'ConcatenatedDoc2Vec' object has no attribute 'corpus_count'

[jasminezz]

Description

study "doc2vec-IMDB.ipynb" step by step ,but when train the model, got the following error:

100000 docs: 25000 train-sentiment, 25000 test-sentiment
Doc2Vec(dm/c,d100,n5,w5,mc2,s0.001,t8)
Doc2Vec(dbow,d100,n5,mc2,s0.001,t8)
Doc2Vec(dm/m,d100,n5,w10,mc2,s0.001,t8)
START 2017-05-22 17:11:40.993000
*0.248200 : 1 passes : Doc2Vec(dm/c,d100,n5,w5,mc2,s0.001,t8) 257.3s 1.1s
*0.240400 : 1 passes : Doc2Vec(dm/c,d100,n5,w5,mc2,s0.001,t8)_inferred 257.3s 5.7s
*0.113640 : 1 passes : Doc2Vec(dbow,d100,n5,mc2,s0.001,t8) 54.2s 0.8s
*0.101200 : 1 passes : Doc2Vec(dbow,d100,n5,mc2,s0.001,t8)_inferred 54.2s 2.6s
*0.176480 : 1 passes : Doc2Vec(dm/m,d100,n5,w10,mc2,s0.001,t8) 75.1s 0.8s
Traceback (most recent call last):
File "D:/PycharmProjects/NLP_demo/doc2vec_train_test.py", line 142, in
train(doc_list, models_by_name, train_docs, test_docs)
File "D:/PycharmProjects/NLP_demo/doc2vec_train_test.py", line 105, in train
train_model.train(doc_list, total_examples=train_model.corpus_count, epochs=train_model.iter)
AttributeError: 'ConcatenatedDoc2Vec' object has no attribute 'corpus_count'
*0.194800 : 1 passes : Doc2Vec(dm/m,d100,n5,w10,mc2,s0.001,t8)_inferred 75.1s 3.3s

Environment:
python3
windows 7

[menshikh-iv]

Thank you @jasminezz, I reproduce this bug
@robotcator could you please fix this bug in this notebook?

[robotcator]

@menshikh-iv okay, thank you. And I will fix it.

[jasminezz]

@robotcator If you fixed it ,please let me know, thank you

[ltn614]

I came across the same bug, python 3.5 on ubuntu 16.04

[ltn614]

@robotcator , hope you fix it ASAP

[robotcator]

@jasminezz Sorry for the late reply, cause I have some other things to do these days. Can you post your doc2vec_train_test.py file and tell me your version of gensim?

[robotcator]

@jasminezz Hi, the old version of word2vec train method did not provide the parameters of count and epochs. But the new version of word2vec force the user to calls the train() using explicit count, epochs(5e9529b). And in fact the 'ConcatenatedDoc2Vec' object didn't have attribute 'corpus_count'(refer to line 394, https://github.com/RaRe-Technologies/gensim/blob/develop/gensim/test/test_doc2vec.py ), and the train method didn't do anything in fact( you should train subcomponents individually). So if you want to call the method train() of ConcatenatedDoc2Vec, you don't have to provide the parameter of count and epochs.

[robotcator]

@menshikh-iv is there need to do modify the class ConcatenatedDoc2Vec's train to provide these parameter to be compatible with the new version word2vec train method (https://github.com/RaRe-Technologies/gensim/blob/develop/gensim/test/test_doc2vec.py)

[menshikh-iv]

@robotcator yes

[flowingrain]

When I tried to train without total_examples and epochs, ipython remind me to specify either total_examples or total_words; when I tried to train with total_examples, it remind me that train() got an unexpected keyword argument 'total_examples'. Is it a inner contradiction?

[gojomo]

@flowingrain Can you be more specific about what class(es) you're using? (Until this issue is fixed in a released gensim, ConcatenatedDoc2Vec will object to these new arguments in its no-op train() method – it's really just a hack to evaluate combined models. But Word2Vec or Doc2Vec train() will in recent versions require them.)

The fact that ConcatenatedDoc2Vec can't (until fixed) be used in the style of the demo notebook is the open bug here. But you can just avoid calling the no-op train() on it as a workaround.

[flowingrain]

@gojomo Oops~
I‘m using ConcatentedDoc2Vec as in the tutorial.
Is simple model not concatenated available?
Let me have a try