__init__.py 5.9 KB
Newer Older
1
#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
D
dzhwinter 已提交
2
#
D
dzhwinter 已提交
3 4 5
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
D
dzhwinter 已提交
6
#
D
dzhwinter 已提交
7
#     http://www.apache.org/licenses/LICENSE-2.0
D
dzhwinter 已提交
8
#
D
dzhwinter 已提交
9 10 11 12 13 14
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

Y
Yang Yu 已提交
15
from __future__ import print_function
P
peizhilin 已提交
16
import os
17
# import all class inside framework into fluid module
18 19
from . import framework
from .framework import *
Y
Yang Yu 已提交
20
# import all class inside executor into fluid module
21 22
from . import executor
from .executor import *
W
Wang Guibao 已提交
23 24 25 26 27 28 29

from . import data_feed_desc
from .data_feed_desc import *

from . import async_executor
from .async_executor import *

30 31
from . import trainer
from . import inferencer
H
Helin Wang 已提交
32

33 34 35 36
from . import io
from . import evaluator
from . import initializer
from . import layers
37
from . import imperative
38 39 40 41 42 43 44 45
from . import contrib
from . import nets
from . import optimizer
from . import backward
from . import regularizer
from . import average
from . import metrics
from . import transpiler
Q
Qiao Longfei 已提交
46
from . import distribute_lookup_table
47 48
from .param_attr import ParamAttr, WeightNormParamAttr
from .data_feeder import DataFeeder
S
sneaxiy 已提交
49
from .core import LoDTensor, LoDTensorArray, CPUPlace, CUDAPlace, CUDAPinnedPlace, Scope, _Scope
50
from .transpiler import DistributeTranspiler, \
G
gongweibao 已提交
51
    memory_optimize, release_memory, DistributeTranspilerConfig
52 53 54 55 56 57 58
from .lod_tensor import create_lod_tensor, create_random_int_lodtensor
from . import clip
from . import profiler
from . import unique_name
from . import recordio_writer
from . import parallel_executor
from .parallel_executor import *
X
Xin Pan 已提交
59 60
from . import compiler
from .compiler import *
X
Xin Pan 已提交
61
from paddle.fluid.layers.math_op_patch import monkey_patch_variable
62 63

Tensor = LoDTensor
Y
Yang Yu 已提交
64

W
Wu Yi 已提交
65
__all__ = framework.__all__ + executor.__all__ + \
66
    trainer.__all__ + inferencer.__all__ + transpiler.__all__ + \
W
Wang Guibao 已提交
67
    parallel_executor.__all__ + lod_tensor.__all__ + \
X
Xin Pan 已提交
68
    data_feed_desc.__all__ + async_executor.__all__ + compiler.__all__ + [
69 70 71 72
        'io',
        'initializer',
        'layers',
        'contrib',
73
        'imperative',
74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94
        'transpiler',
        'nets',
        'optimizer',
        'learning_rate_decay',
        'backward',
        'regularizer',
        'LoDTensor',
        'LoDTensorArray',
        'CPUPlace',
        'CUDAPlace',
        'CUDAPinnedPlace',
        'Tensor',
        'ParamAttr',
        'WeightNormParamAttr',
        'DataFeeder',
        'clip',
        'profiler',
        'unique_name',
        'recordio_writer',
        'Scope',
    ]
95 96


Y
Yang Yu 已提交
97
def __bootstrap__():
98 99
    """
    Enable reading gflags from environment variables.
Y
Yu Yang 已提交
100

101 102 103 104
    Returns:
        None
    """
    import sys
Y
Yang Yu 已提交
105
    import os
J
JiabinYang 已提交
106
    import platform
107
    from . import core
Y
Yang Yu 已提交
108

X
Xin Pan 已提交
109 110
    in_test = 'unittest' in sys.modules

Y
Yang Yu 已提交
111 112 113 114 115 116 117 118 119 120 121 122 123 124 125
    try:
        num_threads = int(os.getenv('OMP_NUM_THREADS', '1'))
    except ValueError:
        num_threads = 1

    if num_threads > 1:
        print(
            'WARNING: OMP_NUM_THREADS set to {0}, not 1. The computation '
            'speed will not be optimized if you use data parallel. It will '
            'fail if this PaddlePaddle binary is compiled with OpenBlas since'
            ' OpenBlas does not support multi-threads.'.format(num_threads),
            file=sys.stderr)
        print('PLEASE USE OMP_NUM_THREADS WISELY.', file=sys.stderr)

    os.environ['OMP_NUM_THREADS'] = str(num_threads)
J
JiabinYang 已提交
126
    sysstr = platform.system()
127
    read_env_flags = [
128
        'check_nan_inf', 'benchmark', 'eager_delete_scope',
129 130
        'initial_cpu_memory_in_mb', 'init_allocated_mem', 'free_idle_memory',
        'paddle_num_threads', "dist_threadpool_size", 'eager_delete_tensor_gb',
S
sneaxiy 已提交
131
        'fast_eager_deletion_mode', 'memory_fraction_of_eager_deletion',
S
fix bug  
sneaxiy 已提交
132
        'allocator_strategy', 'reader_queue_speed_test_mode',
133
        'print_sub_graph_dir', 'pe_profile_fname', 'warpctc_dir',
S
sneaxiy 已提交
134
        'inner_op_parallelism', 'enable_parallel_graph',
M
minqiyang 已提交
135 136
        'multiple_of_cupti_buffer_size', 'enable_subgraph_optimize',
        'tracer_profile_fname'
137
    ]
J
JiabinYang 已提交
138
    if 'Darwin' not in sysstr:
J
JiabinYang 已提交
139 140
        read_env_flags.append('use_pinned_memory')

P
peizhilin 已提交
141 142 143
    if os.name != 'nt':
        read_env_flags.append('cpu_deterministic')

144 145 146
    if core.is_compiled_with_mkldnn():
        read_env_flags.append('use_mkldnn')

147 148 149
    if core.is_compiled_with_ngraph():
        read_env_flags.append('use_ngraph')

Y
update  
Yancey1989 已提交
150 151
    if core.is_compiled_with_dist():
        read_env_flags.append('rpc_deadline')
Q
qiaolongfei 已提交
152
        read_env_flags.append('rpc_server_profile_path')
G
gongweibao 已提交
153
        read_env_flags.append('enable_rpc_profiler')
Q
Qiao Longfei 已提交
154 155 156
        read_env_flags.append('rpc_send_thread_num')
        read_env_flags.append('rpc_get_thread_num')
        read_env_flags.append('rpc_prefetch_thread_num')
157
        read_env_flags.append('rpc_disable_reuse_port')
158 159 160 161
        if core.is_compiled_with_brpc():
            read_env_flags.append('max_body_size')
            #set brpc max body size
            os.environ['FLAGS_max_body_size'] = "2147483647"
Y
update  
Yancey1989 已提交
162

163
    if core.is_compiled_with_cuda():
164
        read_env_flags += [
165
            'fraction_of_gpu_memory_to_use', 'cudnn_deterministic',
166
            'enable_cublas_tensor_op_math', 'conv_workspace_size_limit',
167
            'cudnn_exhaustive_search', 'memory_optimize_debug', 'selected_gpus',
168
            'sync_nccl_allreduce', 'limit_of_tmp_allocation',
D
dzhwinter 已提交
169 170
            'times_excess_than_required_tmp_allocation',
            'enable_inplace_whitelist'
171
        ]
Q
QI JUN 已提交
172 173
    core.init_gflags([sys.argv[0]] +
                     ["--tryfromenv=" + ",".join(read_env_flags)])
Y
Yang Yu 已提交
174
    core.init_glog(sys.argv[0])
X
Xin Pan 已提交
175 176
    # don't init_p2p when in unittest to save time.
    core.init_devices(not in_test)
D
dzhwinter 已提交
177

178

X
Xin Pan 已提交
179 180
# TODO(panyx0718): Avoid doing complex initialization logic in __init__.py.
# Consider paddle.init(args) or paddle.main(args)
X
Xin Pan 已提交
181
monkey_patch_variable()
Y
Yang Yu 已提交
182
__bootstrap__()