https://repo.mxnet.io/gluon/dataset/fashion-mnist/train-labels-idx1-ubyte.gz, DNS: 0.1956 sec, LOAD: 0.8657 sec. Timing for PYPI: https://pypi.python.org/pypi/pip, DNS: 0.0413 sec, LOAD: 1.1598 sec. Timing for Conda: https://repo.continuum.io/pkgs/free/, DNS: 0.0411 sec, LOAD: 0.4481 sec. ```

[feevos]

Some additional information: it seems the error relates to how many times the initial input is passed from the conv layers. It is not directly related to the iteration over the HybridSequential container.

This works irrespective to what is the length of the kernel_sizes:

class Demo(HybridBlock):
    def __init__(self, kernel_sizes = [3,3,3,3],**kwards):
        super().__init__(**kwards)

        with self.name_scope():
            self.net = gluon.nn.HybridSequential()
            for k in kernel_sizes:
                tnet = gluon.nn.HybridSequential()
                for _ in range(3):
                    tnet.add(gluon.nn.Conv2D(32,kernel_size=k,padding=1))
                self.net.add(tnet)
    def hybrid_forward(self, F, input):
        x = input
        for conv in self.net:
            #x = x + conv(input)
            x = x + conv(x) ## <===== CHANGE HERE 

        return x

Runs fine:

nfilters=32
F = 256

net = Demo(kernel_sizes=[3]*100)
net.initialize()
net.hybridize()
xx = nd.random.uniform(shape=[7,nfilters,F,F])
out = net(xx)

[feevos]

Workaround that solves the problem (at some computational cost, I guess...):

class Demo(HybridBlock):
    def __init__(self, kernel_sizes = [3,3,3,3],**kwards):
        super().__init__(**kwards)

        with self.name_scope():
            self.net = gluon.nn.HybridSequential()
            for k in kernel_sizes:
                tnet = gluon.nn.HybridSequential()
                for _ in range(3):
                    tnet.add(gluon.nn.Conv2D(32,kernel_size=k,padding=1))
                self.net.add(tnet)
    def hybrid_forward(self, F, input):
        x = input
        for conv in self.net:
            x = F.identity(x) # <====== CHANGE HERE 
            x = x + conv(input)

        return x

[samskalicky]

@zachgk assign @szha

[feevos]

Dear all,

using the new version of mxnet (2.0) solves this problem:

In [1]: import mxnet as mx 
   ...: from mxnet import nd, gluon 
   ...: from mxnet.gluon import HybridBlock 
   ...: from mxnet import np, npx 
   ...: npx.set_np() 
   ...: class Demo(HybridBlock): 
   ...:     def __init__(self, kernel_sizes = [3]*17,**kwards): 
   ...:         super().__init__(**kwards) 
   ...:          
   ...:         
   ...:         self.net = gluon.nn.HybridSequential() 
   ...:         for k in kernel_sizes: 
   ...:             self.net.add(gluon.nn.Conv2D(32,kernel_size=k,padding=1)) 
   ...:                  
   ...:     def forward(self,input): 
   ...:         x = input 
   ...:         for conv in self.net: 
   ...:             x = x + conv(input) 
   ...:         
   ...:         return x 
   ...:  
   ...: # This reproduces the error.  
   ...: nfilters=32 
   ...: F = 256 
   ...:  
   ...: net = Demo(kernel_sizes=[3]*7) # <=== CHANGE HERE, for length of list < 7 this script runs fine.   
   ...: net.initialize() 
   ...: net.hybridize() 
   ...: xx = np.random.rand(7,nfilters,F,F) 
   ...: out = net(xx)                                                                                                                                             
/usr/local/lib/python3.7/site-packages/joblib/_multiprocessing_helpers.py:45: UserWarning: [Errno 28] No space left on device.  joblib will operate in serial mode
  warnings.warn('%s.  joblib will operate in serial mode' % (e,))
[18:14:54] ../src/storage/storage.cc:199: Using Pooled (Naive) StorageManager for CPU

In [2]: out = net(xx)                                                                                                                                             

In [3]: out.shape                                                                                                                                                 
Out[3]: (7, 32, 256, 256)