Hi, I am new to this task and want to know the relation between mean rank and mAP. I am reproducing the result of link prediction task and trained TransE model as well as Poincare model. When I evaluation those two models, I found that TransE may get higher mean rank and higher MAP ,but Poincare may get lower mean rank and lower MAP. Should it always be lower mean rank and higher MAP? Are there some differences between those two ways of evaluation? Or maybe there is something wrong with my evaluation code :(

mldl@ub1604:~/ub16_prj/poincare-embeddings$ NTHREADS=2 ./train-nouns.sh Using 2 threads slurp: objects=82115, edges=743086 Indexing data json_conf: {"distfn": "poincare", "dim": 10, "lr": 1, "batchsize": 50, "negs": 50} Burnin: lr=0.01 'Traceback (most recent call last):\n File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 57, in _worker_loop\n samples = collate_fn([dataset[i] for i in batch_indices])\n File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 57, in \n samples = collate_fn([dataset[i] for i in batch_indices])\n File "/home/mldl/ub16_prj/poincare-embeddings/model.py", line 185, in getitem\n if n not in self._weights[t]:\nKeyError: tensor(23511)\n' Traceback (most recent call last): File "/home/mldl/ub16_prj/poincare-embeddings/train.py", line 19, in train_mp train(model, data, optimizer, opt, log, rank, queue) File "/home/mldl/ub16_prj/poincare-embeddings/train.py", line 46, in train for inputs, targets in loader: File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 286, in next return self._process_next_batch(batch) File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 307, in _process_next_batch raise batch.exc_type(batch.exc_msg) KeyError: 'Traceback (most recent call last):\n File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 57, in _worker_loop\n samples = collate_fn([dataset[i] for i in batch_indices])\n File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 57, in \n samples = collate_fn([dataset[i] for i in batch_indices])\n File "/home/mldl/ub16_prj/poincare-embeddings/model.py", line 185, in getitem\n if n not in self._weights[t]:\nKeyError: tensor(23511)\n' 'Traceback (most recent call last):\n File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 57, in _worker_loop\n samples = collate_fn([dataset[i] for i in batch_indices])\n File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 57, in \n samples = collate_fn([dataset[i] for i in batch_indices])\n File "/home/mldl/ub16_prj/poincare-embeddings/model.py", line 185, in getitem\n if n not in self._weights[t]:\nKeyError: tensor(23511)\n' Traceback (most recent call last): File "/home/mldl/ub16_prj/poincare-embeddings/train.py", line 19, in train_mp train(model, data, optimizer, opt, log, rank, queue) File "/home/mldl/ub16_prj/poincare-embeddings/train.py", line 46, in train for inputs, targets in loader: File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 286, in next return self._process_next_batch(batch) File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 307, in _process_next_batch raise batch.exc_type(batch.exc_msg) KeyError: 'Traceback (most recent call last):\n File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 57, in _worker_loop\n samples = collate_fn([dataset[i] for i in batch_indices])\n File "/usr/local/lib/python3.6/dist-packages/torch/utils/data/dataloader.py", line 57, in \n samples = collate_fn([dataset[i] for i in batch_indices])\n File "/home/mldl/ub16_prj/poincare-embeddings/model.py", line 185, in getitem\n if n not in self._weights[t]:\nKeyError: tensor(23511)\n'

Hi, I am wondering is it normal for “train-nouns.py” to stop for a relatively long time after many epochs. I have run it for nearly 1 day and it does not have any outputs now.

Hi, I just train your code on pytorch 0.4.0, on the mammals dataset, with your default hyperparameters, I can just get an embedding of MAP around 0.4 after 300 epochs, which is far more less than your reported results 0.927 in your paper. BTW, I just add 6 lines of code in the form of t= int(t) because the version of pytorch, other codes remain the same. I am wondering how could this happen?

I wrote this question on stack overflow here... this is not so much an issue with the repo, as it is a question of how to implement these embeddings in practice.

My assumption for how to implement these is to take a sentence, like """The US and UK could agree a “phenomenal' trade deal after Britain leaves the EU.""", tokenize into a list of synsets [[], Synset('united_states.n.01'), [], Synset('united_kingdom.n.01'), ... ]... but in order to do that, one needs each unique representative synset node of the wordnet, based on the context in which that word lives.

This seems like a pretty difficult aspect of using these embeddings, and I'm wondering what strategies there are to solve this, what literature is available, whether this is totally not how their implemented in practice, or if there are open source projects which can take both the word and the sentence context into account to map a sentence to a list of "optimal" synsets. Seems like this is a critical aspect of using these (and wordnets in general), but not frequently discussed.

BTW many thanks for open-sourcing this, I find this technology really fascinating.

Thank you for sharing this great code. However, I encountered one valueError when I am trying to reproduce the train-mammals.sh. The error information can be as follows:

(poincare) C:\Users\DELL\Github\poincare-embeddings-master>sh train-mammals.sh
Specified hogwild training with GPU, defaulting to CPU...
Using edge list dataloader
Traceback (most recent call last):
  File "embed.py", line 246, in <module>
    main()
  File "embed.py", line 147, in main
    manifold, opt, idx, objects, weights, sparse=opt.sparse
  File "C:\Users\DELL\Github\poincare-embeddings-master\hype\sn.py", line 64, in initialize
    opt.ndproc, opt.burnin > 0, opt.dampening)
  File "hype\graph_dataset.pyx", line 75, in hype.graph_dataset.BatchedDataset.__cinit__
    self._mk_weights(idx, weights)
  File "hype\graph_dataset.pyx", line 81, in hype.graph_dataset.BatchedDataset._mk_weights
    def _mk_weights(self, npc.ndarray[npc.long_t, ndim=2] idx, npc.ndarray[npc.double_t, ndim=1] weights):
ValueError: Buffer dtype mismatch, expected 'long_t' but got 'long'

I am using pytorch 1.0 with cuda 10.0 on windows:

(poincare) C:\Users\DELL\Github\poincare-embeddings-master>python -c "import torch; print(torch.version.cuda)"
10.0
(poincare) C:\Users\DELL\Github\poincare-embeddings-master>nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2018 NVIDIA Corporation
Built on Sat_Aug_25_21:08:04_Central_Daylight_Time_2018
Cuda compilation tools, release 10.0, V10.0.130

I found that the problem comes from earlier steps python setup.py build_ext --inplace. When I run python setup.py build_ext --inplace, it shows the following information:

(poincare) C:\Users\DELL\Github\poincare-embeddings-master>python setup.py build_ext --inplace
Compiling hype/graph_dataset.pyx because it depends on C:\Anaconda3\envs\poincare\lib\site-packages\Cython\Includes\numpy\__init__.pxd.
Compiling hype/adjacency_matrix_dataset.pyx because it depends on C:\Anaconda3\envs\poincare\lib\site-packages\Cython\Includes\numpy\__init__.pxd.
[1/2] Cythonizing hype/adjacency_matrix_dataset.pyx
C:\Anaconda3\envs\poincare\lib\site-packages\Cython\Compiler\Main.py:367: FutureWarning: Cython directive 'language_level' not set, using 2 for now (Py2). This will change in a later release! File: C:\Users\DELL\Github\poincare-embeddings-master\hype\adjacency_matrix_dataset.pyx
  tree = Parsing.p_module(s, pxd, full_module_name)
[2/2] Cythonizing hype/graph_dataset.pyx
C:\Anaconda3\envs\poincare\lib\site-packages\Cython\Compiler\Main.py:367: FutureWarning: Cython directive 'language_level' not set, using 2 for now (Py2). This will change in a later release! File: C:\Users\DELL\Github\poincare-embeddings-master\hype\graph_dataset.pyx
  tree = Parsing.p_module(s, pxd, full_module_name)
running build_ext
building 'hype.graph_dataset' extension
C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\bin\HostX86\x64\cl.exe /c /nologo /Ox /W3 /GL /DNDEBUG /MD -IC:\Anaconda3\envs\poincare\lib\site-packages\numpy\core\include -IC:\Anaconda3\envs\poincare\include -IC:\Anaconda3\envs\poincare\include "-IC:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\ATLMFC\include" "-IC:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\include" "-IC:\Program Files (x86)\Windows Kits\NETFXSDK\4.6.1\include\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.17763.0\ucrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.17763.0\shared" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.17763.0\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.17763.0\winrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.17763.0\cppwinrt" /EHsc /Tphype/graph_dataset.cpp /Fobuild\temp.win-amd64-3.6\Release\hype/graph_dataset.obj -std=c++11
cl: 命令行 warning D9002 :忽略未知选项“-std=c++11”
graph_dataset.cpp
c:\anaconda3\envs\poincare\lib\site-packages\numpy\core\include\numpy\npy_1_7_deprecated_api.h(14) : Warning Msg: Using deprecated NumPy API, disable it with #define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION
hype/graph_dataset.cpp(3033): warning C4244: “=”: 从“Py_ssize_t”转换到“int”，可能丢失数据
hype/graph_dataset.cpp(3418): warning C4244: “=”: 从“__pyx_t_5numpy_long_t”转换到“long”，可能丢失数据
hype/graph_dataset.cpp(3429): warning C4244: “=”: 从“__pyx_t_5numpy_long_t”转换到“long”，可能丢失数据
C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\bin\HostX86\x64\link.exe /nologo /INCREMENTAL:NO /LTCG /DLL /MANIFEST:EMBED,ID=2 /MANIFESTUAC:NO /LIBPATH:C:\Anaconda3\envs\poincare\libs /LIBPATH:C:\Anaconda3\envs\poincare\PCbuild\amd64 "/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\ATLMFC\lib\x64" "/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\lib\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\NETFXSDK\4.6.1\lib\um\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\10\lib\10.0.17763.0\ucrt\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\10\lib\10.0.17763.0\um\x64" /EXPORT:PyInit_graph_dataset build\temp.win-amd64-3.6\Release\hype/graph_dataset.obj /OUT:C:\Users\DELL\Github\poincare-embeddings-master\hype\graph_dataset.cp36-win_amd64.pyd /IMPLIB:build\temp.win-amd64-3.6\Release\hype\graph_dataset.cp36-win_amd64.lib
  正在创建库 build\temp.win-amd64-3.6\Release\hype\graph_dataset.cp36-win_amd64.lib 和对象 build\temp.win-amd64-3.6\Release\hype\graph_dataset.cp36-win_amd64.exp
正在生成代码
已完成代码的生成
building 'hype.adjacency_matrix_dataset' extension
C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\bin\HostX86\x64\cl.exe /c /nologo /Ox /W3 /GL /DNDEBUG /MD -IC:\Anaconda3\envs\poincare\lib\site-packages\numpy\core\include -IC:\Anaconda3\envs\poincare\include -IC:\Anaconda3\envs\poincare\include "-IC:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\ATLMFC\include" "-IC:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\include" "-IC:\Program Files (x86)\Windows Kits\NETFXSDK\4.6.1\include\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.17763.0\ucrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.17763.0\shared" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.17763.0\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.17763.0\winrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.17763.0\cppwinrt" /EHsc /Tphype/adjacency_matrix_dataset.cpp /Fobuild\temp.win-amd64-3.6\Release\hype/adjacency_matrix_dataset.obj -std=c++11
cl: 命令行 warning D9002 :忽略未知选项“-std=c++11”
adjacency_matrix_dataset.cpp
c:\anaconda3\envs\poincare\lib\site-packages\numpy\core\include\numpy\npy_1_7_deprecated_api.h(14) : Warning Msg: Using deprecated NumPy API, disable it with #define NPY_NO_DEPRECATED_API NPY_1_7_API_VERSION
hype/adjacency_matrix_dataset.cpp(3144): warning C4244: “=”: 从“Py_ssize_t”转换到“int”，可能丢失数据
hype/adjacency_matrix_dataset.cpp(5484): warning C4244: “=”: 从“Py_ssize_t”转换到“long”，可能丢失数据
hype/adjacency_matrix_dataset.cpp(5595): warning C4244: “=”: 从“Py_ssize_t”转换到“long”，可能丢失数据
C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\bin\HostX86\x64\link.exe /nologo /INCREMENTAL:NO /LTCG /DLL /MANIFEST:EMBED,ID=2 /MANIFESTUAC:NO /LIBPATH:C:\Anaconda3\envs\poincare\libs /LIBPATH:C:\Anaconda3\envs\poincare\PCbuild\amd64 "/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\ATLMFC\lib\x64" "/LIBPATH:C:\Program Files (x86)\Microsoft Visual Studio\2017\Community\VC\Tools\MSVC\14.16.27023\lib\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\NETFXSDK\4.6.1\lib\um\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\10\lib\10.0.17763.0\ucrt\x64" "/LIBPATH:C:\Program Files (x86)\Windows Kits\10\lib\10.0.17763.0\um\x64" /EXPORT:PyInit_adjacency_matrix_dataset build\temp.win-amd64-3.6\Release\hype/adjacency_matrix_dataset.obj /OUT:C:\Users\DELL\Github\poincare-embeddings-master\hype\adjacency_matrix_dataset.cp36-win_amd64.pyd /IMPLIB:build\temp.win-amd64-3.6\Release\hype\adjacency_matrix_dataset.cp36-win_amd64.lib
  正在创建库 build\temp.win-amd64-3.6\Release\hype\adjacency_matrix_dataset.cp36-win_amd64.lib 和对象 build\temp.win-amd64-3.6\Release\hype\adjacency_matrix_dataset.cp36-win_amd64.exp
正在生成代码
已完成代码的生成

Your suggestion are welcome. Thanks.

Hello there,

I saw this question and I am looking for something similar.

But, in my case, let's say I have a list like:

parent | child -- | -- animal | carnivore carnivore | dog dog | rottweiler dog | huskey dog | pug

From this list I can train a Poincare model and get poincare embeddings for every word/node in this hierarchy. Then, let's say I receive a new word: "dogo argentino". What I want at this point is a way to infer automatically that "dogo argentino" is the child of "dog". Can poincare embeddings be of help here? If ignoring poincare embeddings at all, a thought would be to get fastText embeddings for all the words, and find the word that is the closest (maybe in terms of cosine similarity) to the fastText embedding for "dogo argentino". Then assign the parent of that closest word, to the "dogo argentino" word. Any thoughts would be greatly appreciated. Thank you!

Going through the code instructions on Mac OSX gave the following error:

ModuleNotFoundError: No module named 'torch'

Which was resolved by modifying python3 to python.

Hi, I'm dealing with embedding of hierarchical structures (a tree) and I used your algorithm. It works, but when I visualise the 2D or 3D embedding the root of the tree is not in the center of the disk. Is there some command or flag to add to the code? Or is the root supposed to be in the center on his own?

Hi, I try to use this tool follow the readme file. But the latest version does not include hypernymysuite file. My system return me

"from hypernymysuite.base import HypernymySuiteModel ModuleNotFoundError: No module named 'hypernymysuite'"

Hi, I trained a 100 dimension model with train-nouns.sh in poincare manifold. I tested the model by calculating the tensor.norm of some nodes and poincare.distance between some nodes.

However, for Gauss, Brain, Blue and many other nodes near to the edge, the norm of them become 1.0000, which means infinity in poincare ball. In addition, if I calculate the poincare distance between these nodes near to the edge and other nodes, the distance will be always 12.2061, which is a little confusing.

Thanks for your attention!

Hi there, I have a quite large dataset and GPU usage that does not exceed 24 hours. In this case, I need to restore from checkpoint and continue to train the old models. The current situation is that, each time I try to do so, it is shown that the dimension of parameters is not correct--originally I train a 2-dimension embedding but now the dim of parameter is 21. Any thoughts on this? Any advice?

Hello, apologies in advance if this is a silly question. I was just looking at mammal_closure.csv because i want to do something similar with my own data and run like train-mammals.sh. I was wondering, what do the numbers in the file indicate? Such as vixen.n.02, mastiff.n.01 - what are the n.01 and n.02? Thanks!

Hello, I am Tianyu. I just tried to use my data to train a new model using embed.py. However, after I get the dl model, what is the next step for me to get the embeddings? I don't find clear steps for future work. Thanks a lot.

In the formula angle_at_u(...), dist is used squared in the numerator, and unsquared in the denominator. Thus, this does not expect distance(...) to return a squared norm but a true norm.

You can notice the inconsistency if you put norm(...) and distance(...) next to each other:

Hi,

I could be wrong, but I have the impression the entailment cones energy function is wrong, and optimizes the inverted order relationship between concepts.

If I understood correctly...:

• In our data files, id2 should be the generic concept, and id1 the more specific one.
• After training norm(embedding(id2)) should usually be lower than norm(embedding(id1)) as cones have more space on the edge, so generic concepts should be close to the origin and specific concepts should be far from it.

However, after training, I observe the reverse to be true:

• Generic concepts are as far as possible from the center.
• In their cones, one can find even-more-generic concepts, and no more specific concepts.

I'm either doing something very wrong, or have incorrect assumptions, or the energy function is reversed compared to what it should be according to the entailment cones article, no?