Added accumulation of loggers' metrics for the same steps (#1278)

* `add_argparse_args` method fixed (argument types added)

* autopep8 fixes

* --gpus=0 removed from test (for ci tests)

* Update pytorch_lightning/trainer/trainer.py

Co-Authored-By: Joe Davison <joe@huggingface.co>

* test_with_accumulate_grad_batches added

* agg_and_log_metrics logic added to the base logger class

* small format fix

* agg metrics strategies removed (not to complicate stuff)

* agg metrics: handle zero step

* autopep8

* changelog upd

* flake fix

* metrics aggregators factored out, metrics_agg.py added + tests

* metrics agg default value added

* Update pytorch_lightning/loggers/metrics_agg.py

Co-Authored-By: Jirka Borovec <Borda@users.noreply.github.com>

* metrics aggregators factored out, metrics_agg.py added + tests

* metrics agg default value added

* Update pytorch_lightning/loggers/metrics_agg.py

Co-Authored-By: Jirka Borovec <Borda@users.noreply.github.com>

* remove .item which causes sync issues (#1254)

* remove .item which causes sync issues

* fixed gradient acc sched

* fixed gradient acc sched

* test_metrics_agg.py removed (all tested in doctrings), agg metrics refactored

* test_metrics_agg.py removed (all tested in doctrings), agg metrics refactored

* autopep8

* loggers base.py types fixed

* test

* test

* metrics aggregation for loggers: each key now has a specific function (or default one)

* metrics aggregation for loggers: each key now has a specific function (or default one)

* docstrings upd

* manual typehints removed from docstrings

* batch_size decreased for test `test_with_accumulate_grad_batches`

* extend running accum

* refactor

* fix tests

* fix tests

* allowed_types generator scoped

* trainer.py distutils was imported twice, fixed

* TensorRunningAccum refactored

* TensorRunningAccum added to change log (Changed)

* change log pull link added

Co-authored-by: Joe Davison <joe@huggingface.co>
Co-authored-by: Jirka Borovec <Borda@users.noreply.github.com>
Co-authored-by: William Falcon <waf2107@columbia.edu>
Co-authored-by: J. Borovec <jirka.borovec@seznam.cz>

CHANGELOG.md

@@ -8,6 +8,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Added
 
+- Added same step loggers' metrics aggregation ([#1278](https://github.com/PyTorchLightning/pytorch-lightning/pull/1278))
 - Added parity test between a vanilla MNIST model and lightning model ([#1284](https://github.com/PyTorchLightning/pytorch-lightning/pull/1284))
 - Added parity test between a vanilla RNN model and lightning model ([#1351](https://github.com/PyTorchLightning/pytorch-lightning/pull/1351))
 - Added Reinforcement Learning - Deep Q-network (DQN) lightning example ([#1232](https://github.com/PyTorchLightning/pytorch-lightning/pull/1232))
@@ -30,6 +31,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Changed
 
+- Changed (renamed and refatored) `TensorRunningMean` -> `TensorRunningAccum`: running accumulations were generalized. ([#1278](https://github.com/PyTorchLightning/pytorch-lightning/pull/1278))
 - Changed `progress_bar_refresh_rate` trainer flag to disable progress bar when set to 0. ([#1108](https://github.com/PyTorchLightning/pytorch-lightning/pull/1108))
 - Enhanced `load_from_checkpoint` to also forward params to the model ([#1307](https://github.com/PyTorchLightning/pytorch-lightning/pull/1307))
 - Updated references to self.forward() to instead use the `__call__` interface. ([#1211](https://github.com/PyTorchLightning/pytorch-lightning/pull/1211))

pytorch_lightning/loggers/base.py

@@ -1,9 +1,12 @@
 import argparse
+import functools
+import operator
 from abc import ABC, abstractmethod
 from argparse import Namespace
 from functools import wraps
-from typing import Union, Optional, Dict, Iterable, Any, Callable, List
+from typing import Union, Optional, Dict, Iterable, Any, Callable, List, Sequence, Mapping, Tuple
 
+import numpy as np
 import torch
 
 
@@ -25,22 +28,119 @@ def wrapped_fn(self, *args, **kwargs):
 class LightningLoggerBase(ABC):
     """Base class for experiment loggers."""
 
-    def __init__(self):
+    def __init__(
+            self,
+            agg_key_funcs: Optional[Mapping[str, Callable[[Sequence[float]], float]]] = None,
+            agg_default_func: Callable[[Sequence[float]], float] = np.mean
+    ):
+        """
+        Args:
+            agg_key_funcs:
+                Dictionary which maps a metric name to a function, which will
+                aggregate the metric values for the same steps.
+            agg_default_func:
+                Default function to aggregate metric values. If some metric name
+                is not presented in the `agg_key_funcs` dictionary, then the
+                `agg_default_func` will be used for aggregation.
+
+        Notes:
+            `agg_key_funcs` and `agg_default_func` are used only when one logs metrics with
+            `LightningLoggerBase.agg_and_log_metrics` method.
+        """
         self._rank = 0
+        self._prev_step = -1
+        self._metrics_to_agg: List[Dict[str, float]] = []
+        self._agg_key_funcs = agg_key_funcs if agg_key_funcs else {}
+        self._agg_default_func = agg_default_func
+
+    def update_agg_funcs(
+            self,
+            agg_key_funcs: Optional[Mapping[str, Callable[[Sequence[float]], float]]] = None,
+            agg_default_func: Callable[[Sequence[float]], float] = np.mean
+    ):
+        """Update aggregation methods.
+
+        Args:
+            agg_key_funcs:
+                Dictionary which maps a metric name to a function, which will
+                aggregate the metric values for the same steps.
+            agg_default_func:
+                Default function to aggregate metric values. If some metric name
+                is not presented in the `agg_key_funcs` dictionary, then the
+                `agg_default_func` will be used for aggregation.
+        """
+        if agg_key_funcs:
+            self._agg_key_funcs.update(agg_key_funcs)
+        if agg_default_func:
+            self._agg_default_func = agg_default_func
 
     @property
     @abstractmethod
     def experiment(self) -> Any:
         """Return the experiment object associated with this logger"""
 
+    def _aggregate_metrics(
+            self, metrics: Dict[str, float], step: Optional[int] = None
+    ) -> Tuple[int, Optional[Dict[str, float]]]:
+        """Aggregates metrics.
+
+        Args:
+            metrics: Dictionary with metric names as keys and measured quantities as values
+            step: Step number at which the metrics should be recorded
+
+        Returns:
+            sStep and aggregated metrics. The return value could be None. In such case, metrics
+            are added to the aggregation list, but not aggregated yet.
+        """
+        # if you still receiving metric from the same step, just accumulate it
+        if step == self._prev_step:
+            self._metrics_to_agg.append(metrics)
+            return step, None
+
+        # compute the metrics
+        agg_step, agg_mets = self._finalize_agg_metrics()
+
+        # as new step received reset accumulator
+        self._metrics_to_agg = [metrics]
+        self._prev_step = step
+        return agg_step, agg_mets
+
+    def _finalize_agg_metrics(self):
+        """Aggregate accumulated metrics. This shall be called in close."""
+        # compute the metrics
+        if not self._metrics_to_agg:
+            agg_mets = None
+        elif len(self._metrics_to_agg) == 1:
+            agg_mets = self._metrics_to_agg[0]
+        else:
+            agg_mets = merge_dicts(self._metrics_to_agg, self._agg_key_funcs, self._agg_default_func)
+        return self._prev_step, agg_mets
+
+    def agg_and_log_metrics(self, metrics: Dict[str, float], step: Optional[int] = None):
+        """Aggregates and records metrics.
+        This method doesn't log the passed metrics instantaneously, but instead
+        it aggregates them and logs only if metrics are ready to be logged.
+
+        Args:
+            metrics: Dictionary with metric names as keys and measured quantities as values
+            step: Step number at which the metrics should be recorded
+        """
+        agg_step, metrics_to_log = self._aggregate_metrics(metrics=metrics, step=step)
+
+        if metrics_to_log is not None:
+            self.log_metrics(metrics=metrics_to_log, step=agg_step)
+
     @abstractmethod
     def log_metrics(self, metrics: Dict[str, float], step: Optional[int] = None):
-        """Record metrics.
+        """Records metrics.
+        This method logs metrics as as soon as it received them. If you want to aggregate
+        metrics for one specific `step`, use the `agg_and_log_metrics` method.
 
         Args:
             metrics: Dictionary with metric names as keys and measured quantities as values
             step: Step number at which the metrics should be recorded
         """
+        pass
 
     @staticmethod
     def _convert_params(params: Union[Dict[str, Any], Namespace]) -> Dict[str, Any]:
@@ -131,7 +231,10 @@ def finalize(self, status: str) -> None:
 
     def close(self) -> None:
         """Do any cleanup that is necessary to close an experiment."""
-        pass
+        agg_step, metrics_to_log = self._finalize_agg_metrics()
+
+        if metrics_to_log is not None:
+            self.log_metrics(metrics=metrics_to_log, step=agg_step)
 
     @property
     def rank(self) -> int:
@@ -200,3 +303,48 @@ def name(self) -> str:
     @property
     def version(self) -> str:
         return '_'.join([str(logger.version) for logger in self._logger_iterable])
+
+
+def merge_dicts(
+        dicts: Sequence[Mapping],
+        agg_key_funcs: Optional[Mapping[str, Callable[[Sequence[float]], float]]] = None,
+        default_func: Callable[[Sequence[float]], float] = np.mean
+) -> Dict:
+    """Merge a sequence with dictionaries into one dictionary by aggregating the
+    same keys with some given function.
+
+    Args:
+        dicts:
+            Sequence of dictionaries to be merged.
+        agg_key_funcs:
+            Mapping from key name to function. This function will aggregate a
+            list of values, obtained from the same key of all dictionaries.
+            If some key has no specified aggregation function, the default one
+            will be used. Default is: None (all keys will be aggregated by the
+            default function).
+        default_func:
+            Default function to aggregate keys, which are not presented in the
+            `agg_key_funcs` map.
+
+    Returns:
+        Dictionary with merged values.
+
+    Examples:
+        >>> import pprint
+        >>> d1 = {'a': 1.7, 'b': 2.0, 'c': 1}
+        >>> d2 = {'a': 1.1, 'b': 2.2, 'v': 1}
+        >>> d3 = {'a': 1.1, 'v': 2.3}
+        >>> dflt_func = min
+        >>> agg_funcs = {'a': np.mean, 'v': max}
+        >>> pprint.pprint(merge_dicts([d1, d2, d3], agg_funcs, dflt_func))
+        {'a': 1.3, 'b': 2.0, 'c': 1, 'v': 2.3}
+    """
+
+    keys = list(functools.reduce(operator.or_, [set(d.keys()) for d in dicts]))
+    d_out = {}
+    for k in keys:
+        fn = agg_key_funcs.get(k, default_func) if agg_key_funcs else default_func
+        agg_val = fn([v for v in [d_in.get(k) for d_in in dicts] if v is not None])
+        d_out[k] = agg_val
+
+    return d_out

pytorch_lightning/trainer/logging.py

@@ -71,7 +71,7 @@ def log_metrics(self, metrics, grad_norm_dic, step=None):
             step = step if step is not None else self.global_step
         # log actual metrics
         if self.proc_rank == 0 and self.logger is not None:
-            self.logger.log_metrics(scalar_metrics, step=step)
+            self.logger.agg_and_log_metrics(scalar_metrics, step=step)
             self.logger.save()
 
     def add_tqdm_metrics(self, metrics):

pytorch_lightning/trainer/supporters.py

@@ -1,13 +1,12 @@
 import torch
 
 
-class TensorRunningMean(object):
-    """
-    Tracks a running mean without graph references.
-    Round robbin for the mean
+class TensorRunningAccum(object):
+    """Tracks a running accumulation values (min, max, mean) without graph
+    references.
 
     Examples:
-        >>> accum = TensorRunningMean(5)
+        >>> accum = TensorRunningAccum(5)
         >>> accum.last(), accum.mean()
         (None, None)
         >>> accum.append(torch.tensor(1.5))
@@ -18,8 +17,8 @@ class TensorRunningMean(object):
         (tensor(2.5000), tensor(2.))
         >>> accum.reset()
         >>> _= [accum.append(torch.tensor(i)) for i in range(13)]
-        >>> accum.last(), accum.mean()
-        (tensor(12.), tensor(10.))
+        >>> accum.last(), accum.mean(), accum.min(), accum.max()
+        (tensor(12.), tensor(10.), tensor(8.), tensor(12.))
     """
     def __init__(self, window_length: int):
         self.window_length = window_length
@@ -29,13 +28,16 @@ def __init__(self, window_length: int):
         self.rotated: bool = False
 
     def reset(self) -> None:
-        self = TensorRunningMean(self.window_length)
+        """Empty the accumulator."""
+        self = TensorRunningAccum(self.window_length)
 
     def last(self):
+        """Get the last added element."""
         if self.last_idx is not None:
             return self.memory[self.last_idx]
 
     def append(self, x):
+        """Add an element to the accumulator."""
         # ensure same device and type
         if self.memory.device != x.device or self.memory.type() != x.type():
             x = x.to(self.memory)
@@ -54,5 +56,20 @@ def append(self, x):
             self.rotated = True
 
     def mean(self):
+        """Get mean value from stored elements."""
+        return self._agg_memory('mean')
+
+    def max(self):
+        """Get maximal value from stored elements."""
+        return self._agg_memory('max')
+
+    def min(self):
+        """Get minimal value from stored elements."""
+        return self._agg_memory('min')
+
+    def _agg_memory(self, how: str):
         if self.last_idx is not None:
-            return self.memory.mean() if self.rotated else self.memory[:self.current_idx].mean()
+            if self.rotated:
+                return getattr(self.memory, how)()
+            else:
+                return getattr(self.memory[:self.current_idx], how)()

pytorch_lightning/trainer/trainer.py

@@ -28,7 +28,7 @@
 from pytorch_lightning.trainer.logging import TrainerLoggingMixin
 from pytorch_lightning.trainer.model_hooks import TrainerModelHooksMixin
 from pytorch_lightning.trainer.optimizers import TrainerOptimizersMixin
-from pytorch_lightning.trainer.supporters import TensorRunningMean
+from pytorch_lightning.trainer.supporters import TensorRunningAccum
 from pytorch_lightning.trainer.training_io import TrainerIOMixin
 from pytorch_lightning.trainer.training_loop import TrainerTrainLoopMixin
 from pytorch_lightning.trainer.training_tricks import TrainerTrainingTricksMixin
@@ -342,7 +342,7 @@ def __init__(
 
         # training bookeeping
         self.total_batch_idx = 0
-        self.running_loss = TensorRunningMean(window_length=20)
+        self.running_loss = TensorRunningAccum(window_length=20)
         self.batch_idx = 0
         self.tqdm_metrics = {}
         self.callback_metrics = {}
@@ -551,20 +551,19 @@ def add_argparse_args(cls, parent_parser: ArgumentParser) -> ArgumentParser:
 
         allowed_types = (str, float, int, bool)
         # TODO: get "help" from docstring :)
-        for arg, arg_types, arg_default in cls.get_init_arguments_and_types():
-            if arg not in depr_arg_names:
-                for allowed_type in allowed_types:
-                    if allowed_type in arg_types:
-                        if allowed_type is bool:
-                            allowed_type = lambda x: bool(distutils.util.strtobool(x))
-                        parser.add_argument(
-                            f'--{arg}',
-                            default=arg_default,
-                            type=allowed_type,
-                            dest=arg,
-                            help='autogenerated by pl.Trainer'
-                        )
-                        break
+        for arg, arg_types, arg_default in (at for at in cls.get_init_arguments_and_types()
+                                            if at[0] not in depr_arg_names):
+            for allowed_type in (at for at in allowed_types if at in arg_types):
+                if isinstance(allowed_type, bool):
+                    allowed_type = lambda x: bool(distutils.util.strtobool(x))
+                parser.add_argument(
+                    f'--{arg}',
+                    default=arg_default,
+                    type=allowed_type,
+                    dest=arg,
+                    help='autogenerated by pl.Trainer'
+                )
+                break
 
         return parser
 

pytorch_lightning/trainer/training_loop.py

@@ -147,7 +147,7 @@ def training_step(self, batch, batch_idx):
 from pytorch_lightning.loggers import LightningLoggerBase
 from pytorch_lightning.overrides.data_parallel import LightningDistributedDataParallel, LightningDataParallel
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
-from pytorch_lightning.trainer.supporters import TensorRunningMean
+from pytorch_lightning.trainer.supporters import TensorRunningAccum
 
 try:
     from apex import amp
@@ -337,7 +337,7 @@ def train(self):
                 self.accumulation_scheduler.on_epoch_start(self, self.get_model())
 
                 # stores accumulated grad fractions per batch
-                self.batch_loss_value = TensorRunningMean(
+                self.batch_loss_value = TensorRunningAccum(
                     window_length=self.accumulate_grad_batches
                 )