Re-Implement file parser to also allow loading from in-memory buffers

[seijikun]

The current libsvm only allows loading trained svm-models from a file with the given name. Especially for use in mobile applications, parsing a model from a buffer would be much appreciated.

The PR consists of two commits:

• Commit 1: Implement the parsing using C++ streams, with an abstract source. This should allow to fairly easily implement parsing from arbitrary sources, such as network sockets, files, buffers, or maybe even a on-the-fly unpacked gzip stream. Add a new public svm_parse_model_from_buffer which uses the new parser to parse svm-models from a given buffer.
• Commit 2: Refactor svm_load_model to use the new parser.

Correctness of the parser was tested with the following code - which simply compared the same model parsed from a buffer and a file (in between both commits):

int main() {
    struct svm_model* svmModelFile = svm_load_model("/tmp/cache/activity.model");
    struct svm_model* svmModel = svm_parse_model_from_buffer(MODEL.data(), MODEL.size());
    assert((svmModel == nullptr && svmModelFile == nullptr) || (svmModel != nullptr && svmModelFile != nullptr));
    if(svmModel == nullptr) { return 1; }

    assert(svmModel->l == svmModelFile->l);
    assert(svmModel->nr_class == svmModelFile->nr_class);
    assert(memcmp(&svmModel->param, &svmModelFile->param, sizeof(svm_parameter)) == 0);
    assert(svmModel->free_sv == svmModelFile->free_sv);

    size_t nr_class_permut = svmModel->nr_class * (svmModel->nr_class - 1) / 2;
    for(size_t i = 0; i < nr_class_permut; ++i) {
        assert(svmModel->probA[i] == svmModelFile->probA[i]);
        assert(svmModel->probB[i] == svmModelFile->probB[i]);
        assert(svmModel->rho[i] == svmModelFile->rho[i]);
    }
    for(size_t i = 0; i < svmModel->nr_class; ++i) {
        assert(svmModel->label[i] == svmModelFile->label[i]);
        assert(svmModel->nSV[i] == svmModelFile->nSV[i]);
    }

    for(size_t c = 0; c < svmModel->nr_class - 1; ++c) {
        for(size_t i = 0; i < svmModel->l; ++i) {
            assert(svmModel->sv_coef[c][i] == svmModelFile->sv_coef[c][i]);
        }
    }

    for(size_t i = 0; i < svmModel->l; ++i) {
        struct svm_node* supportVector = svmModelFile->SV[i];
        size_t j = 0;
        for(j; supportVector[j].index != -1; ++j) {
            assert(svmModel->SV[i][j].index == supportVector[j].index);
            assert(svmModel->SV[i][j].value == supportVector[j].value);
        }
        assert(svmModel->SV[i][j].index == supportVector[j].index);
    }

    svm_free_and_destroy_model(&svmModel);
    svm_free_and_destroy_model(&svmModelFile);
}

[KyleLin123456]

@seijikun Hi, I would like to ask about the change in https://github.com/cjlin1/libsvm/pull/170/commits/adbbfe236387f14080b4012775ab271fa2aa57c9 Why would you unset the the parameter "stderr" in line 292? Is there any meaningful benefit? Thank you so much for the answer.

[cjlin1]

We take the grid.py change for accepting file names with spaces. Other changes are not taken.