__init__.py 7.5 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 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33
import sys

# The legacy core need to be removed before "import core",
# in case of users installing paddlepadde without -U option
core_suffix = 'so'
if os.name == 'nt':
    core_suffix = 'pyd'

legacy_core = os.path.abspath(os.path.dirname(
    __file__)) + os.sep + 'core.' + core_suffix
if os.path.exists(legacy_core):
    sys.stderr.write('Deleting legacy file ' + legacy_core + '\n')
    try:
        os.remove(legacy_core)
    except Exception as e:
        raise e

34
# import all class inside framework into fluid module
35 36
from . import framework
from .framework import *
Y
Yang Yu 已提交
37
# import all class inside executor into fluid module
38 39
from . import executor
from .executor import *
W
Wang Guibao 已提交
40 41 42 43

from . import data_feed_desc
from .data_feed_desc import *

44 45 46
from . import dataset
from .dataset import *

H
Huihuang Zheng 已提交
47 48
from .data import *

49
from . import trainer_desc
50
from . import inferencer
H
Helin Wang 已提交
51

52 53 54 55
from . import io
from . import evaluator
from . import initializer
from . import layers
L
lujun 已提交
56
from . import dygraph
57 58 59 60
from . import contrib
from . import nets
from . import optimizer
from . import backward
61
from .backward import gradients
62 63 64 65
from . import regularizer
from . import average
from . import metrics
from . import transpiler
66
from . import incubate
67
from .input import embedding, one_hot
Q
Qiao Longfei 已提交
68
from . import distribute_lookup_table
69 70
from .param_attr import ParamAttr, WeightNormParamAttr
from .data_feeder import DataFeeder
S
sneaxiy 已提交
71
from .core import LoDTensor, LoDTensorArray, CPUPlace, CUDAPlace, CUDAPinnedPlace, Scope, _Scope
72 73
from .incubate import fleet
from .incubate import data_generator
74
from .transpiler import DistributeTranspiler, \
G
gongweibao 已提交
75
    memory_optimize, release_memory, DistributeTranspilerConfig
76 77
from .lod_tensor import create_lod_tensor, create_random_int_lodtensor
from . import clip
78
from . import dygraph_grad_clip
79 80 81 82
from . import profiler
from . import unique_name
from . import parallel_executor
from .parallel_executor import *
X
Xin Pan 已提交
83 84
from . import compiler
from .compiler import *
X
Xin Pan 已提交
85
from paddle.fluid.layers.math_op_patch import monkey_patch_variable
J
Jiabin Yang 已提交
86
from . import install_check
87 88
from .dygraph.nn import *
from .dygraph.layers import *
89 90

Tensor = LoDTensor
Y
Yang Yu 已提交
91

W
Wu Yi 已提交
92
__all__ = framework.__all__ + executor.__all__ + \
93
    trainer_desc.__all__ + inferencer.__all__ + transpiler.__all__ + \
W
Wang Guibao 已提交
94
    parallel_executor.__all__ + lod_tensor.__all__ + \
95
    data_feed_desc.__all__ + compiler.__all__ + backward.__all__ + [
96 97
        'io',
        'initializer',
98 99
        'embedding',
        'one_hot',
100 101
        'layers',
        'contrib',
H
Huihuang Zheng 已提交
102
        'data',
L
lujun 已提交
103
        'dygraph',
104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119
        'transpiler',
        'nets',
        'optimizer',
        'learning_rate_decay',
        'backward',
        'regularizer',
        'LoDTensor',
        'LoDTensorArray',
        'CPUPlace',
        'CUDAPlace',
        'CUDAPinnedPlace',
        'Tensor',
        'ParamAttr',
        'WeightNormParamAttr',
        'DataFeeder',
        'clip',
120
        'dygraph_grad_clip',
121 122 123
        'profiler',
        'unique_name',
        'Scope',
J
Jiabin Yang 已提交
124
        'install_check',
125
    ]
126 127


Y
Yang Yu 已提交
128
def __bootstrap__():
129 130
    """
    Enable reading gflags from environment variables.
Y
Yu Yang 已提交
131

132 133 134 135
    Returns:
        None
    """
    import sys
Y
Yang Yu 已提交
136
    import os
J
JiabinYang 已提交
137
    import platform
138
    from . import core
Y
Yang Yu 已提交
139

X
Xin Pan 已提交
140 141
    in_test = 'unittest' in sys.modules

Y
Yang Yu 已提交
142 143 144 145 146 147 148 149 150 151 152 153 154 155 156
    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
157
    sysstr = platform.system()
158
    read_env_flags = [
P
pkpk 已提交
159 160
        'check_nan_inf', 'fast_check_nan_inf', 'benchmark',
        'eager_delete_scope', 'initial_cpu_memory_in_mb', 'init_allocated_mem',
161 162 163 164 165 166 167
        'paddle_num_threads', 'dist_threadpool_size', 'eager_delete_tensor_gb',
        'fast_eager_deletion_mode', 'memory_fraction_of_eager_deletion',
        'allocator_strategy', 'reader_queue_speed_test_mode',
        'print_sub_graph_dir', 'pe_profile_fname', 'inner_op_parallelism',
        'enable_parallel_graph', 'fuse_parameter_groups_size',
        'multiple_of_cupti_buffer_size', 'fuse_parameter_memory_size',
        'tracer_profile_fname', 'dygraph_debug'
168
    ]
J
JiabinYang 已提交
169
    if 'Darwin' not in sysstr:
J
JiabinYang 已提交
170 171
        read_env_flags.append('use_pinned_memory')

P
peizhilin 已提交
172 173 174
    if os.name != 'nt':
        read_env_flags.append('cpu_deterministic')

175 176 177
    if core.is_compiled_with_mkldnn():
        read_env_flags.append('use_mkldnn')

178 179 180
    if core.is_compiled_with_ngraph():
        read_env_flags.append('use_ngraph')

Y
update  
Yancey1989 已提交
181
    if core.is_compiled_with_dist():
Q
Qiao Longfei 已提交
182
        #env for rpc
Y
update  
Yancey1989 已提交
183
        read_env_flags.append('rpc_deadline')
184
        read_env_flags.append('rpc_retry_times')
Q
qiaolongfei 已提交
185
        read_env_flags.append('rpc_server_profile_path')
G
gongweibao 已提交
186
        read_env_flags.append('enable_rpc_profiler')
Q
Qiao Longfei 已提交
187 188 189
        read_env_flags.append('rpc_send_thread_num')
        read_env_flags.append('rpc_get_thread_num')
        read_env_flags.append('rpc_prefetch_thread_num')
190
        read_env_flags.append('rpc_disable_reuse_port')
Q
Qiao Longfei 已提交
191

192 193
        read_env_flags.append('worker_update_interval_secs')

Q
Qiao Longfei 已提交
194
        # env for communicator
Q
Qiao Longfei 已提交
195 196
        read_env_flags.append('communicator_independent_recv_thread')
        read_env_flags.append('communicator_send_queue_size')
197
        read_env_flags.append('communicator_min_send_grad_num_before_recv')
Q
Qiao Longfei 已提交
198
        read_env_flags.append('communicator_thread_pool_size')
Q
Qiao Longfei 已提交
199
        read_env_flags.append('communicator_max_merge_var_num')
200
        read_env_flags.append('communicator_merge_sparse_bucket')
Q
Qiao Longfei 已提交
201
        read_env_flags.append('communicator_fake_rpc')
Q
Qiao Longfei 已提交
202
        read_env_flags.append('communicator_send_wait_times')
203
        read_env_flags.append('communicator_merge_sparse_grad')
204 205 206 207
        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
208

209
    if core.is_compiled_with_cuda():
210
        read_env_flags += [
211 212
            'fraction_of_gpu_memory_to_use', 'initial_gpu_memory_in_mb',
            'reallocate_gpu_memory_in_mb', 'cudnn_deterministic',
213
            'enable_cublas_tensor_op_math', 'conv_workspace_size_limit',
214
            'cudnn_exhaustive_search', 'selected_gpus', 'sync_nccl_allreduce',
215 216
            'cudnn_batchnorm_spatial_persistent', 'gpu_allocator_retry_time',
            'local_exe_sub_scope_limit'
217
        ]
Q
QI JUN 已提交
218 219
    core.init_gflags([sys.argv[0]] +
                     ["--tryfromenv=" + ",".join(read_env_flags)])
Y
Yang Yu 已提交
220
    core.init_glog(sys.argv[0])
X
Xin Pan 已提交
221 222
    # don't init_p2p when in unittest to save time.
    core.init_devices(not in_test)
D
dzhwinter 已提交
223

224

X
Xin Pan 已提交
225 226
# TODO(panyx0718): Avoid doing complex initialization logic in __init__.py.
# Consider paddle.init(args) or paddle.main(args)
X
Xin Pan 已提交
227
monkey_patch_variable()
Y
Yang Yu 已提交
228
__bootstrap__()