I'm trying to train a PassiveAggressiveClassifier using TfidVectorizer with partial_fit technique in the script below:
Code Updated:
a, ta = [], []
r, tr = [], []
g = []
vect = HashingVectorizer(ngram_range=(1,4))
model = PassiveAggressiveClassifier()
with open('files', 'rb') as f:
for line in f:
line = line.strip()
with open('gau-' + line + '.csv', 'rb') as csvfile:
reader = csv.DictReader(csvfile)
for row in reader:
res = row['gau']
g.append(res)
cls = np.unique(g)
print(len(cls))
with open('gau-' + line + '.csv', 'rb') as csvfile:
reader = csv.DictReader(csvfile)
i = 0
j = True
for row in reader:
arr = row['text']
res = row['gau']
a.append(arr)
if(len(res) > 0):
r.append(int(res))
i = i + 1
if i % 400 == 0:
training_set = vect.fit_transform(a)
print(training_set.shape)
training_result = np.array(r)
model = model.partial_fit(
training_set, training_result, classes=cls)
a, r, i = [], [], 0
print(model)
testing_set = vect.transform(ta)
testing_result = np.array(tr)
predicted = model.predict(testing_set)
print "Result to be predicted: "+testing_result
print "Prediction: "+predicted
There are multiple CSV files each containing 4k-5k records and I am trying to fit 400 records at a time using the partial_fit function. When I ran this code, I ran into the following error:
Result to be predicted: 1742
Prediction: 2617
How do I resolve this issue? The records in my CSV files are of variable length.
UPDATE:
Replacing TfidVectorizer with HashingVectorizer, I successfully created my model, but now while executing prediction on my test data the predictions generated were all incorrect.
My training data contains millions of lines of csv files and each line contains at most 4k-5k words of text.
So Is there any problem with my approach i.e. can these algorithms can be used with my data?
I'm trying to train PassiveAggressiveClassifier using TfidVectorizer with partial_fit technique with below script:
You can't, because TfidfVectorizer does not work for online learning. You want HashingVectorizer for that.
As for what exactly is going on in your code, the problem is here:
training_set = vect.fit_transform(a)
print(training_set.shape)
training_result = np.array(r)
model = model.partial_fit(training_set, training_result, classes=cls)
You are refitting your TF-IDF object at each step. So there is nothing stopping you from having a dictionary size at one iteration and another at the next iteration, which is exactly the error you are getting.
You can try a few things if you insist on using TF-IDF:
Append zeroes / trim the vector returned by fit_transform to make the length of the first one: very unlikely to work well;
Call fit on the TF-IDF object with an initial data set (preferably a large one) and then call transform on the others. This might work better, but I still suggest the HashingVectorizer.
This is what i understand from your problem.
1) You have a requirement to apply the partial fit model to do the online training.
2) Your feature space is so huge.
If I got it right then I faced the same problem. And if you will use the HashingVectorizer, there are high chances of key collision.
HashingVectorizer doc
There are also a couple of cons (vs using a CountVectorizer with an
in-memory vocabulary): there is no way to compute the inverse
transform (from feature indices to string feature names) which can be
a problem when trying to introspect which features are most important
to a model. there can be collisions: distinct tokens can be mapped to
the same feature index. However in practice this is rarely an issue if
n_features is large enough (e.g. 2 ** 18 for text classification
problems). no IDF weighting as this would render the transformer
stageful.
If key will collide then there are chances of reduction in accuracy.
In my online training, firstly i trained the classifier with partial_fit like this.
classifier = MultinomialNB(alpha=alpha_optimized).partial_fit(X_train_tfidf,y_train,classes=np.array([0,1]))
On second day i load the pickled classifier, count_vect and tfidf of first day training set. Then I only applied the transform on count_vet and tfidf. And it worked
X_train_counts = count_vect.transform(x_train)
X_train_tfidf = tfidf.transform(X_train_counts)
pf_classifier.partial_fit(X_train_tfidf,y_train)
In case of any doubt please reply.
As a solution I would say that using HashingVectorizer can fit your needs as you can set the number of features in the constructor.
You may prefer to use TfidfVectorizer and maybe it is more suitable for your situation. I let the answer until someone give something that use a more useful for you.
Hope there will be. Don't forget to accept the one you choose
For those who HashingVectorizer doesn't meet their needs, see a possible alternative in my answer to this related question here. It's basically a custom implementation of partial_fit for TfidfVectorizer and CountVectorizer.
Two comments relating to the specific discussion here:
OP's issue requires that the dimension of the output vector be identical after every call of partial_fit. In general it is expected that every Scikit-Learn estimator that implements partial_fit be able to work within a pipeline after that partial_fit is called, so for vectorizers this means not changing the output dimension since other estimators in the pipeline may not necessarily be able to handle the change. I think this is why partial_fit has not yet been implemented in Scikit-Learn for these vectorizers (see discussion on an active PR), since partial_fit will presumably update the vocabulary which will definitely change the output dimension.
So the solution proposed by my answer (a partial_fit method for TfidfVectorizer) would only solve the first part of OP's needs which is incremental learning. To solve the second part it may be possible to pad the output sequence with zeros into a predetermined vector. It's not ideal, since it would fail when the vocabulary exceeds that limit.
