def get_peer_latencies():
"""Returns the vector V of round-trip time from this peer to all other peers.
For the peer of rank i, V[j] is the RTT from i to j (j != i), V[i] = 0.
"""
return _op_lib.kungfu_get_peer_latencies(
cluster_size=current_cluster_size())
def get_neighbour_mask(edges):
"""Compute a bool vector of neighbours for the current peer.
For the peer of rank i, v[j] = true if (i, j) is an edge of the MST,
otherwise v[j] = false.
"""
return _op_lib.kungfu_get_neighbour_mask(
edges, self_rank=current_rank(), cluster_size=current_cluster_size())
def get_peer_latencies(local_step=None):
"""Returns the vector V of round-trip time from this peer to all other peers.
For the peer of rank i, V[j] is the RTT from i to j (j != i), V[i] = 0.
"""
# FIXME: don't require input
if local_step is None:
import tensorflow as tf
local_step = tf.Variable(tf.zeros([], tf.int64), trainable=False)
return _op_lib.kungfu_get_peer_latencies(
local_step, cluster_size=current_cluster_size())
def _parse_schedule(schedule, batch_size, num_train):
# schedule is of the form
# f1;e1;f2;e2;f3;e3
tokens = schedule.split(",")
print("Num train: " + str(num_train))
print("Batch size: " + str(batch_size))
to_gs = lambda epoch: int(epoch * num_train /
(batch_size * current_cluster_size()))
pairs = [(to_gs(int(t.split(":")[0])), float(t.split(":")[1]))
for t in tokens]
steps, fractions = zip(*pairs)
print("Steps: " + str(steps))
print("Fractions: " + str(fractions))
return steps, fractions
