Important: This notebook will only work with fastai-0.7.x. Do not try to run any fastai-1.x code from this path in the repository because it will load fastai-0.7.x



In [ ]:

    
%reload_ext autoreload
%autoreload 2
%matplotlib inline

from fastai.nlp import *
from sklearn.linear_model import LogisticRegression
from sklearn.svm import LinearSVC
from torchtext import vocab, data, datasets

IMBD dataset and the sentiment classification task

The large movie view dataset contains a collection of 50,000 reviews from IMDB. The dataset contains an even number of positive and negative reviews. The authors considered only highly polarized reviews. A negative review has a score ≤ 4 out of 10, and a positive review has a score ≥ 7 out of 10. Neutral reviews are not included in the dataset. The dataset is divided into training and test sets. The training set is the same 25,000 labeled reviews.

The sentiment classification task consists of predicting the polarity (positive or negative) of a given text.

To get the dataset, in your terminal run the following commands:

wget http://ai.stanford.edu/~amaas/data/sentiment/aclImdb_v1.tar.gz

gunzip aclImdb_v1.tar.gz

tar -xvf aclImdb_v1.tar

Tokenizing



In [ ]:

    
sl=1000
vocab_size=200000



In [ ]:

    
PATH='data/aclImdb/'

names = ['neg','pos']
trn,trn_y = texts_labels_from_folders(f'{PATH}train',names)
val,val_y = texts_labels_from_folders(f'{PATH}test',names)

Here is the text of the first review:



In [ ]:

    
trn[0]









    Out[ ]:





"Story of a man who has unnatural feelings for a pig. Starts out with a opening scene that is a terrific example of absurd comedy. A formal orchestra audience is turned into an insane, violent mob by the crazy chantings of it's singers. Unfortunately it stays absurd the WHOLE time with no general narrative eventually making it just too off putting. Even those from the era should be turned off. The cryptic dialogue would make Shakespeare seem easy to a third grader. On a technical level it's better than you might think with some good cinematography by future great Vilmos Zsigmond. Future stars Sally Kirkland and Frederic Forrest can be seen briefly."



In [ ]:

    
trn_y[0]









    Out[ ]:





0

CountVectorizer converts a collection of text documents to a matrix of token counts (part of sklearn.feature_extraction.text). Here is how you specify parameters to the CountVectorizer. We will be working with the top 200000 unigrams, bigrams and trigrams.



In [ ]:

    
veczr = CountVectorizer(ngram_range=(1,3), tokenizer=tokenize, max_features=vocab_size)

In the next line fit_transform(trn) computes the vocabulary and other hyparameters learned from the training set. It also transforms the training set. Since we have to apply the same transformation to your validation set, the second line uses just the method transform(val). trn_term_doc and val_term_doc are sparse matrices. trn_term_doc[i] represents training document $i$ and it is binary (it has a $1$ for each vocabulary n-gram present in document $i$ and $0$ otherwise).



In [ ]:

    
trn_term_doc = veczr.fit_transform(trn)
val_term_doc = veczr.transform(val)



In [ ]:

    
trn_term_doc.shape









    Out[ ]:





(25000, 200000)



In [ ]:

    
veczr.get_params()









    Out[ ]:





{'analyzer': 'word',
 'binary': False,
 'decode_error': 'strict',
 'dtype': numpy.int64,
 'encoding': 'utf-8',
 'input': 'content',
 'lowercase': True,
 'max_df': 1.0,
 'max_features': 200000,
 'min_df': 1,
 'ngram_range': (1, 3),
 'preprocessor': None,
 'stop_words': None,
 'strip_accents': None,
 'token_pattern': '(?u)\\b\\w\\w+\\b',
 'tokenizer': <function fastai.nlp.tokenize>,
 'vocabulary': None}



In [ ]:

    
# here is the vocabulary
vocab = veczr.get_feature_names()



In [ ]:

    
vocab[50:55]









    Out[ ]:





['! " and', '! " as', '! " at', '! " but', '! " for']

Weighted Naive Bayes

Our first model is a version of logistic regression with Naive Bayes features described here. For every document we compute binarized features as described above. Each feature if multiplied by a log-count ratio (see below for explanation). A logitic regression model is then trained to predict sentiment.

Here is how to define log-count ratio for a feature $f$:

$\text{log-count ratio} = \log \frac{\text{ratio of feature $f$ in positive documents}}{\text{ratio of feature $f$ in negative documents}}$

where ratio of feature $f$ in positive documents is the number of times a positive document has a feature divided by the number of positive documents.



In [ ]:

    
# Here is how we get a model from a bag of words
md = TextClassifierData.from_bow(trn_term_doc, trn_y, val_term_doc, val_y, sl)



In [ ]:

    
learner = md.dotprod_nb_learner()
learner.fit(0.02, 1, wds=1e-5, cycle_len=1)









    



    Warning: no model found for 'en'

    Only loading the 'en' tokenizer.


    Warning: no model found for 'en'

    Only loading the 'en' tokenizer.







    





 
 










    



[ 0.      0.068   0.1223  0.9165]



In [ ]:

    
learner = md.dotprod_nb_learner()
learner.fit(0.02, 1, wds=1e-6)









    





 
 










    



[ 0.      0.0581  0.1073  0.9235]

unigram

Here is use CountVectorizer with a different set of parameters. In particular ngram_range by default is set to (1, 1)so we will get unigram features. Note that we are specifiying our own tokenize function.



In [ ]:

    
veczr =  CountVectorizer(tokenizer=tokenize)
trn_term_doc = veczr.fit_transform(trn)
val_term_doc = veczr.transform(val)

Here is how to compute the $\text{log-count ratio}$ r.



In [ ]:

    
x=trn_term_doc
y=trn_y

p = x[y==1].sum(0)+1
q = x[y==0].sum(0)+1
r = np.log((p/p.sum())/(q/q.sum()))
b = np.log(len(p)/len(q))

Here is the formula for Naive Bayes.



In [ ]:

    
pre_preds = val_term_doc @ r.T + b
preds = pre_preds.T>0
(preds==val_y).mean()









    Out[ ]:





0.80740000000000001



In [ ]:

    
pre_preds = val_term_doc.sign() @ r.T + b
preds = pre_preds.T>0
(preds==val_y).mean()









    Out[ ]:





0.82623999999999997

Here is how we can fit regularized logistic regression where the features are the unigrams.



In [ ]:

    
m = LogisticRegression(C=0.1, fit_intercept=False, dual=True)
m.fit(x, y)
preds = m.predict(val_term_doc)
(preds==val_y).mean()









    Out[ ]:





0.88260000000000005

bigram with NB features

Similar to the model before but with bigram features.



In [ ]:

    
veczr =  CountVectorizer(ngram_range=(1,2), tokenizer=tokenize)
trn_term_doc = veczr.fit_transform(trn)
val_term_doc = veczr.transform(val)



In [ ]:

    
y=trn_y
x=trn_term_doc.sign()
val_x = val_term_doc.sign()
p = x[y==1].sum(0)+1
q = x[y==0].sum(0)+1
r = np.log((p/p.sum())/(q/q.sum()))
b = np.log(len(p)/len(q))

Here we fit regularized logistic regression where the features are the bigrams. Bigrams are giving us 2% boost.



In [ ]:

    
m = LogisticRegression(C=0.1, fit_intercept=False)
m.fit(x, y);

preds = m.predict(val_x)
(preds.T==val_y).mean()









    Out[ ]:





0.90271999999999997

Here is the $\text{log-count ratio}$ r.



In [ ]:

    
r









    Out[ ]:





matrix([[ 0.6863,  0.6863, -0.7   , ...,  0.6863, -0.7   , -0.7   ]])

Here we fit regularized logistic regression where the features are the bigrams multiplied by the $\text{log-count ratio}$. We are getting an extra boost for the normalization.



In [ ]:

    
x_nb = x.multiply(r)
m = LogisticRegression(dual=True, C=1, fit_intercept=False)
m.fit(x_nb, y);



In [ ]:

    
w = m.coef_.T
preds = (val_x_nb @ w + m.intercept_)>0
(preds.T==val_y).mean()









    Out[ ]:





0.91479999999999995

This is an interpolation between Naive Bayes the regulaized logistic regression approach.



In [ ]:

    
beta=0.25

val_x_nb = val_x.multiply(r)
w = (1-beta)*m.coef_.mean() + beta*m.coef_.T
preds = (val_x_nb @ w + m.intercept_)>0
(preds.T==val_y).mean()









    Out[ ]:





0.91639999999999999



In [ ]:

    
w2 = w.T[0]*r.A1



In [ ]:

    
preds = (val_x @ w2 + m.intercept_)>0
(preds.T==val_y).mean()









    Out[ ]:





0.91479999999999995

References

Baselines and Bigrams: Simple, Good Sentiment and Topic Classification. Sida Wang and Christopher D. Manning pdf

Unused helpers



In [ ]:

    
class EzLSTM(nn.LSTM):
    def __init__(self, input_size, hidden_size, *args, **kwargs):
        super().__init__(input_size, hidden_size, *args, **kwargs)
        self.num_dirs = 2 if self.bidirectional else 1
        self.input_size = input_size
        self.hidden_size = hidden_size
        
    def forward(self, x):
        h0 = c0 = Variable(torch.zeros(self.num_dirs,x.size(1),self.hidden_size)).cuda()
        outp,_ = super().forward(x, (h0,c0))
        return outp[-1]



In [ ]:

    
def init_wgts(m, last_l=-2):
    c = list(m.children())
    for l in c:
        if isinstance(l, nn.Embedding): 
            l.weight.data.uniform_(-0.05,0.05)
        elif isinstance(l, (nn.Linear, nn.Conv1d)):
            xavier_uniform(l.weight.data, gain=calculate_gain('relu'))
            l.bias.data.zero_()
    xavier_uniform(c[last_l].weight.data, gain=calculate_gain('linear'));

class SeqSize(nn.Sequential):
    def forward(self, x):
        for l in self.children():
            x = l(x)
            print(x.size())
        return x

End



In [ ]: