CODEWITHSUNDEEP
pythontensorflowmachine-learningdeep-learning
이한빈
이한빈

Reputation: 55

Tensorflow: How can I design custom cost function?

I am a student studying Object Detection. I implemented a simple object finding model with FCN. The accuracy is 95% but the recall is 0. In short, the model has determined that there are no objects for all areas. I used softmax_cross_entropy as the error function. However, if the model judges that there is no object for the part where the object exists, I want to make a large error. I think I have to design the loss function myself, what should I do?

def model(X, P):
    x = tf.cast(X, tf.float32)
    x = x / 255.0

    net = slim.conv2d(x, 32, kernel_size =(3, 3))
    net = slim.max_pool2d(net, (2, 2)) # 128 160
    net = slim.conv2d(net, 64, kernel_size =(3, 3))
    net = slim.max_pool2d(net, (2, 2)) # 64 80
    net = slim.conv2d(net, 128, kernel_size =(3, 3))
    net = slim.max_pool2d(net, (2, 2)) # 32 40
    net = slim.conv2d(net, 256, kernel_size =(3, 3))
    net = slim.max_pool2d(net, (2, 2)) # 16 20
    net = slim.conv2d(net, 512, kernel_size=(3, 3))
    net = slim.max_pool2d(net, (2, 2)) # 8 10
    net = tf.nn.dropout(net, keep_prob=P)
    net = slim.conv2d(net, 2, kernel_size=(3, 3))
    net = slim.max_pool2d(net, (2, 2)) # 4 5

X = tf.placeholder(tf.uint8 ,[None, 320, 256, 3])
Y = tf.placeholder(tf.int64, [None, 20])
P = tf.placeholder(tf.float32)
y = tf.reshape(Y, (-1, 5, 4))
logits = model(X,P)
arg = tf.argmax(logits, -1)
with tf.name_scope('Optimizer'):
    cost = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=y))
    optimizer = tf.train.AdadeltaOptimizer(0.001).minimize(cost)

enter image description here

Upvotes: 2

Views: 202

Answers (1)

lhk
lhk

Reputation: 30236

You are using sparse_softmax_cross_entropy, this already allows you to add weights to your loss:

tf.losses.sparse_softmax_cross_entropy(
    labels,
    logits,
    weights=1.0,
    ...

weights acts as a coefficient for the loss. If a scalar is provided, then the loss is simply scaled by the given value. If weights is a tensor of shape [batch_size], then the loss weights apply to each corresponding sample.

If your y is binary (1 for object, 0 for empty), you could simply multiply this with a desired factor and pass it in as weights:

penalty_empty = 0.5
penalty_object = 1

weights = penalty_empty + y * (penalty_object - penalty_empty)

Upvotes: 2

Related Questions