Keras class_weight in multi-label binary classific

Question

Having trouble using class_weight for my multi-label problem. That is, each label is either 0 or 1, but there are many labels for each input sample.

The code (with random data for MWE purposes):

import tensorflow as tf
from keras.models import Sequential, Model
from keras.layers import Input, Concatenate, LSTM, Dense
from keras import optimizers
from keras.utils import to_categorical
from keras import backend as K
import numpy as np

# from http://www.deepideas.net/unbalanced-classes-machine-learning/
def sensitivity(y_true, y_pred):
        true_positives = tf.reduce_sum(tf.round(K.clip(y_true * y_pred, 0, 1)))
        possible_positives = tf.reduce_sum(tf.round(K.clip(y_true, 0, 1)))
        return true_positives / (possible_positives + K.epsilon())

# from http://www.deepideas.net/unbalanced-classes-machine-learning/    
def specificity(y_true, y_pred):
        true_negatives = tf.reduce_sum(K.round(K.clip((1-y_true) * (1-y_pred), 0, 1)))
        possible_negatives = tf.reduce_sum(K.round(K.clip(1-y_true, 0, 1)))
        return true_negatives / (possible_negatives + K.epsilon())

def to_train(a_train, y_train):
        hours_np = [np.arange(a_train.shape[1])]*a_train.shape[0]
        train_hours = to_categorical(hours_np)
        n_samples = a_train.shape[0]
        n_classes = 4
        features_in = np.zeros((n_samples, n_classes))
        supp_feat = np.random.choice(n_classes, n_samples)
        features_in[np.arange(n_samples), supp_feat] = 1

        #This model has 3 separate inputs
        seq_model_in = Input(shape=(1,),batch_shape=(1, 1, a_train.shape[2]), name='seq_model_in')
        feat_in = Input(shape=(1,), batch_shape=(1, features_in.shape[1]), name='feat_in')
        feat_dense = Dense(1)(feat_in)
        hours_in = Input(shape=(1,), batch_shape=(1, 1, train_hours.shape[2]), name='hours_in')

        #Model intermediate layers
        t_concat = Concatenate(axis=-1)([seq_model_in, hours_in])
        lstm_layer = LSTM(1, batch_input_shape=(1, 1, (a_train.shape[2]+train_hours.shape[2])), return_sequences=False, stateful=True)(t_concat)
        merged_after_lstm = Concatenate(axis=-1)([lstm_layer, feat_dense]) #may need another Dense() after
        dense_merged = Dense(a_train.shape[2], activation="sigmoid")(merged_after_lstm)

        #Define input and output to create model, and compile
        model = Model(inputs=[seq_model_in, feat_in, hours_in], outputs=dense_merged)
        model.compile(loss='binary_crossentropy', optimizer='adam', metrics=[sensitivity, specificity])

        class_weights = {0.:1., 1.:118.}
        seq_length = 23

        #TRAINING (based on http://philipperemy.github.io/keras-stateful-lstm/)
        for epoch in range(2):
            for i in range(a_train.shape[0]):
                    y_true_1 = np.expand_dims(y_train[i,:], axis=1)
                    y_true = np.swapaxes(y_true_1, 0, 1)
                    #print 'y_true', y_true.shape
                    for j in range(seq_length-1):
                            input_1 = np.expand_dims(np.expand_dims(a_train[i][j], axis=1), axis=1)
                            input_1 = np.reshape(input_1, (1, 1, a_train.shape[2]))
                            input_2 = np.expand_dims(np.array(features_in[i]), axis=1)
                            input_2 = np.swapaxes(input_2, 0, 1)
                            input_3 = np.expand_dims(np.array([train_hours[i][j]]), axis=1)
                            tr_loss, tr_sens, tr_spec = model.train_on_batch([input_1, input_2, input_3], y_true, class_weight=class_weights)
                    model.reset_states()
       return 0

a_train = np.random.normal(size=(50,24,5625))
y_train = a_train[:, -1, :]
a_train = a_train[:, :-1, :]
y_train[y_train > 0.] = 1.
y_train[y_train < 0.] = 0.
to_train(a_train, y_train)

The error that I'm getting is:

ValueError: `class_weight` must contain all classes in the data. The classes set([330]) exist in the data but not in `class_weight`.

The value inside 'set([...])' changes at each run. But as I said, the only two classes in the data are 0 and 1; there are just multiple labels per sample. So for example, one response (y_train) looks like this:

print y_train[0,:]
#[ 0.  0.  1. ...,  0.  1.  0.]

How can I use class_weights for a multi-label problem in Keras?

Answer 1

For multi-label I found two options:

1. create multi-output model, 1 output per 1 label and pass standard class_weight dictionary
2. create weights_aware_binary_crossentropy loss which can calculate mask based on passed list of class_weight dictionaries and y_true and do:

K.binary_crossentropy(y_true, y_pred) * mask

Answer 2

you can create a call back which appends the index of the label to a list for example:

y = [[0,1,0,1,1],[0,1,1,0,0]]

would create a list: category_list = [1, 3, 4, 1, 2]

where each instance of a label is tallied in the category_list

then you can use

weighted_list = class_weight.compute_class_weight('balanced', np.unique(category_list), category_list)

Then just transform weighted_list to a dictionary to use in Keras.

Answer 3

Yep. It's a known bug in keras (issue #8011). Basically, keras code assumes one-hot encoding, when determines the number of classes, not multi-label ordinal encoding.

keras/engine/training.py:

# if 2nd dimension is greater than 1, it must be one-hot encoded, 
# so let's just get the max index...
if y.shape[1] > 1:
  y_classes = y.argmax(axis=1)

I can't think of a better workaround right now, other than set y_true[:, 1] = 1, i.e. "reserve" the 1 position in y to be always one. This will cause y_classes = 1 (which is the right value in binary classification).

Why does it work? The code fails when y_true[i] gets the value like [0, 0, ..., 0, 1, ...] with some number of leading zeros. Keras implementation (mistakenly) estimates the number of classes via the index of the max element, which turns out to be some j > 1 for which y[i][j] = 1. This makes Keras engine think there are more than 2 classes, so the provided class_weights are wrong. Setting y_true[i][1] = 1 makes sure that j <= 1 (because np.argmax picks the smallest max index), which allows to bypass keras guards.