A question about only_new

[thu-wangz17]

Hi,I have a question about the keyward only_new in teach method.In the document,it says that to make sure that you train only on newly queried labels, pass only_new=True to the .teach() method of the learner.Then if I set only_new=False with Pytorch models in modAL workflows,the models will be retrained with all the labelled data or continue learning with all the labelled data?If the workflow is the last one,it seems that the model will overfit on the repetitive data point,which makes me confused. Thank you.

[cosmic-cortex]

Hi!

only_new=True will train the model on the data which is passed to it in that function call.

Traditionally, when active learning was used in combination with more classical methods like random forest, SVM and others (essentialy anything which is not a neural network), the entire model was retrained on the old data + new data after the query. Since neural networks train incrementally because of the batch gradient descent, the traditional method is not particularly the best for them. (It would be very costly also.)

So, when you set only_new=True, you basically perform a training step on a single batch.

[thu-wangz17]

Thank you for the reply.Actually,since I set train_split=CVSplit(0.2) in NeuralNet,but with active learning,the model is only teached with 1 data point,then the validation set will be empty if only_new=True.However when I set only_new=False,the code runs.

import torch
from torch import nn
from modAL.models import ActiveLearner
from skorch import NeuralNet
from skorch.dataset import CVSplit
from modAL.uncertainty import uncertainty_sampling

class net(nn.Module):
    def __init__(self):
        super(net, self).__init__()
        self.model = nn.Sequential(
            nn.Linear(30, 256), 
            nn.ReLU(), 
            nn.Linear(256, 1), 
        )

    def forward(self, x):
        out = self.model(x)
        return out

device = 'cuda' if torch.cuda.is_available() else 'cpu'

model = NeuralNet(
    module=net, 
    criterion=nn.MSELoss,
    optimizer=torch.optim.Adam, 
    lr=3e-4, 
    max_epochs=10, 
    batch_size=32, 
    train_split=CVSplit(0.2), 
    verbose=1, 
    device=device
)

learner = ActiveLearner(
    estimator=model, 
    query_strategy=uncertainty_sampling,
    X_training=torch.rand(50, 30), 
    y_training=torch.rand(50, 1)
)

pool = torch.rand(20000, 30)

n_queries = 5

for i in range(n_queries):
    print('Query no. %d' % (i + 1))
    query_idx, query_inst = learner.query(pool, n_instances=1)
    learner.teach(
        X=pool[query_idx], 
        y=torch.rand(1, 1), 
#         only_new=True
    )
    pool = np.delete(pool, query_idx, axis=0)

If only_new=True,the following error will be returned:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-4-67dbdfd957b1> in <module>
     50         X=pool[query_idx],
     51         y=torch.rand(1, 1),
---> 52         only_new=True
     53     )
     54     pool = np.delete(pool, query_idx, axis=0)

c:\users\anaconda3\lib\site-packages\modAL\models\learners.py in teach(self, X, y, bootstrap, only_new, **fit_kwargs)
     97             self._fit_to_known(bootstrap=bootstrap, **fit_kwargs)
     98         else:
---> 99             self._fit_on_new(X, y, bootstrap=bootstrap, **fit_kwargs)
    100
    101

c:\users\anaconda3\lib\site-packages\modAL\models\base.py in _fit_on_new(self, X, y, bootstrap, **fit_kwargs)
    129
    130         if not bootstrap:
--> 131             self.estimator.fit(X, y, **fit_kwargs)
    132         else:
    133             bootstrap_idx = np.random.choice(range(X.shape[0]), X.shape[0], replace=True)

c:\users\anaconda3\lib\site-packages\skorch\net.py in fit(self, X, y, **fit_params)
    852             self.initialize()
    853
--> 854         self.partial_fit(X, y, **fit_params)
    855         return self
    856

c:\users\anaconda3\lib\site-packages\skorch\net.py in partial_fit(self, X, y, classes, **fit_params)
    811         self.notify('on_train_begin', X=X, y=y)
    812         try:
--> 813             self.fit_loop(X, y, **fit_params)
    814         except KeyboardInterrupt:
    815             pass

c:\users\anaconda3\lib\site-packages\skorch\net.py in fit_loop(self, X, y, epochs, **fit_params)
    715
    716         dataset_train, dataset_valid = self.get_split_datasets(
--> 717             X, y, **fit_params)
    718         on_epoch_kwargs = {
    719             'dataset_train': dataset_train,

c:\users\anaconda3\lib\site-packages\skorch\net.py in get_split_datasets(self, X, y, **fit_params)
   1199         if self.train_split:
   1200             dataset_train, dataset_valid = self.train_split(
-> 1201                 dataset, y, **fit_params)
   1202         else:
   1203             dataset_train, dataset_valid = dataset, None

c:\users\anaconda3\lib\site-packages\skorch\dataset.py in __call__(self, dataset, y, groups)
    323             args = args + (to_numpy(y),)
    324
--> 325         idx_train, idx_valid = next(iter(cv.split(*args, groups=groups)))
    326         dataset_train = torch.utils.data.Subset(dataset, idx_train)
    327         dataset_valid = torch.utils.data.Subset(dataset, idx_valid)

c:\users\anaconda3\lib\site-packages\sklearn\model_selection\_split.py in split(self, X, y, groups)
   1327         """
   1328         X, y, groups = indexable(X, y, groups)
-> 1329         for train, test in self._iter_indices(X, y, groups):
   1330             yield train, test
   1331

c:\users\anaconda3\lib\site-packages\sklearn\model_selection\_split.py in _iter_indices(self, X, y, groups)
   1435         n_train, n_test = _validate_shuffle_split(
   1436             n_samples, self.test_size, self.train_size,
-> 1437             default_test_size=self._default_test_size)
   1438
   1439         rng = check_random_state(self.random_state)

c:\users\anaconda3\lib\site-packages\sklearn\model_selection\_split.py in _validate_shuffle_split(n_samples, test_size, train_size, default_test_size)
   1803             'resulting train set will be empty. Adjust any of the '
   1804             'aforementioned parameters.'.format(n_samples, test_size,
-> 1805                                                 train_size)
   1806         )
   1807

ValueError: With n_samples=1, test_size=0.2 and train_size=None, the resulting train set will be empty. Adjust any of the aforementioned parameters.

Thus I'm confused about the mechanism with only_new=False.