import theano
import theano.tensor as T
import numpy as np
# predictions is the argmax of the posterior
def accuracy_instance(predictions, targets, n=[1, 2, 3, 4, 5, 10], \
nb_classes=5, nb_samples_per_class=10, batch_size=1):
accuracy_0 = theano.shared(np.zeros((batch_size, nb_samples_per_class), \
dtype=theano.config.floatX))
indices_0 = theano.shared(np.zeros((batch_size, nb_classes), \
dtype=np.int32))
batch_range = T.arange(batch_size)
def step_(p, t, acc, idx):
acc = T.inc_subtensor(acc[batch_range, idx[batch_range, t]], T.eq(p, t))
idx = T.inc_subtensor(idx[batch_range, t], 1)
return (acc, idx)
(raw_accuracy, _), _ = theano.foldl(step_, sequences=[predictions.dimshuffle(1, 0), \
targets.dimshuffle(1, 0)], outputs_info=[accuracy_0, indices_0])
accuracy = T.mean(raw_accuracy / nb_classes, axis=0)
return accuracy
