Algorithm 2 inconsistent with code about fedprox HOT 7 CLOSED

The updates of FedProx is defined here: https://github.com/litian96/FedProx/blob/master/flearn/optimizer/pgd.py#L29-L36
where var is the current local model you are optimizing for, and vstar is the current global model which is the starting point of solving for the local subproblem.

from fedprox.

Thank you, It makes sense now!

from fedprox.

Hi Tian,

Thanks for previous reply, but there is one place I still don't figure out. The paper shows objective function is optimizing Fk(w) and ||w - w^t|| simultaneously, but the code shows that we are calculating loss and ||w - w^t|| seperately. I found issue #19 #18 and #6 are quite useful, but not exactly what I need.

Take mnist task as an example, the code first calculate the cross entropy loss in
grads_and_vars = optimizer.compute_gradients(loss), this part does not consider ||w - w^t||.
By calling var_update = state_ops.assign_sub(var, lr_t*(grad + mu_t*(var-vstar))) in
train_op = optimizer.apply_gradients(grads_and_vars, global_step=tf.train.get_global_step()) we consider the proximal term.

  def create_model(self, optimizer):
      """Model function for Logistic Regression."""
      features = tf.placeholder(tf.float32, shape=[None, 784], name='features')
      labels = tf.placeholder(tf.int64, shape=[None,], name='labels')
      logits = tf.layers.dense(inputs=features, units=self.num_classes, kernel_regularizer=tf.contrib.layers.l2_regularizer(0.001))
      predictions = {
          "classes": tf.argmax(input=logits, axis=1),
          "probabilities": tf.nn.softmax(logits, name="softmax_tensor")
          }
      loss = tf.losses.sparse_softmax_cross_entropy(labels=labels, logits=logits)
      # set_trace()
      grads_and_vars = optimizer.compute_gradients(loss)
      grads, _ = zip(*grads_and_vars)
      train_op = optimizer.apply_gradients(grads_and_vars, global_step=tf.train.get_global_step())
      eval_metric_ops = tf.count_nonzero(tf.equal(labels, predictions["classes"]))
      return features, labels, train_op, grads, eval_metric_ops, loss

Therefore, it seems that the code is calculating the gradient from Fk's perspective, and then subtract a proximal part to get the final gradient. I also debuged the code to check who has called pgd's _apply_dense(), it shows that this method is only used by optimizer.apply_gradient(), instead of compute_gradient().

Thank you in advance :)

from fedprox.

You are exactly right that we first need to obtain the gradients of \nabla F_k (w_local) + \mu (w_local - w_global), and then apply the gradients to w_local.
What is your question?

from fedprox.

My question is compute_gradient() only calculates \nabla F_k(w_local). How can we find argmin_w h_k without considering \nabla (w_local - w_global)?

I know we considered \mu (w_local - w_global) in apply_gradient(), but this proximal term not necessary leads to argmin h_k.

from fedprox.

The gradient of h_k is \nabla F_k (w_local) + \mu (w_local - w_global), which is what we apply to update w_local (specifically, grad + mu_t * (var-vstar) ).

from fedprox.

Ohhhhh, I see, really appreciate for the clarification :)

from fedprox.