Update/dependencies

[rssdev10]

Update dependencies. Reduce version requirements.

• accompanying code changes

[rssdev10]

Hi, I'm trying to make TextModels compilable again. Result of testing:

Test Summary:                  | Pass  Error  Total   Time
All tests                      |   93      3     96  23.4s
  📂 crf.jl                    |           3      3   3.4s
    crf                        |           3      3   3.3s
      Loss function            |           1      1   1.0s
      Viterbi Decode           |           1      1   0.7s
      CRF with Flux Layers     |           1      1   1.6s
  📂 ner.jl                    |   13            13   5.0s
    NER                        |   13            13   5.0s
  📂 pos.jl                    |   14            14   1.0s
    POS                        |   14            14   1.0s
  📂 sentiment.jl              |    6             6   0.6s
  📂 averagePerceptronTagger.jl |   13            13   7.4s
    Average Perceptron Tagger  |   13            13   7.4s
  📂 ulmfit.jl                 |   47            47   5.8s
    Custom layers              |   38            38   3.0s
    Language model             |    7             7   1.6s
    Text Classifier            |    2             2   0.8s
ERROR: LoadError: Some tests did not pass: 93 passed, 0 failed, 3 errored, 0 broken.

ulmfit was updated for Flux >=0.13 and, based on the unit testing this works fine.

The biggest issue is crf. I'm not sure how that was working before looking in to the unit test. The problem is https://github.com/JuliaText/TextModels.jl/blob/master/test/crf.jl#L7 :

        input_seq = [rand(4) for i in 1:3]
        c = CRF(2)

        scores = []
        push!(scores, score_sequence(c, input_seq, [onehot(1, 1:2), onehot(1, 1:2), onehot(1, 1:2)]))
        #...
        
        init_α = fill(-10000, (c.n + 2, 1))
        init_α[c.n + 1] = 0

        s1 = sum(exp.(scores))
        s2 = exp(forward_score(c, input_seq, init_α))

        @test (s1 - s2) / max(s1, s2) <= 0.00000001

score_sequence cannot be performed because of input_seq contains the array of 4 elements but each of onehot(1, 1:2) generates just 2 labels. So, onecold() inside score_sequence throws mismatch of size.

From other side,

        init_α = fill(-10000, (c.n + 2, 1))
        init_α[c.n + 1] = 0

        s2 = exp(forward_score(c, input_seq, init_α))

this code speaks about expected 4 elements in input_seq = [rand(4) for i in 1:3] but not 2 as mentioned number of labels in CRF(2) and onehot(1, 1:2)...

@AdarshKumar712 any suggestions about this?