With some help from a previous question I devised the following implementation of IoU:
def iou(y_pred_batch, y_true_batch):
intersection = tf.zeros(())
union = tf.zeros(())
y_pred_batch = np.argmax(y_pred_batch, axis=-1)
y_true_batch = np.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
I use the following lines to test the code:
sess = tf.InteractiveSession()
y_true_batch = np.asarray([np.random.rand(imRows, imCols, num_classes) for i in range(2)])
y_pred_batch = np.asarray([np.random.rand(imRows, imCols, num_classes) for i in range(2)])
print (iou(y_true_batch, y_pred_batch).eval())
sess.close()
This produces a value of ~0.02, which is to be expected from randomly initialized values. However, when I use this metric in my keras model, the metric returns 1.0000 from epoch 1 onwards, which is obviously wrong. I have no idea why this is and any help would be appreciated.
just changed the
np.argmax()
to
from keras import backend as K
K.argmax()
The reason being when you compute using np.argmax() no tensor is created ,the code should be in the language of tensors. you need to perform everything in terms of tensor operations in keras.
for keras testing.
y_true = np.asarray([np.random.rand(4,4, 4) for i in range(2)])
y_pred = np.asarray([np.random.rand(4, 4, 4) for i in range(2)])
iou_value = iou(
K.variable(y_true),
K.variable(y_pred),
).eval(session=K.get_session())
print('iou', iou)