ModuleNotFoundError after calling pythainlp.sentiment(string, 'ulmfit')

[LXZE]

Describe the bug
After the installation of Pythainlp version 1.7 inside Google Colab, I have called a function pythainlp.sentiment(string, 'ulmfit') and then an error occurred as below example.

To Reproduce
Steps to reproduce the behaviour:

1. Create new Google colab's notebook, setting notebook to use Python3 and Hardware accelerator as GPU
2. Code as below

!pip install pythainlp
!pip install torchvision
!pip install https://github.com/fastai/fastai/archive/1.0.22.zip
import pythainlp
test_string = 'ประโยคทดสอบ'
print(pythainlp.sentiment(test_string, 'ulmfit'))

Note
After I've tried to install torchvision and fastai in the latest version and v1.0.22 (according to this issue).
The fastai library still causes an error ModuleNotFoundError: No module named 'fastai.lm_rnn' So, I raise this issue.

Expected behaviour
According to pythainlp 1.7 document on sentiment analysis,
This method should return a prediction result from a string and not raise an error after calling sentiment(string, 'ulmfit) method without any additional code that importing all of the required libraries.

Example
Here is a google colab's notebook link

Desktop

• OS: macOS High Sierra 10.13.4
• Python Version: 3.6 (Google Colab)
• Version: pythainlp=1.7

Suggestion
As this issue might be about Fastai's problem, I suggest that this method should automatically install a torch library and the specific version of Fastai, or make a model become compatible with the latest version.

[wannaphong]

Please use fastai 0.7 (PyThaiNLP 1.7). We updata to Fastai 1.0 in the next release.
We will remove sentiment in the next release (PyThaiNLP 2.0).

[LXZE]

I've tried fastai library version 0.7 with the same code and this error occurred

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-2-3f17b1579a73> in <module>()
      3 test_string = 'ประโยคทดสอบ'
      4 
----> 5 print(pythainlp.sentiment(test_string, 'ulmfit'))

/usr/local/lib/python3.6/dist-packages/pythainlp/sentiment/__init__.py in sentiment(text, engine)
     36                 return classifier.classify(featurized_test_sentence)
     37         elif engine=='ulmfit':
---> 38                 from pythainlp.sentiment import ulmfit_sent
     39                 tag=ulmfit_sent.get_sentiment(text)
     40                 sa=""
...
...
...
/usr/local/lib/python3.6/dist-packages/torchtext/data/field.py in Field()
    116     # numeric type.
    117     dtypes = {
--> 118         torch.float32: float,
    119         torch.float: float,
    120         torch.float64: float,

AttributeError: module 'torch' has no attribute 'float32'

Ps. So, are you going to remove the sentiment analysis function on next release and separate this task to other library or something?

[wannaphong]

@LXZE Fastai 0.7 need PyTorch 0.4.1.
You can train the sentiment analysis from https://github.com/PyThaiNLP/pythainlp/blob/dev/notebooks/ulmfit_sentiment_example.ipynb in the next release (PyThaiNLP 2.0). Because we don't have enough contributors to maintain this function.

[AlgoBeach]

I think NLP for Thai language is an opportunity. Would like to know better what this guy is doing. Do you think you could talk to him? I will coach you before if you like. Phil From: Wannaphong Phatthiyaphaibun [mailto:notifications@github.com] Sent: Thursday, January 24, 2019 8:35 AM To: PyThaiNLP/pythainlp Cc: Subscribed Subject: Re: [PyThaiNLP/pythainlp] ModuleNotFoundError after calling pythainlp.sentiment(string, 'ulmfit') (#172) @LXZE<https://github.com/LXZE> Fastai 0.7 need PyTorch 0.4.1. You can train the sentiment analysis from https://github.com/PyThaiNLP/pythainlp/blob/dev/notebooks/ulmfit_sentiment_example.ipynb in the next release (PyThaiNLP 2.0). — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub<#172 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AdzaxmShvErfoTBN2bOA3tGGpDGF2OQtks5vGeCogaJpZM4aNFhJ>.

[cstorm125]

@LXZE Seems like there is a bug with the versions of pre-1.0 fastai. I've tried combing through the commits but still no luck. We are removing sentiment analysis because

1. It is redundant with text classification.
2. We trained the sentiment on product reviews which can hardly generalize to other types of sentiment. It would be better if you just train on your own domain-dependent datasets. We are also working on releasing a few diverse training datasets for this purpose.

[LXZE]

Understood. I would train the model according to the given example on my own later. Closing this issue now.