timvieira / crf Goto Github PK

Traceback (most recent call last):
File "crf/example.py", line 144, in
main()
File "crf/example.py", line 138, in main
run(proportion=[0.7, 0.3], load='model.pkl~')
File "crf/example.py", line 104, in run
validate(crf)
File "crf/example.py", line 93, in validate
llh = sum(map(crf.likelihood, iterview(train, msg='llh'))) / len(train)
File "/crf/crf/crf.py", line 100, in likelihood
(g0, g) = self.log_potentials(x)
File "/crf/crf/crf.py", line 58, in log_potentials
g0[y] = W[f[0,None,y]].sum()
TypeError: 'NoneType' object has no attribute 'getitem'

i ran this on repl.it after installing arsenal and crf package successfully.

Matplotlib created a temporary config/cache directory at /tmp/matplotlib-ykbg_s95 because the default path (/config/matplotlib) is not a writable directory; it is highly recommended to set the MPLCONFIGDIR environment variable to a writable directory, in particular to speed up the import of Matplotlib and to better support multiprocessing.
Traceback (most recent call last):
  File "crf/example.py", line 160, in <module>
    main()
  File "crf/example.py", line 154, in main
    run_test()
  File "crf/example.py", line 126, in run_test
    (L, A) = build_domain(train)
  File "/home/runner/SorrowfulKnowledgeableOrigin/crf/crf/stringcrf.py", line 20, in build_domain
    A.stop_growth()
AttributeError: 'Alphabet' object has no attribute 'stop_growth'```

Hi!

Thanks for sharing this code. I wanted to know if it is possible to use this code to clean classifier predictions where each pixel gets a probability value of a particular class. If so, is there an example that I can look at?

Thanks!

hi, whether the basecrf.py is suitable for general crf model or the aspect of NLP?

Really apperciate your work, i have investigated your code but have some confusion part.
in the code

    def sgd(self, data, iterations=20, a0=10, validate=None):
        """ Parameter estimation with stochastic gradient descent (sgd). """
        self.preprocess(data)
        W = self.W
        for i in range(iterations):
            rate = a0 / (i + 1)**0.501
            for x in iterview(data, msg='Iteration %s' % i):
                for k, v in self.expectation(x).items():
                    W[k] -= rate*v
                for k in x.target_features:
                    W[k] += rate
            if validate:
                validate(self, i)

• Is it a parameter restore or fix for the warm-up stage? for the result the gradient at warm-up stage may fluctuate violently, so restore parameter by add some scales can mitigate this.
• But why not fix the gradient(here is the expectation logit) directly, is there some theory analysis about these two method?

Hi!

Thanks for this code. What's the License for it? Would you consider an MIT license or a compatible?

Hi, I'm confused about the feature function. In your code, is feature function defined in the method FeatureVectorSequence?

ubgpu@ubgpu:/github/crf$ python example.py run 100
Traceback (most recent call last):
File "example.py", line 138, in
main(proportion=[0.7, 0.3], iterations=args.interations)
AttributeError: 'Namespace' object has no attribute 'interations'
ubgpu@ubgpu:/github/crf$ echo $PYTHONPATH
/home/ubgpu/github/arsenal:/usr/local/lib/python2.7/dist-packages
ubgpu@ubgpu:~/github/crf$

I think the tagged_reference.txt is the math reference data from https://people.cs.umass.edu/~mccallum/data/cora-refs.tar.gz, i am wondering whether the data need process or the script can run with raw data?

Hi Tim,

I was using your code as a base for a course homework I'm TAing.

I was looking at your forward code and I was wondering if you had a rationale for how you use the logexpsum.

Looking at Sutton & McCallum [1] and taking alpha as forward recursion variable, it seems that each forward recursion is:

alpha(thisstate) = sum_laststate exp(transition_potential)*alpha(laststate)

You are adding the alpha(laststate) to the transition potential, exponentiating, summing, and logging (and some interesting max operator thing going on).

Can you really combine the alpha and transition in the exponential? I don't see how you push an exponential function back to the outside in each alpha to allow for such an operation.

Thanks!

[1] Sutton & McCallum. Page 15. Section 1.3.3. Eq. 1.37. http://people.cs.umass.edu/~mccallum/papers/crf-tutorial.pdf

(side note, I believe we met at MASC this year)

Best,
Brian

hello, i know that hmm can predict the label of one observed sequence such as predict the word of one speech.can crf do the same thing ...the steps are like:
1.first train the different crf models with diffenrent set of samples
2.input the sequence that want predicting to different models and get the probabilities of diffenrent models
3.find the model that has the maximum probability,and the label the sequence with the model's label.

thanks a lot!

Hey,pal ! I wanna program the crf in python,but I still diidn't understand it and being puzzled. Could you send me some tutorials of CRF or tell me how to learn it please? Any guidance would be greatly appreciated! Thanks in advance!

Hi, After we done some experiment, we found that, In stringcrf2.py the size of W (crf feature function parameter matrix) should be (A+L)_L instead of A_L. After fixing size issue and a related line of code, we can achieve higher llh on same dataset and also monotonically increase after iteration. However, it could be highly possible that we are wrong, so please check that. We are looking forward to hear from you, thanks