34 lines
1.4 KiB
Python
34 lines
1.4 KiB
Python
|
import numpy as np
|
||
|
|
import tensorflow as tf
|
||
|
|
from mpi4py import MPI
|
||
|
|
|
||
|
|
class MpiAdamOptimizer(tf.train.AdamOptimizer):
|
||
|
|
"""Adam optimizer that averages gradients across mpi processes."""
|
||
|
|
|
||
|
|
def __init__(self, comm, **kwargs):
|
||
|
|
self.comm = comm
|
||
|
|
tf.train.AdamOptimizer.__init__(self, **kwargs)
|
||
|
|
|
||
|
|
def compute_gradients(self, loss, var_list, **kwargs):
|
||
|
|
grads_and_vars = tf.train.AdamOptimizer.compute_gradients(self, loss, var_list, **kwargs)
|
||
|
|
grads_and_vars = [(g, v) for g, v in grads_and_vars if g is not None]
|
||
|
|
flat_grad = tf.concat([tf.reshape(g, (-1,)) for g, v in grads_and_vars], axis=0)
|
||
|
|
shapes = [v.shape.as_list() for g, v in grads_and_vars]
|
||
|
|
sizes = [int(np.prod(s)) for s in shapes]
|
||
|
|
|
||
|
|
_task_id, num_tasks = self.comm.Get_rank(), self.comm.Get_size()
|
||
|
|
buf = np.zeros(sum(sizes), np.float32)
|
||
|
|
|
||
|
|
def _collect_grads(flat_grad):
|
||
|
|
self.comm.Allreduce(flat_grad, buf, op=MPI.SUM)
|
||
|
|
np.divide(buf, float(num_tasks), out=buf)
|
||
|
|
return buf
|
||
|
|
|
||
|
|
avg_flat_grad = tf.py_func(_collect_grads, [flat_grad], tf.float32)
|
||
|
|
avg_flat_grad.set_shape(flat_grad.shape)
|
||
|
|
avg_grads = tf.split(avg_flat_grad, sizes, axis=0)
|
||
|
|
avg_grads_and_vars = [(tf.reshape(g, v.shape), v)
|
||
|
|
for g, (_, v) in zip(avg_grads, grads_and_vars)]
|
||
|
|
|
||
|
|
return avg_grads_and_vars
|