logger manager (PaddlePaddle#45909) (PaddlePaddle#46087)

uniform logger manager in FleetAPI.
hidde API under distributed/utils which users don't need.

python/paddle/distributed/__init__.py

@@ -65,7 +65,6 @@
 from paddle.fluid.dygraph.parallel import ParallelEnv  # noqa: F401
 
 from . import cloud_utils  # noqa: F401
-from . import utils  # noqa: F401
 
 from .sharding import *  # noqa: F401
 

python/paddle/distributed/auto_parallel/converter.py

@@ -16,7 +16,7 @@
 import warnings
 import logging
 import numpy as np
-from .utils import get_logger
+from ..utils.log_utils import get_logger
 
 
 class Converter(object):

python/paddle/distributed/auto_parallel/dist_saver.py

@@ -27,7 +27,7 @@
 from .utils import get_dist_attr
 from .converter import Converter
 from .process_group import _g_process_group_map
-from .utils import get_logger
+from ..utils.log_utils import get_logger
 
 
 def check_filename(re_exp, filename):

python/paddle/distributed/auto_parallel/parallelizer.py

@@ -24,7 +24,7 @@
 import time
 import paddle
 from paddle.fluid.backward import append_backward
-from paddle.distributed.utils import get_logger
+from paddle.distributed.utils.log_utils import get_logger
 from paddle.distributed.fleet import cloud_utils
 import paddle.fluid.core as core
 from paddle.fluid import program_guard

python/paddle/distributed/cloud_utils.py

@@ -14,10 +14,8 @@
 
 import os
 import paddle
-from paddle.distributed.utils import get_cluster
-from paddle.distributed.utils import logger
-from paddle.distributed.utils import get_gpus
-from paddle.distributed.utils import get_cluster_from_args
+from paddle.distributed.utils.launch_utils import get_cluster, get_gpus, get_cluster_from_args
+from paddle.distributed.utils.launch_utils import logger
 
 __all__ = []
 

python/paddle/distributed/fleet/__init__.py

@@ -32,6 +32,7 @@
 from .model import distributed_model
 from .optimizer import distributed_optimizer
 from .scaler import distributed_scaler
+from .utils import log_util
 
 __all__ = [  #noqa
     "CommunicateTopology", "UtilBase", "HybridCommunicateGroup",
@@ -90,5 +91,7 @@
 shrink = fleet.shrink
 get_hybrid_communicate_group = fleet.get_hybrid_communicate_group
 distributed_scaler = distributed_scaler
-
+set_log_level = log_util.set_log_level
+get_log_level_code = log_util.get_log_level_code
+get_log_level_name = log_util.get_log_level_name
 from .. import auto_parallel as auto

python/paddle/distributed/fleet/elastic/manager.py

@@ -26,13 +26,9 @@
 from paddle.distributed.fleet import cloud_utils
 from paddle.distributed.fleet import launch_utils
 
-logger = logging.getLogger("ELASTIC")
-logger.setLevel(logging.INFO)
-formatter = logging.Formatter(
-    fmt='%(name)s %(levelname)s %(asctime)s %(message)s')
-ch = logging.StreamHandler()
-ch.setFormatter(formatter)
-logger.addHandler(ch)
+from paddle.distributed.utils.log_utils import get_logger
+
+logger = get_logger("INFO", "ELASTIC")
 
 ELASTIC_EXIT_CODE = 101
 ELASTIC_AUTO_PARALLEL_EXIT_CODE = 102
@@ -354,7 +350,7 @@ def pre_hook(self):
                                     stderr=subprocess.PIPE,
                                     shell=True).communicate()
         if err:
-            logger.warn("pre_hook exec failed")
+            logger.warning("pre_hook exec failed")
         else:
             logger.info(f"pre_hook exec result: {out.decode('utf-8').strip()}")
 

python/paddle/distributed/fleet/fleet.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import copy
-import warnings
 import paddle
 import os
 from types import MethodType
@@ -32,6 +31,8 @@
 from .meta_parallel import model_parallel_random_seed
 from paddle import _C_ops, _legacy_C_ops
 from paddle.fluid import core
+from .utils.log_util import logger, set_log_level
+import logging
 
 __all__ = []
 
@@ -54,7 +55,7 @@ def apply_ir_passes(main_program, startup_program, config):
         # RawProgramOptimizer also inserts coalesce_tensor
         # into program. These two procedures may conflict
         # in which vars are to be fused.
-        warnings.warn(
+        logger.warning(
             'Currently, the fuse_all_optimizer_ops pass has conflict with fuse_all_reduce_ops pass. Disable the fuse_all_optimizer_ops pass temporarily.'
         )
         build_strategy.fuse_all_optimizer_ops = False
@@ -83,7 +84,7 @@ def __impl__(*args, **kwargs):
 
         if cls._role_maker is not None and cls._role_maker._is_non_distributed(
         ) is True:
-            warnings.warn(
+            logger.warning(
                 "%s() function doesn't work when use non_distributed fleet." %
                 (func.__name__))
             return
@@ -165,7 +166,11 @@ def __init__(self):
         self._context = {}
         self.user_defined_optimizer = paddle.optimizer.Optimizer(0.0)
 
-    def init(self, role_maker=None, is_collective=False, strategy=None):
+    def init(self,
+             role_maker=None,
+             is_collective=False,
+             strategy=None,
+             log_level="INFO"):
         """
         Initialize role_maker in Fleet.
 
@@ -182,6 +187,8 @@ def init(self, role_maker=None, is_collective=False, strategy=None):
                 GPU.The default value is False.The default value is False.
             strategy (DistributedStrategy): Extra properties for distributed training. 
                 For details, please refer to paddle.distributed.fleet.DistributedStrategy. Default: None.
+            log_level (Integer, String, optional): A ``Integer`` or ``String`` Variable determining how hight
+                the logging level is. Default is "INFO".
 
 
         Returns:
@@ -217,7 +224,18 @@ def init(self, role_maker=None, is_collective=False, strategy=None):
                 strategy = fleet.DistributedStrategy()
                 fleet.init(strategy=strategy)
 
+        Examples5:
+
+            .. code-block:: python
+
+                import paddle.distributed.fleet as fleet
+                strategy = fleet.DistributedStrategy()
+                fleet.init(log_level = "DEBUG")
+
         """
+
+        set_log_level(log_level)
+
         if strategy is None:
             strategy = DistributedStrategy()
         self._user_defined_strategy = copy.deepcopy(strategy)
@@ -261,12 +279,12 @@ def init(self, role_maker=None, is_collective=False, strategy=None):
                 self._hcg = tp.HybridCommunicateGroup(self._topology)
                 return
             if parallel_helper._is_parallel_ctx_initialized():
-                warnings.warn(
+                logger.warning(
                     "The dygraph parallel environment has been initialized.")
             else:
                 # FLAGS_nccl_nrings is used for dynamic graph multi-stream communication
                 if "FLAGS_nccl_nrings" in os.environ:
-                    warnings.warn(
+                    logger.warning(
                         "You have set the environment variable FLAGS_nccl_nrings "
                         "outside the program, so the nccl_comm_num in "
                         "DistributedStrategy will not take effect here.")
@@ -281,7 +299,7 @@ def init(self, role_maker=None, is_collective=False, strategy=None):
                 if tp._HYBRID_PARALLEL_GROUP is None:
                     self._init_hybrid_parallel_env()
                 else:
-                    warnings.warn(
+                    logger.warning(
                         "The dygraph hybrid parallel environment has been initialized."
                     )
         elif self._is_collective:
@@ -850,9 +868,6 @@ def save_inference_model(self,
                 fleet.init_server()
 
         """
-        # warnings.warn(
-        #     "'save_inference_model' is a deprecated, will be deleted after v2.2.0, Please use fleet.save instead."
-        # )
 
         self._runtime_handle._save_inference_model(executor, dirname,
                                                    feeded_var_names,
@@ -902,10 +917,6 @@ def save_persistables(self, executor, dirname, main_program=None, mode=0):
                 fleet.save_persistables(exe, "dirname", paddle.static.default_main_program())
 
         """
-        # warnings.warn(
-        #     "'save_persistables' is a deprecated, will be deleted after v2.2.0, Please use fleet.save instead."
-        # )
-
         self._runtime_handle._save_persistables(executor, dirname, main_program,
                                                 mode)
 
@@ -1015,7 +1026,7 @@ def distributed_optimizer(self, optimizer, strategy=None):
 
         if strategy is not None:
             if self._is_collective:
-                warnings.warn(
+                logger.warning(
                     "It is recommended to use DistributedStrategy "
                     "in fleet.init(). The strategy here is only for compatibility. "
                     "If the strategy in fleet.distributed_optimizer() is "
@@ -1304,8 +1315,9 @@ def _minimize_impl(self,
             copy_user_defined_strategy, can_not_apply_optimizer_list)
 
         context["valid_strategy"] = copy.deepcopy(valid_strategy)
-        # print("valid_strategy:", context["valid_strategy"])
-        # print("user_defined_strategy:", context["user_defined_strategy"])
+        logger.debug("valid_strategy: " + str(context["valid_strategy"]))
+        logger.debug("user_defined_strategy: " +
+                     str(context["user_defined_strategy"]))
 
         applied_meta_list = self.strategy_compiler._get_applied_meta_list()
         applied_graph_list = self.strategy_compiler._get_applied_graph_list()
@@ -1335,17 +1347,19 @@ def _minimize_impl(self,
                                                         no_grad_set=no_grad_set)
 
         if meta_optimizer:
-            # print("before minimize program id:", id(loss.block.program))
+            logger.debug("before minimize program id: " +
+                         str(id(loss.block.program)))
             optimize_ops, params_grads = meta_optimizer.minimize(
                 loss, startup_program, parameter_list, no_grad_set=no_grad_set)
-            # print("after minimize program id:", id(loss.block.program))
-
+            logger.debug("after minimize program id: " +
+                         str(id(loss.block.program)))
             default_program = paddle.static.default_main_program()
-            # print("default program id:", id(default_program))
+            logger.debug("default program id: " + str(id(default_program)))
 
             if id(default_program) != id(loss.block.program):
                 paddle.fluid.framework.switch_main_program(loss.block.program)
-            # print("default program id after switch:", id(default_program))
+            logger.debug("default program id after switch: " +
+                         str(id(default_program)))
 
         else:
             optimize_ops, params_grads = self.user_defined_optimizer.minimize(
@@ -1355,7 +1369,8 @@ def _minimize_impl(self,
         context["program_params_grads"] = params_grads
 
         if graph_optimizer:
-            # print("before graph minimize program id:", id(loss.block.program))
+            logger.debug("before graph minimize program id: " +
+                         str(id(loss.block.program)))
             optimize_ops, params_grads = graph_optimizer.minimize(
                 loss, startup_program, parameter_list, no_grad_set=no_grad_set)
             # since we do not encourage users to use graph operations
@@ -1454,7 +1469,8 @@ def _minimize_losses_impl(self,
                 if v or k not in opt_info:
                     opt_info[k] = v
             program._fleet_opt = opt_info
-            # print("fleet base opt info:", id(program), program._fleet_opt)
+            logger.debug("fleet base opt info: " + str(id(program)) +
+                         str(program._fleet_opt))
 
         if self._runtime_handle is None:
             self._runtime_handle = RuntimeFactory()._create_runtime(context)

python/paddle/distributed/fleet/meta_optimizers/sharding_optimizer.py

@@ -30,15 +30,8 @@
 from .sharding import utils
 # FIXME: import *
 from .sharding.utils import *
-
 import logging
-
-logger = logging.getLogger(__name__)
-formatter = logging.Formatter(fmt='%(asctime)s %(levelname)-8s %(message)s',
-                              datefmt='%Y-%m-%d %H:%M:%S')
-ch = logging.StreamHandler()
-ch.setFormatter(formatter)
-logger.addHandler(ch)
+from ..utils.log_util import logger
 
 __all__ = []
 

python/paddle/distributed/fleet/meta_parallel/sharding/group_sharded_stage2.py

@@ -33,7 +33,7 @@
 import paddle
 from paddle import nn
 from paddle.distributed import collective
-from paddle.distributed.utils import get_logger
+from paddle.distributed.utils.log_utils import get_logger
 
 from .group_sharded_storage import GradStorage
 from .group_sharded_optimizer_stage2 import GroupShardedOptimizerStage2

python/paddle/distributed/fleet/optimizer.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import copy
-import warnings
 import paddle
 import os
 import numpy as np
@@ -22,6 +21,7 @@
 from .meta_optimizers import HybridParallelOptimizer, HeterParallelOptimizer
 from paddle.fluid import core
 from paddle.distributed import fleet
+from .utils.log_util import logger
 
 
 def _dygraph_distributed_optimizer(optimizer, strategy=None):
@@ -52,7 +52,7 @@ def _dygraph_distributed_optimizer(optimizer, strategy=None):
 
     if strategy is not None:
         if fleet_env._is_collective:
-            warnings.warn(
+            logger.warning(
                 "It is recommended to use DistributedStrategy "
                 "in fleet_env.init(). The strategy here is only for compatibility. "
                 "If the strategy in fleet_env.distributed_optimizer() is "

python/paddle/distributed/fleet/recompute/recompute.py

@@ -22,13 +22,7 @@
 from paddle.fluid.framework import in_dygraph_mode
 
 import logging
-
-logger = logging.getLogger(__name__)
-formatter = logging.Formatter(fmt='%(asctime)s %(levelname)-8s %(message)s',
-                              datefmt='%Y-%m-%d %H:%M:%S')
-ch = logging.StreamHandler()
-ch.setFormatter(formatter)
-logger.addHandler(ch)
+from ..utils.log_util import logger
 
 __all__ = []
 
@@ -49,7 +43,7 @@ def detach_variable(inputs):
 def check_recompute_necessary(inputs):
     if not any(input_.stop_gradient == False for input_ in inputs
                if isinstance(input_, (core.eager.Tensor, paddle.Tensor))):
-        logger.warn(
+        logger.warning(
             "[Recompute]: None of the inputs to current recompute block need grad, "
             "therefore there is NO need to recompute this block in backward !")
 

python/paddle/distributed/fleet/utils/hybrid_parallel_util.py

@@ -14,7 +14,6 @@
 import os
 import six
 import numpy as np
-import warnings
 
 from paddle import framework
 import paddle