File size: 7,086 Bytes
7393a38
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
# coding: utf-8

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import six
import tensorflow as tf
import tensorflow.contrib as tc

from tensorflow.python.training import device_setter
from tensorflow.python.framework import device as pydev
from tensorflow.core.framework import node_def_pb2

from utils import dtype


def local_device_setter(num_devices=1,
                        ps_device_type='cpu',
                        worker_device='/cpu:0',
                        ps_ops=None,
                        ps_strategy=None):
    if ps_ops is None:
        ps_ops = ['Variable', 'VariableV2', 'VarHandleOp']

    if ps_strategy is None:
        ps_strategy = device_setter._RoundRobinStrategy(num_devices)
    if not six.callable(ps_strategy):
        raise TypeError("ps_strategy must be callable")

    def _local_device_chooser(op):
        current_device = pydev.DeviceSpec.from_string(op.device or "")

        node_def = op if isinstance(op, node_def_pb2.NodeDef) else op.node_def
        if node_def.op in ps_ops:
            ps_device_spec = pydev.DeviceSpec.from_string(
                '/{}:{}'.format(ps_device_type, ps_strategy(op)))

            ps_device_spec.merge_from(current_device)
            return ps_device_spec.to_string()
        else:
            worker_device_spec = pydev.DeviceSpec.from_string(worker_device or "")
            worker_device_spec.merge_from(current_device)
            return worker_device_spec.to_string()

    return _local_device_chooser


def _maybe_repeat(x, n):
    if isinstance(x, list):
        assert len(x) == n
        return x
    else:
        return [x] * n


def _reshape_output(outputs):
    # assumption: or outputs[0] are all tensor lists/tuples,
    #             or outputs[0] are dictionaries
    if isinstance(outputs[0], (tuple, list)):
        outputs = list(zip(*outputs))
        outputs = tuple([list(o) for o in outputs])
    else:
        if not isinstance(outputs[0], dict):
            return outputs

        assert isinstance(outputs[0], dict), \
            'invalid data type %s' % type(outputs[0])

        combine_outputs = {}
        for key in outputs[0]:
            combine_outputs[key] = [o[key] for o in outputs]
        outputs = combine_outputs

    return outputs


# Data-level parallelism
def data_parallelism(device_type, num_devices, fn, *args, **kwargs):
    # Replicate args and kwargs
    if args:
        new_args = [_maybe_repeat(arg, num_devices) for arg in args]
        # Transpose
        new_args = [list(x) for x in zip(*new_args)]
    else:
        new_args = [[] for _ in range(num_devices)]

    new_kwargs = [{} for _ in range(num_devices)]

    for k, v in kwargs.items():
        vals = _maybe_repeat(v, num_devices)

        for i in range(num_devices):
            new_kwargs[i][k] = vals[i]

    fns = _maybe_repeat(fn, num_devices)

    # Now make the parallel call.
    outputs = []
    for i in range(num_devices):
        worker = "/{}:{}".format(device_type, i)
        if device_type == 'cpu':
            _device_setter = local_device_setter(worker_device=worker)
        else:
            _device_setter = local_device_setter(
                ps_device_type='gpu',
                worker_device=worker,
                ps_strategy=tc.training.GreedyLoadBalancingStrategy(
                    num_devices, tc.training.byte_size_load_fn)
            )

        with tf.variable_scope(tf.get_variable_scope(), reuse=bool(i != 0),
                               dtype=tf.as_dtype(dtype.floatx())):
            with tf.name_scope("tower_%d" % i):
                with tf.device(_device_setter):
                    outputs.append(fns[i](*new_args[i], **new_kwargs[i]))

    return _reshape_output(outputs)


def parallel_model(model_fn, features, devices, use_cpu=False):
    device_type = 'gpu'
    num_devices = len(devices)

    if use_cpu:
        device_type = 'cpu'
        num_devices = 1

    outputs = data_parallelism(device_type, num_devices, model_fn, features)

    return outputs


def average_gradients(tower_grads, mask=None):
    """Modified from Bilm"""

    # optimizer for single device
    if len(tower_grads) == 1:
        return tower_grads[0]

    # calculate average gradient for each shared variable across all GPUs
    def _deduplicate_indexed_slices(values, indices):
        """Sums `values` associated with any non-unique `indices`."""
        unique_indices, new_index_positions = tf.unique(indices)
        summed_values = tf.unsorted_segment_sum(
            values, new_index_positions,
            tf.shape(unique_indices)[0])
        return summed_values, unique_indices

    average_grads = []
    for grad_and_vars in zip(*tower_grads):
        # Note that each grad_and_vars looks like the following:
        #   ((grad0_gpu0, var0_gpu0), ... , (grad0_gpuN, var0_gpuN))
        # We need to average the gradients across each GPU.

        g0, v0 = grad_and_vars[0]

        if g0 is None:
            # no gradient for this variable, skip it
            tf.logging.warn("{} has no gradient".format(v0.name))
            average_grads.append((g0, v0))
            continue

        if isinstance(g0, tf.IndexedSlices):
            # If the gradient is type IndexedSlices then this is a sparse
            #   gradient with attributes indices and values.
            # To average, need to concat them individually then create
            #   a new IndexedSlices object.
            indices = []
            values = []
            for g, v in grad_and_vars:
                indices.append(g.indices)
                values.append(g.values)
            all_indices = tf.concat(indices, 0)
            if mask is None:
                avg_values = tf.concat(values, 0) / len(grad_and_vars)
            else:
                avg_values = tf.concat(values, 0) / tf.reduce_sum(mask)
            # deduplicate across indices
            av, ai = _deduplicate_indexed_slices(avg_values, all_indices)
            grad = tf.IndexedSlices(av, ai, dense_shape=g0.dense_shape)
        else:
            # a normal tensor can just do a simple average
            grads = []
            for g, v in grad_and_vars:
                # Add 0 dimension to the gradients to represent the tower.
                expanded_g = tf.expand_dims(g, 0)
                # Append on a 'tower' dimension which we will average over
                grads.append(expanded_g)

            # Average over the 'tower' dimension.
            grad = tf.concat(grads, 0)
            if mask is not None:
                grad = tf.boolean_mask(
                    grad, tf.cast(mask, tf.bool), axis=0)
            grad = tf.reduce_mean(grad, 0)

        # the Variables are redundant because they are shared
        # across towers. So.. just return the first tower's pointer to
        # the Variable.
        v = grad_and_vars[0][1]
        grad_and_var = (grad, v)

        average_grads.append(grad_and_var)

    assert len(average_grads) == len(list(zip(*tower_grads)))

    return average_grads