Do we have to set eval_size with a fraction ? about nolearn HOT 15 CLOSED

@dnouri I am sorry if I didn't describe my question clearly. Let me try again. Suppose I use the example code you provided above, that is :

    net = nolearn.lasagne.NeuralNet(..., eval_size = 0.2, verbose = 1)
    Xt, Xv, yt, yv = net.train_test_split(X, y, net.eval_size)
    net.fit(Xt, yt)

My question is that when the net is looping over training examples(i.e. Xt), it will print the training loss, validation loss and validation accuracy since verbose is set 1. Will this validation accuracy be calculated by Xv, yv? Or in another way: the net.fit( ) call will again separate another 20% of Xt as validation data and use this to calculate the accuracy. In such scenario, the actual training data size is X.shape[0] * 80% * 80%? Please clarify me. Thank you .

from nolearn.

@dnouri Also, what if I would like to use a validation dataset generated previously by myself instead of separating from X_train?

from nolearn.

The way that the train_test_split method is set up, you can just call it again and get the same split. So to get the loss on the training set after training finished, you could do something like this:

    Xt, Xv, yt, yv = net.train_test_split(X, y, net.eval_size)
    yp = net.predict_proba(Xt)

Please reopen if you think there's an issue, or if I've misunderstood. Note it's also possible to override with your own train_test_split entirely, to do whatever you want.

from nolearn.

Also, I'm noticing that the issue title seemingly has nothing to do with the issue that you describe. What's going on?

from nolearn.

The validation loss will be calculated with the same Xv and yv.

from nolearn.

@dnouri , then here comes my confusion. Why don't just explicitly pass Xt, Xv, yt, yv to fit() function ?
Oh, I see. May be we should keep the same style as sklearn pipeline provides ?

from nolearn.

Yes, same style as sklearn.

from nolearn.

In that case, you might want to sublass NeuralNet and implement your own train_test_split.

from nolearn.

@dnouri I see. I have also noticed that you have updated the train_test_split() function in the Github. But the pip install nolearn version keeps the same. In the Github version, there is a check for eval_size. As I understand, we can pass eval_size = None in the Github version. And in such case, the validation accuracy will be calculated actually by X_train ?

from nolearn.

There will be no meaningful validation accuracy checks then. To use the Git version, download from Github and run python setup.py develop.

from nolearn.

@dnouri No meaningful! No wonder that I pass eval_size = None and it prints out Nan. Also thanks for your tips of using Git version.

from nolearn.

@dnouri Dainel. There is one question that obsessed me for days. Inspired by your posted Kaggle Facial Detection Blog, I participated in the Kaggle MNIST competition and get 99.4% accuracy in submission by using the same CNN architecture as you suggested in the blog. While I am trying to classify a TinnyImageNet dataset with 200 classes. Still I am using the same network, and I define a subclass named myNeuralNet by inheriting nolearn.lasagne.NeuralNet by a slight modification based on our previous discussion. That is, when eval_size = None we will pass X_val and y_val to myNeuralNet so that I can set my own validation dataset in the train_test_split function. Then here comes my confusion. I just use 500 data examples(X_train) for training, and deliberately set X_val = X_train. In such case, I am expecting overfitting these 500 training examples after a few epochs. If everything goes right, it will print a higher and higher validation accuracy overing looping epochs. However this is what I get.

In case the image collapses, the url is: http://imgur.com/mYlrS8i
My snippet code for defining CNN looks like this:

conv = myNeuralNet(
    layers = [
        ('input', layers.InputLayer),

        ('conv1', Conv2DLayer),
        ('pool1', MaxPool2DLayer),
        ('dropout1', layers.DropoutLayer),

        ('conv2', Conv2DLayer),
        ('pool2', MaxPool2DLayer), 
        ('dropout2', layers.DropoutLayer),      

        ('conv3', Conv2DLayer),
        ('pool3', MaxPool2DLayer), 
        ('dropout3', layers.DropoutLayer),   

        ('hidden4', layers.DenseLayer),
        ('dropout4', layers.DropoutLayer),

        ('hidden5', layers.DenseLayer),

        ('output', layers.DenseLayer),
        ],
    # layer parameters:
    input_shape = (None, 3, 64, 64),  # 64x64 input pixels per batch

    conv1_num_filters = 32, conv1_filter_size = (3, 3), pool1_ds = (2, 2), dropout1_p = 0.1,
    conv2_num_filters = 64, conv2_filter_size=(2, 2), pool2_ds=(2, 2), dropout2_p = 0.2,
    conv3_num_filters = 128, conv3_filter_size = (2, 2), pool3_ds = (2, 2), dropout3_p = 0.3, 
    hidden4_num_units = 500, dropout4_p = 0.5, 
    hidden5_num_units = 500,
    output_num_units = 200,  # 200 labels

    conv1_nonlinearity = rectify, conv2_nonlinearity = rectify, conv3_nonlinearity = rectify,
    hidden4_nonlinearity = rectify, hidden5_nonlinearity = rectify,
    output_nonlinearity = softmax,  # output layer uses softmax function

    # optimization method:
    #update = nesterov_momentum,
    update = rmsprop,
    update_learning_rate = 0.001,
    #update_momentum = 0.9,

    eval_size = None,
    X_valid = X_train,
    y_valid = y_train,


    on_epoch_finished = [early_stopping],
    on_training_finished = [early_stopping.load_best_weights],

    batch_iterator_train = BatchIterator(batch_size = 20),
    batch_iterator_test = BatchIterator(batch_size = 20),

    max_epochs = MAX_EPOCHS,  # we want to train this many epochs
    verbose = 1,
    )

conv.fit(X_val, y_val)

So, do you see anything wrong or any suggestions are welcomed. Thank you.

from nolearn.

What is your
on_training_finished = [early_stopping.load_best_weights],
doing? I assume you are just loading the best weights after all the training epochs are finished. Can you post the code for that, maybe it is doing something weird.

from nolearn.

@msegala Thank you for your reply and trying to help. Actually I am using the early_stopping code provided by @cancan101 in previous issues. And here is the code:

class EarlyStopping(object):
    """
    Early stopping strategy which is used to prevent from overfitting the training data.
    """
    def __init__(self, patience = 100):
        self.patience = patience
        self.best_valid = np.inf
        self.best_valid_epoch = 0
        self.best_valid_accuracy = 0
        self.best_weights = None

    def __call__(self, nn, train_history):
        current_valid = train_history[-1]['valid_loss']
        current_epoch = train_history[-1]['epoch']
        current_valid_accuracy = train_history[-1]['valid_accuracy']

        # Use validation loss validation or accuracy 
        if current_valid_accuracy > self.best_valid_accuracy:
        #if current_valid < self.best_valid:
            self.best_valid = current_valid
            self.best_valid_epoch = current_epoch
            self.best_valid_accuracy = current_valid_accuracy
            self.best_weights = [w.get_value() for w in nn.get_all_params()]
        elif self.best_valid_epoch + self.patience < current_epoch:
            if nn.verbose:
                print("Early stopping.")
                print("Best valid loss was {:.6f} at epoch {} with accuracy {}.".format(
                    self.best_valid, self.best_valid_epoch, self.best_valid_accuracy))
            nn.load_weights_from(self.best_weights)
            if nn.verbose:
                print("Weights set.")
            raise StopIteration()

    def load_best_weights(self, nn, train_history):
        print("Training stage finishes. Best valid loss was {:.6f} at epoch {} with accuracy {}.".format(
            self.best_valid, self.best_valid_epoch, self.best_valid_accuracy))
        nn.load_weights_from(self.best_weights) 

from nolearn.

@dnouri Danniel, still waiting for your advice and suggestions.

from nolearn.