优化部分warning显示

fastNLP/core/callbacks/has_monitor_callback.py

@@ -78,11 +78,11 @@ class MonitorUtility:
             return monitor_value
         # 第一次运行
         if isinstance(self.monitor, str) and self._real_monitor == self.monitor and use_monitor != self.monitor:
-            logger.warning(f"We can not find `{self.monitor}` in the evaluation result (with keys as {list(results.keys())}), "
-                           f"we use the `{use_monitor}` as the monitor for `{self.__class__.__name__}`.")
+            logger.rank_zero_warning(f"We can not find `{self.monitor}` in the evaluation result (with keys as "
+                                     f"{list(results.keys())}), we use the `{use_monitor}` as the monitor.", once=True)
         # 检测到此次和上次不同。
         elif isinstance(self.monitor, str) and self._real_monitor != self.monitor and use_monitor != self._real_monitor:
-            logger.warning(f"Change of monitor detected for `{self.__class__.__name__}`. "
+            logger.rank_zero_warning(f"Change of monitor detected for `{self.__class__.__name__}`. "
                            f"The expected monitor is:`{self.monitor}`, last used monitor is:"
                            f"`{self._real_monitor}` and current monitor is:`{use_monitor}`. Please consider using a "
                            f"customized monitor function when the evaluation results are varying between validation.")

fastNLP/core/callbacks/topk_saver.py

@@ -32,7 +32,7 @@ class Saver:
         :param kwargs: 更多需要传递给 Trainer.save() 或者 Trainer.save_model() 接口的参数。
         """
         if folder is None:
-            logger.warning(
+            logger.rank_zero_warning(
                 "Parameter `folder` is None, and we will use the current work directory to find and load your model.")
             folder = Path.cwd()
         folder = Path(folder)

fastNLP/core/controllers/evaluator.py

@@ -221,7 +221,7 @@ class Evaluator:
     @evaluate_batch_loop.setter
     def evaluate_batch_loop(self, loop: Loop):
         if self.evaluate_batch_step_fn is not None:
-            logger.warning("`evaluate_batch_step_fn` was customized in the Evaluator initialization, it will be ignored "
+            logger.rank_zero_warning("`evaluate_batch_step_fn` was customized in the Evaluator initialization, it will be ignored "
                            "when the `evaluate_batch_loop` is also customized.")
         self._evaluate_batch_loop = loop
 

fastNLP/core/controllers/trainer.py

@@ -305,7 +305,7 @@ class Trainer(TrainerEventTrigger):
         else:
             if self.driver.is_distributed():
                 if catch_KeyboardInterrupt:
-                    logger.warning("Parameter `catch_KeyboardInterrupt` can only be False when you are using multi-device "
+                    logger.rank_zero_warning("Parameter `catch_KeyboardInterrupt` can only be False when you are using multi-device "
                                    "driver. And we are gonna to set it to False.")
                 catch_KeyboardInterrupt = False
 
@@ -535,7 +535,7 @@ class Trainer(TrainerEventTrigger):
                     _not_called_callback_fns.append(each_callback_fn)
 
         if check_mode:
-            logger.warning("You have customized your 'batch_step_fn' in the 'train_batch_loop' and also use these "
+            logger.rank_zero_warning("You have customized your 'batch_step_fn' in the 'train_batch_loop' and also use these "
                            f"callback_fns: {_not_called_callback_fns}, but it seems that"
                            "you don't call the corresponding callback hook explicitly in your 'batch_step_fn'.")
             # 对于 'batch_step_fn' 来讲，其只需要在第一次的 step 后进行检测即可，因此在第一次检测后将 check_batch_step_fn 置为 pass

fastNLP/core/drivers/paddle_driver/initialize_paddle_driver.py

@@ -69,7 +69,7 @@ def initialize_paddle_driver(driver: str, device: Optional[Union[str, int, List[
         if not isinstance(device, List):
             return PaddleSingleDriver(model, device, **kwargs)
         else:
-            logger.warning("Notice you are using `paddle` driver but your chosen `device` are multi gpus, we will use"
+            logger.rank_zero_warning("Notice you are using `paddle` driver but your chosen `device` are multi gpus, we will use"
                             "`Fleetriver` by default. But if you mean using `PaddleFleetDriver`, you should choose parameter"
                             "`driver` as `PaddleFleetDriver`.")
             return PaddleFleetDriver(model, device, **kwargs)
@@ -77,7 +77,7 @@ def initialize_paddle_driver(driver: str, device: Optional[Union[str, int, List[
         if not isinstance(device, List):
             if device == "cpu":
                 raise ValueError("You are using `fleet` driver, but your chosen `device` is 'cpu'.")
-            logger.warning("Notice you are using `fleet` driver, but your chosen `device` is only one gpu, we will"
+            logger.rank_zero_warning("Notice you are using `fleet` driver, but your chosen `device` is only one gpu, we will"
                             "still use `PaddleFleetDriver` for you, but if you mean using `PaddleSingleDriver`, you should "
                             "choose `paddle` driver.")
             return PaddleFleetDriver(model, [device], **kwargs)

fastNLP/core/drivers/paddle_driver/paddle_driver.py

@@ -72,7 +72,7 @@ class PaddleDriver(Driver):
         :param set_to_none: 用来判断是否需要将梯度直接置为 None；Paddle中这个参数无效。
         """
         if set_to_none:
-            logger.warning_once("Parameter `set_to_none` does nothing in paddle since grad cannot be set directly.")
+            logger.rank_zero_warning("Parameter `set_to_none` does nothing in paddle since grad cannot be set directly.")
         for optimizer in self.optimizers:
             optimizer.clear_grad()
 
@@ -256,7 +256,7 @@ class PaddleDriver(Driver):
                     if dataloader_args.batch_size is not None:
                         num_consumed_batches = num_consumed_batches * dataloader_args.batch_size
                     else:  # 有可能 batch_size 为 None，就只有损失精度了
-                        logger.warning("fastNLP cannot get batch_size, we have to save based on `num_consumed_samples`, "
+                        logger.rank_zero_warning("fastNLP cannot get batch_size, we have to save based on `num_consumed_samples`, "
                                      "it may cause missing some samples when reload.")
                         num_consumed_batches = sampler_states['num_consumed_samples']
                 sampler_states['num_consumed_samples'] = num_consumed_samples_array[num_consumed_batches]
@@ -266,7 +266,7 @@ class PaddleDriver(Driver):
                     sampler_states['num_consumed_samples'] = sampler.num_replicas * dataloader_args.batch_size \
                                                              * num_consumed_batches
                 else:
-                    logger.warning("fastNLP cannot get batch_size, we have to save based on `num_consumed_samples`, "
+                    logger.rank_zero_warning("fastNLP cannot get batch_size, we have to save based on `num_consumed_samples`, "
                                  "it may cause missing some samples when reload.")
         else:
             raise RuntimeError(
@@ -329,7 +329,7 @@ class PaddleDriver(Driver):
             self.grad_scaler.load_state_dict(grad_scaler_state_dict)
             logger.debug("Load grad_scaler state dict...")
         elif not isinstance(self.grad_scaler, DummyGradScaler):
-            logger.warning(f"Checkpoint {folder} is not trained with fp16=True, while resume to a fp16=True training, "
+            logger.rank_zero_warning(f"Checkpoint {folder} is not trained with fp16=True, while resume to a fp16=True training, "
                            f"the training process may be unstable.")
 
         # 4. 恢复 sampler 的状态；

fastNLP/core/drivers/paddle_driver/utils.py

@@ -51,7 +51,7 @@ def paddle_seed_everything(seed: Optional[int] = None, workers: bool = False) ->
         seed = int(seed)
 
     if not (min_seed_value <= seed <= max_seed_value):
-        logger.warning("Your seed value is two big or two small for numpy, we will choose a random seed for "
+        logger.rank_zero_warning("Your seed value is two big or two small for numpy, we will choose a random seed for "
                         "you.")
 
         # rank_zero_warn(f"{seed} is not in bounds, numpy accepts from {min_seed_value} to {max_seed_value}")

fastNLP/core/drivers/torch_driver/torch_driver.py

@@ -197,7 +197,7 @@ class TorchDriver(Driver):
                     if dataloader_args.batch_size is not None:
                         num_consumed_batches = num_consumed_batches * dataloader_args.batch_size
                     else:  # 有可能 batch_size 为 None，就只有损失精度了
-                        logger.warning("fastNLP cannot get batch_size, we have to save based on `num_consumed_samples`, "
+                        logger.rank_zero_warning("fastNLP cannot get batch_size, we have to save based on `num_consumed_samples`, "
                                      "it may cause missing some samples when reload.")
                         num_consumed_batches = sampler_states['num_consumed_samples']
                 sampler_states['num_consumed_samples'] = num_consumed_samples_array[num_consumed_batches]
@@ -207,7 +207,7 @@ class TorchDriver(Driver):
                     sampler_states['num_consumed_samples'] = sampler.num_replicas * dataloader_args.batch_size \
                                                              * num_consumed_batches
                 else:
-                    logger.warning("fastNLP cannot get batch_size, we have to save based on `num_consumed_samples`, "
+                    logger.rank_zero_warning("fastNLP cannot get batch_size, we have to save based on `num_consumed_samples`, "
                                  "it may cause missing some samples when reload.")
 
             states['sampler_states'] = sampler_states

fastNLP/core/drivers/torch_driver/utils.py

@@ -60,7 +60,7 @@ def torch_seed_everything(seed: Optional[int] = None, workers: bool = False) ->
         seed = int(seed)
 
     if not (min_seed_value <= seed <= max_seed_value):
-        logger.warning("Your seed value is two big or two small for numpy, we will choose a random seed for you.")
+        logger.rank_zero_warning("Your seed value is two big or two small for numpy, we will choose a random seed for you.")
 
         seed = _select_seed_randomly(min_seed_value, max_seed_value)
 
@@ -162,7 +162,7 @@ def _build_fp16_env(dummy=False):
         if not torch.cuda.is_available():
             raise RuntimeError("No cuda")
         if torch.cuda.get_device_capability(0)[0] < 7:
-            logger.warning(
+            logger.rank_zero_warning(
                 "NOTE: your device does NOT support faster training with fp16, "
                 "please switch to FP32 which is likely to be faster"
             )

fastNLP/core/log/logger.py

@@ -124,18 +124,21 @@ class FastNLPLogger(logging.Logger, metaclass=LoggerSingleton):
                 self._log(WARNING, msg, args, **kwargs)
             self._warning_msgs.add(msg)
 
-    def rank_zero_warning(self, msg, *args, **kwargs):
+    def rank_zero_warning(self, msg, *args, once=False, **kwargs):
         """
         只在 rank 0 上 warning 。
 
         :param msg:
         :param args:
+        :param once: 是否只 warning 一次
         :param kwargs:
         :return:
         """
         if os.environ.get(FASTNLP_GLOBAL_RANK, '0') == '0':
+            if once and msg in self._warning_msgs:
+                return
             if self.isEnabledFor(WARNING):
-                # kwargs = self._add_rank_info(kwargs)
+                kwargs = self._add_rank_info(kwargs)
                 self._log(WARNING, msg, args, **kwargs)
 
     def warn(self, msg, *args, **kwargs):

fastNLP/core/utils/cache_results.py

@@ -15,6 +15,7 @@ __all__ = [
 
 from fastNLP.core.log.logger import logger
 from fastNLP.core.log.highlighter import ColorHighlighter
+from .utils import _get_fun_msg
 
 
 class FuncCallVisitor(ast.NodeVisitor):
@@ -306,7 +307,7 @@ def cache_results(_cache_fp, _hash_param=True, _refresh=False, _verbose=1, _chec
                     if verbose == 1:
                         logger.info("Read cache from {} (Saved on {}).".format(cache_filepath, save_time))
                     if check_hash and old_hash_code != new_hash_code:
-                        logger.warning(f"The function `{func.__name__}` is different from its last cache (Save on {save_time}). The "
+                        logger.warning(f"The function {_get_fun_msg(func)} is different from its last cache (Save on {save_time}). The "
                                       f"difference may caused by the sourcecode change.",
                                        extra={'highlighter': ColorHighlighter('red')})
                     refresh_flag = False

fastNLP/core/utils/utils.py

@@ -239,7 +239,7 @@ def check_user_specific_params(user_params: Dict, fn: Callable):
     fn_arg_names = get_fn_arg_names(fn)
     for arg_name, arg_value in user_params.items():
         if arg_name not in fn_arg_names:
-            logger.warning(f"Notice your specific parameter `{arg_name}` is not used by function `{fn.__name__}`.")
+            logger.rank_zero_warning(f"Notice your specific parameter `{arg_name}` is not used by function `{fn.__name__}`.")
     return user_params