Hi, I am having trouble in training a QuaNet quantifier for multiclass (20) data. Everything works fine with where my dataset only has 2 classes. It looks like the ACC quantifier is not able to aggregate from more than 2 classes?

The classifier is built and trained as with the code below

classifier = LSTMnet(dataset.vocabulary_size, dataset.n_classes)
learner = NeuralClassifierTrainer(classifier)
learner.fit(*dataset.training.Xy)

where it has all the default configurations

{'embedding_size': 100, 'hidden_size': 256, 'repr_size': 100, 'lstm_class_nlayers': 1, 'drop_p': 0.5}

Then I tried to train QuaNet with following code

model = QuaNetTrainer(learner, qp.environ['SAMPLE_SIZE'])
model.fit(dataset.training, fit_learner=False)

and it showed that QuaNet is built as

QuaNetModule( (lstm): LSTM(120, 64, batch_first=True, dropout=0.5, bidirectional=True) (dropout): Dropout(p=0.5, inplace=False) (ff_layers): ModuleList( (0): Linear(in_features=208, out_features=1024, bias=True) (1): Linear(in_features=1024, out_features=512, bias=True) ) (output): Linear(in_features=512, out_features=20, bias=True) )

And then the error occured in model.fit().

Attached is the error I get.

Traceback (most recent call last): File "quanet-test.py", line 181, in model.fit(dataset.training, fit_learner=False) File "/home/vickys/.local/lib/python3.6/site-packages/quapy/method/neural.py", line 126, in fit self.epoch(train_data_embed, train_posteriors, self.tr_iter, epoch_i, early_stop, train=True) File "/home/vickys/.local/lib/python3.6/site-packages/quapy/method/neural.py", line 182, in epoch quant_estims = self.get_aggregative_estims(sample_posteriors) File "/home/vickys/.local/lib/python3.6/site-packages/quapy/method/neural.py", line 145, in get_aggregative_estims prevs_estim.extend(quantifier.aggregate(predictions)) File "/home/vickys/.local/lib/python3.6/site-packages/quapy/method/aggregative.py", line 238, in aggregate return ACC.solve_adjustment(self.Pte_cond_estim_, prevs_estim) File "/home/vickys/.local/lib/python3.6/site-packages/quapy/method/aggregative.py", line 246, in solve_adjustment adjusted_prevs = np.linalg.solve(A, B) File "<array_function internals>", line 6, in solve File "/usr/local/lib64/python3.6/site-packages/numpy/linalg/linalg.py", line 394, in solve r = gufunc(a, b, signature=signature, extobj=extobj) ValueError: solve1: Input operand 1 has a mismatch in its core dimension 0, with gufunc signature (m,m),(m)->(m) (size 2 is different from 20)

Thank you!

The parameter fit_leaner is not used in the function:

def fit(self, data: LabelledCollection, fit_learner=True):

and the learner is fitted every time:

self.learner.fit(*classifier_data.Xy)

In the last part of the Methods wiki page, where it says:

from classification.neural import NeuralClassifierTrainer, CNNnet

I think it should say:

from quapy.classification.neural import NeuralClassifierTrainer, LSTMnet

I think there is the function init_hidden:

def init_hidden(self, set_size):
        opt = self.hyperparams
        var_hidden = torch.zeros(opt['lstm_nlayers'], set_size, opt['lstm_hidden_size'])
        var_cell = torch.zeros(opt['lstm_nlayers'], set_size, opt['lstm_hidden_size'])
        if next(self.lstm.parameters()).is_cuda:
            var_hidden, var_cell = var_hidden.cuda(), var_cell.cuda()
        return var_hidden, var_cell

Where it says opt['lstm_hidden_size'] should be opt['hidden_size']

This transformation function is applied to each intermediate estimate.

Why should someone want to transform the prior between two iterations? A transformation of the prior is a heuristic, yet effective way of promoting desired properties of the solution. For instance,

• small values could be enhanced if the data is extremely imbalanced
• small values could be reduced if the user is looking for a sparse solution
• neighboring values could be averaged if the user is looking for a smooth solution
• the function could also leave the prior unaltered and just be used as a callback for logging the progress of the method

I hope this feature is useful. Let me know what you think!

I found another problem with a parameter. When using LSTMnet with QuaNet two parameters overlap (lstm_nlayers). I have renamed the one in the LSTMnet to lstm_class_nlayers.

The code seems to be tied up to using only 'cuda', which by default uses the first gpu in the system ('cuda:0'). It would be handy to be able to tell the library in which cuda gpu you want to train (cuda:0, cuda:1, etc).