def request_model(peer_ranks, variables, mode, peer_selection_strategy):
import tensorflow as tf
var_shapes = [var.shape for var in variables]
var_sizes = [var.shape.num_elements() for var in variables]
# Remove self rank from the list
peer_ranks.remove(current_rank())
if mode == 'async':
print("Request a model asynchronously.")
request_model = _op_lib.async_request_model(
variables,
self_rank=current_rank(),
ranks=peer_ranks,
var_type_size=variables[0].dtype.size,
var_sizes=var_sizes,
shapes=var_shapes,
peer_selection_strategy=peer_selection_strategy)
elif mode == 'sync':
print("Request a model synchronously.")
request_model = _op_lib.request_model(
variables,
self_rank=current_rank(),
ranks=peer_ranks,
var_type_size=variables[0].dtype.size,
var_sizes=var_sizes,
shapes=var_shapes,
peer_selection_strategy=peer_selection_strategy)
else:
raise Exception("Invalid type of model request mode")
return request_model
def model_averaging(peer_ranks, variables, mode, peer_selection_strategy):
import tensorflow as tf
var_sizes = [var.shape.num_elements() for var in variables]
# Remove self rank from the list
peer_ranks.remove(current_rank())
if mode == 'async':
print(
"Applying model averaging with a model requested asynchronously.")
model_averaging = _op_lib.async_model_averaging(
variables,
self_rank=current_rank(),
ranks=peer_ranks,
var_type_size=variables[0].dtype.size,
var_sizes=var_sizes,
peer_selection_strategy=peer_selection_strategy)
elif mode == 'sync':
print("Applying model averaging with a model requested synchronously.")
model_averaging = _op_lib.model_averaging(
variables,
self_rank=current_rank(),
ranks=peer_ranks,
var_type_size=variables[0].dtype.size,
var_sizes=var_sizes,
peer_selection_strategy=peer_selection_strategy)
else:
raise Exception("Invalid type of model request mode.")
return model_averaging
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())
