rykov8
commented
7 years ago @neil454 sorry for late reply.
I believe, that you need to change the last activation to sigmoid (you have already done it) and use binary_crossentropy as your loss function for conf_loss and average it along classes, some example code (I don't guarantee that it works) for the method in class MultiboxLoss (mainly taken from Keras binary crossentropy loss):
def _multilabel_loss(self, y_true, y_pred):
"""Compute multilabel loss.
# Arguments
y_true: Ground truth targets,
tensor of shape (?, num_boxes, num_classes).
y_pred: Predicted logits,
tensor of shape (?, num_boxes, num_classes).
# Returns
multilabel_loss: Multilabel loss, tensor of shape (?, num_boxes).
"""
# this is just to be sure not to compute log(0)
y_pred = tf.clip_by_value(y_pred, 1e-15, 1 - 1e-15)
# convert to logits
logits = tf.log(y_pred / (1 - y_pred))
# binary crossentropy
multilabel_loss = tf.nn.sigmoid_cross_entropy_with_logits(labels=y_true,
logits=logits)
# average along classes (not sure, probably, tf crossentropy above do this itself)
multilabel_loss = tf.reduce_mean(multilabel_loss, reduction_indices=-1)
return multilabel_loss
neil454
oarriaga
Hello, thank you so much for this Keras implementation of SSD!
I have successfully ported the caffe weights for SSD trained on the COCO dataset (300x300 input, 80+1 classes), and now I'm trying to utilize these weights to help retrain SSD on my specific problem.
I need SSD to output 200-some attributes instead of 81 object classes, and since one object can have multiple attributes, I need SSD to output class scores that don't sum to 1.
So I tried just re-training without any major changes (had to randomly initialize the weights of 6 layers that relied on COCO's 81 class output, but loaded the rest just fine), and my training loss was stuck at around 200.
I then realized this would never train because the class score outputs are normalized to sum to 1, so I changed the Activation function on the last layer of SSD from "softmax" to "sigmoid" (maybe I should use "tanh" instead?), and I'm currently training successfully I think, but I won't know for a while. The loss started at 32 and is now down to 7 after 8000 samples, and still decreasing nicely.
Anyways, I was just wondering about SSD's custom loss function, since I see it uses a softmax loss for conf_loss, in ssd_training.py. Should I change this to some other loss function? If so, which one?