Models: [Deeplab v3+] Modify Loss function

[lizleo]

@aquariusjay
Hi Jay,
I want to modify the loss function in train_utils.py, for example, change "tf.losses.softmax_cross_entropy()" to "tf.losses.softmax_cross_entropy() + sigmoid_cross_entropy_with_logits()", but I didn't see the assignment operation like "loss = tf.losses.softmax_cross_entropy()", so I didn't know how to modify the loss function. Could you please help me? Many thanks for your help.

[aquariusjay]

Hi lizleo,

Thanks for bringing up this issue.
I think you could do something similar to what we do here:
Specifically, you first compute the losses for softmax and for sigmoid respectively, and then you call tf.losses.add_loss(loss_sum).

Cheers,

[lizleo]

Hi @aquariusjay ,

Thanks for your reply. Now I know how to modify the loss function. Thanks again for your help.

Best,

[aquariusjay]

Sure thing. Glad to know that you figure it out.

Cheers,