Custom Keras Metric throwing error

Question

While trying to implement Intersection over Union (IoU), I've run into a python/keras error that I can't seem to place. In a seperate file I define the following metric:

def computeIoU(y_pred_batch, y_true_batch):
    print y_true_batch.shape[0]
    return np.mean(np.asarray([imageIoU(y_pred_batch[i], y_true_batch[i]) for i in range(y_true_batch.shape[0])]))

def imageIoU(y_pred, y_true):
    y_pred = np.argmax(y_pred, axis=2)
    y_true = np.argmax(y_true, axis=2)
    inter = 0
    union = 0
    for x in range(imCols):
        for y in range(imRows):
            for i in range(num_classes):
                inter += (y_pred[y][x] == y_true[y][x] == i)
                union += (y_pred[y][x] == i or y_true[y][x] == i)
    print inter
    print union
    return float(inter)/union

In the main file I have imported the function and use the metric as follows:

fcn32_model.compile(optimizer=sgd, loss='categorical_crossentropy', metrics=['accuracy', computeIoU])

The error that is thrown is

TypeError: __int__ should return int object

After implementing the above algorithm with the suggested Keras/tf syntax from answers here and on another question the code was changed to:

def iou(y_pred_batch, y_true_batch):
    intersection = tf.zeros(())
    union = tf.zeros(())
    y_pred_batch = K.argmax(y_pred_batch, axis=-1)
    y_true_batch = K.argmax(y_true_batch, axis=-1)
    for i in range(num_classes):
        iTensor = tf.to_int64(tf.fill(y_pred_batch.shape, i))
        intersection = tf.add(intersection, tf.to_float(tf.count_nonzero(tf.logical_and(K.equal(y_true_batch, y_pred_batch), K.equal(y_true_batch, iTensor)))))
        union = tf.add(union, tf.to_float(tf.count_nonzero(tf.logical_or(K.equal(y_true_batch, iTensor), K.equal(y_pred_batch, iTensor)))))
    return intersection/union

score 2 · Answer 1 · answered Apr 16 '18 at 13:32

2

The issue seems to be that you are trying to compute in plain integer,instead its a keras variable.

intersection = K.sum(K.abs(y_true * y_pred), axis=-1)
union_sum = K.sum(K.abs(y_true) + K.abs(y_pred), axis=-1)
IOU = (intersection) / (union_sum- intersection)

answered Apr 16 '18 at 13:32

thefifthjack005

573
3
9

I don't see how this produces the same as my initial function. The size is still 2-dimensional instead of a single value and the values from all classes are taken into account instead of the class that is dominant at a certain pixel. Can you elaborate on these things? – Kroshtan Apr 16 '18 at 14:32
1

@Kroshtan Remember that the loss function is computed *directly in the computational graph*, so the code should be in the language of *tensors*. Your code in question computes *python integers*, it is outside of computational graph. This is perhaps the most annoying limitation of *Keras/Tensorflow* - you need to express everything in terms of tensor operations! – Chan Kha Vu Apr 16 '18 at 20:59

Custom Keras Metric throwing error

1 Answers1