8000 What is the usage of SECOND_ORDER_GRAD_ITER=0 and self.total_loss1? · Issue #11 · JWSoh/MZSR · GitHub
[go: up one dir, main page]
Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

What is the usage of SECOND_ORDER_GRAD_ITER=0 and self.total_loss1? #11

Open
qhanson opened this issue Jun 2, 2020 · 3 comments
Open

What is the usage of SECOND_ORDER_GRAD_ITER=0 and self.total_loss1? #11

qhanson opened this issue Jun 2, 2020 · 3 comments

Comments

@qhanson
Copy link
qhanson commented Jun 2, 2020

What is the usage of SECOND_ORDER_GRAD_ITER=0 and self.total_loss1?
As for SECOND_ORDER_GRAD_ITER=0:

if step == SECOND_ORDER_GRAD_ITER:
       second_grad=sess.run(self.second_grad_on)

If we have finished pre-training on large scale datasets, I think it is useless in this meta-transfer learning step.
As for self.total_loss1:

self.total_loss1 = tf.reduce_sum(self.lossesa) / tf.to_float(self.META_BATCH_SIZE)
self.pretrain_op = tf.train.AdamOptimizer(self.META_LR).minimize(self.total_loss1)
        
self.gvs = self.opt.compute_gradients(self.weighted_total_losses2)
self.metatrain_op= self.opt.apply_gradients(self.gvs)

sess.run(self.metatrain_op, feed_dict=feed_dict)

In this meta-transfer learning step, total_loss1 is never used for optimizers. Is it correct?
@JWSoh
Copy link
Owner
JWSoh commented Jun 8, 2020

MAML requires 2nd-order gradients, which requires large computation.
For the fast training, we use 1st-order approximation of the gradients at the beginning of the training, and then after SECOND_ORDER_GRAD_ITER, we use the full gradients.

Therefore, SECOND_ORDER_GRAD_ITER is to decide how many steps to approximate the gradients within 1st-order.

For the self.total_loss1, you are right. it is not used for the training. You may ignore that loss and corresponding optimizer.

@BassantTolba1234
Copy link

Dear Sir,
Amazing work ! Congratulation!!
please , I have a question.can you kindly provide me with the full path I should insert of checkpoint the trained large scale training model to be able to use it as a pre-trained to meta transfer training? as it says that there is no check point file
I'm waiting for your reply.
Thanks in advance

@BassantTolba1234
Copy link

Please can you kindly explain me how to calculate this weight loss ?

def get_loss_weights(self):
loss_weights = tf.ones(shape=[self.TASK_ITER]) * (1.0/self.TASK_ITER)
decay_rate = 1.0 / self.TASK_ITER / (10000 / 3)
min_value= 0.03 / self.TASK_ITER

    loss_weights_pre = tf.maximum(loss_weights[:-1] - (tf.multiply(tf.to_float(self.global_step), decay_rate)), min_value)

    loss_weight_cur= tf.minimum(loss_weights[-1] + (tf.multiply(tf.to_float(self.global_step),(self.TASK_ITER- 1) * decay_rate)), 1.0 - ((self.TASK_ITER - 1) * min_value))
    loss_weights = tf.concat([[loss_weights_pre], [[loss_weight_cur]]], axis=1)
    return loss_weights

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees
No one assigned
Labels
None yet
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
3 participants
@JWSoh @qhanson @BassantTolba1234
0