dithyrambe opened a new issue #17275: CosineEmbeddingLoss won't work after hybridize (ie. with symbol API) URL: https://github.com/apache/incubator-mxnet/issues/17275 ## Description `CosineEmbeddingLoss` seems to have issue in its symbol API implementation. Loss works fine until I hybridize my model. Looks like its some reshaping issue. Below is the stacktrace and the snippet to reproduce. ``` Traceback (most recent call last): File "tmp.tmp", line 9, in <module> loss(x, x, labels) File "/opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/gluon/block.py", line 548, in __call__ out = self.forward(*args) File "/opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/gluon/block.py", line 915, in forward return self._call_cached_op(x, *args) File "/opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/gluon/block.py", line 821, in _call_cached_op out = self._cached_op(*cargs) File "/opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/_ctypes/ndarray.py", line 150, in __call__ ctypes.byref(out_stypes))) File "/opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/base.py", line 253, in check_call raise MXNetError(py_str(_LIB.MXGetLastError())) mxnet.base.MXNetError: Error in operator cosineembeddingloss0__div0: [12:59:44] /home/travis/build/dmlc/mxnet-distro/mxnet-build/3rdparty/mshadow/../../src/operator/tensor/../elemwise_op_common.h:135: Check failed: assign(&dattr, vec.at(i)): Incompatible attr in node cosineembeddingloss0__div0 at 1-th input: expected [2], got [1,2] Stack trace: [bt] (0) /opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x2795cb) [0x7fbbb2fc35cb] [bt] (1) /opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x3e8e63) [0x7fbbb3132e63] [bt] (2) /opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x3e9755) [0x7fbbb3133755] [bt] (3) /opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x9068f1) [0x7fbbb36508f1] [bt] (4) /opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x23c8262) [0x7fbbb5112262] [bt] (5) /opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x23cab2a) [0x7fbbb5114b2a] [bt] (6) /opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/libmxnet.so(+0x23e6b0f) [0x7fbbb5130b0f] [bt] (7) /opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/libmxnet.so(mxnet::CachedOp::CheckDynamicShapeExists(mxnet::Context const&, std::vector<mxnet::NDArray*, std::allocator<mxnet::NDArray*> > const&, bool)+0x3e3) [0x7fbbb5137ad3] [bt] (8) /opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet/libmxnet.so(mxnet::CachedOp::Forward(std::shared_ptr<mxnet::CachedOp> const&, std::vector<mxnet::NDArray*, std::allocator<mxnet::NDArray*> > const&, std::vector<mxnet::NDArray*, std::allocator<mxnet::NDArray*> > const&)+0xa5a) [0x7fbbb513e74a] ``` ## To Reproduce ```python import mxnet as mx x = mx.random.randn(2, 2) labels = mx.nd.ones(2) loss = mx.gluon.loss.CosineEmbeddingLoss() print(loss(x, x, labels)) # works fine loss.hybridize() print(loss(x, x, labels)) # shape issue ``` ## Attempt to solve I manage in to get around it by explicitly calling `F.reshape` instead of `<tensor>.reshape` in `CosineEmbeddingLoss._cosine_similarity` ```python import mxnet as mx from mxnet import ndarray from mxnet.gluon.loss import CosineEmbeddingLoss class FixedCosineEmbeddingLoss(CosineEmbeddingLoss): def _cosine_similarity(self, F, x, y, axis=-1): # Calculates the cosine similarity between 2 vectors x_norm = F.reshape(F.norm(x, axis=axis), shape=(-1, 1)) y_norm = F.reshape(F.norm(y, axis=axis), shape=(-1, 1)) x_dot_y = F.reshape(F.sum(x * y, axis=axis), shape=(-1, 1)) if F is ndarray: eps_arr = F.array([1e-12]) else: eps_arr = F.full((1, 1), 1e-12) return (x_dot_y / F.broadcast_maximum(x_norm * y_norm, eps_arr)) x = mx.random.randn(2, 2) labels = mx.nd.ones(2) loss = FixedCosineEmbeddingLoss() print(loss(x, x, labels)) loss.hybridize() print(loss(x, x, labels)) # works fine now ``` ## Environment ``` curl --retry 10 -s https://raw.githubusercontent.com/dmlc/gluon-nlp/master/tools/diagnose.py | python ----------Python Info---------- Version : 3.6.9 Compiler : GCC 7.3.0 Build : ('default', 'Jul 30 2019 19:07:31') Arch : ('64bit', '') ------------Pip Info----------- Version : 19.3.1 Directory : /opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/pip ----------MXNet Info----------- Version : 1.5.1 Directory : /opt/miniconda3/envs/deep-search/lib/python3.6/site-packages/mxnet Num GPUs : 0 Commit Hash : c9818480680f84daa6e281a974ab263691302ba8 ----------System Info---------- Platform : Linux-5.0.0-37-generic-x86_64-with-debian-buster-sid system : Linux node : lbctp release : 5.0.0-37-generic version : #40~18.04.1-Ubuntu SMP Thu Nov 14 12:06:39 UTC 2019 ----------Hardware Info---------- machine : x86_64 processor : x86_64 Architecture : x86_64 Mode(s) opératoire(s) des processeurs : 32-bit, 64-bit Boutisme : Little Endian Processeur(s) : 8 Liste de processeur(s) en ligne : 0-7 Thread(s) par cœur : 2 Cœur(s) par socket : 4 Socket(s) : 1 Nœud(s) NUMA : 1 Identifiant constructeur : GenuineIntel Famille de processeur : 6 Modèle : 142 Nom de modèle : Intel(R) Core(TM) i7-8550U CPU @ 1.80GHz Révision : 10 Vitesse du processeur en MHz : 2599.998 Vitesse maximale du processeur en MHz : 4000,0000 Vitesse minimale du processeur en MHz : 400,0000 BogoMIPS : 3984.00 Virtualisation : VT-x Cache L1d : 32K Cache L1i : 32K Cache L2 : 256K Cache L3 : 8192K Nœud NUMA 0 de processeur(s) : 0-7 Drapaux : 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 ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid 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: 0.0096 sec, LOAD: 0.6905 sec. Timing for GluonNLP GitHub: https://github.com/dmlc/gluon-nlp, DNS: 0.0007 sec, LOAD: 0.6103 sec. Timing for GluonNLP: http://gluon-nlp.mxnet.io, DNS: 0.0799 sec, LOAD: 0.5383 sec. Timing for D2L: http://d2l.ai, DNS: 0.0188 sec, LOAD: 0.2801 sec. Timing for D2L (zh-cn): http://zh.d2l.ai, DNS: 0.0181 sec, LOAD: 0.3634 sec. Timing for FashionMNIST: https://repo.mxnet.io/gluon/dataset/fashion-mnist/train-labels-idx1-ubyte.gz, DNS: 0.0261 sec, LOAD: 0.7186 sec. Timing for PYPI: https://pypi.python.org/pypi/pip, DNS: 0.0098 sec, LOAD: 0.3788 sec. Timing for Conda: https://repo.continuum.io/pkgs/free/, DNS: 0.0140 sec, LOAD: 0.0724 sec. ``` Thanks for having a look. Let me know if I should make a PR with the tiny modification. Regards, Dithyrambe
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services
