About loss function

[dyl0101]

Hello,Thank you for sharing your code.I want to know some detail about the loss function. In the function,what does logits and factor_lambda mean? Isboundary_target the predicted result? Thanks. `def refine_loss(logits, labels, boundary_target): gamma1 = 0.5 gamma2 = 1-gamma1 factor_lambda= 1.5

dy_logits, dx_logits = tf.image.image_gradients(logits)
dy_labels, dx_labels = tf.image.image_gradients(labels)

# magnitudes of logits and labels gradients
Mpred = tf.sqrt(tf.square(dy_logits)+tf.square(dx_logits))
Mimg = tf.sqrt(tf.square(dy_labels)+tf.square(dx_labels))

# define cos loss and mag loss
cosL = (1-tf.abs(dx_labels*dx_logits+dy_labels*dy_logits))*Mpred
magL = tf.maximum(factor_lambda*Mimg-Mpred,0)

# define mask
M_bound = boundary_target/255.

# define total refine loss
refineLoss = (gamma1*cosL + gamma2*magL)*M_bound
return tf.reduce_mean(refineLoss)
# return  tf.reduce_mean(refineLoss), Mpred, Mimg`

[lewisluk]

Hello,Thank you for sharing your code.I want to know some detail about the loss function. In the function,what does logits and factor_lambda mean? Isboundary_target the predicted result? Thanks. `def refine_loss(logits, labels, boundary_target): gamma1 = 0.5 gamma2 = 1-gamma1 factor_lambda= 1.5

dy_logits, dx_logits = tf.image.image_gradients(logits)
dy_labels, dx_labels = tf.image.image_gradients(labels)

# magnitudes of logits and labels gradients
Mpred = tf.sqrt(tf.square(dy_logits)+tf.square(dx_logits))
Mimg = tf.sqrt(tf.square(dy_labels)+tf.square(dx_labels))

# define cos loss and mag loss
cosL = (1-tf.abs(dx_labels*dx_logits+dy_labels*dy_logits))*Mpred
magL = tf.maximum(factor_lambda*Mimg-Mpred,0)

# define mask
M_bound = boundary_target/255.

# define total refine loss
refineLoss = (gamma1*cosL + gamma2*magL)*M_bound
return tf.reduce_mean(refineLoss)
# return  tf.reduce_mean(refineLoss), Mpred, Mimg`

logits: model predicted result

factor_lambda: a coefficient mentioned in the paper, "λ is a factor that balances distributional differences between image and output confidence map. In our experiment, λ = 1.5"

boundary target: get from mask by using canny edge detection

By the way, I haven't fully achieved the result of their paper, so please be reminded that there might be errors in my code. I will add this in the readme.