improve err msg for PolynomialDecay LR scheduler (#5143)

* improve err msg for PolynomialDecay LR scheduler

* Update CHANGELOG.md

Co-authored-by: Dirk Groeneveld <dirkg@allenai.org>

Co-authored-by: Dirk Groeneveld <dirkg@allenai.org>

CHANGELOG.md

@@ -37,6 +37,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - The `GradientDescentTrainer` no longer leaves stray model checkpoints around when it runs out of patience.
 - Fixed `cached_path()` for "hf://" files.
+- Improved the error message for the `PolynomialDecay` LR scheduler when `num_steps_per_epoch` is missing.
 
 
 ## [v2.3.1](https://github.com/allenai/allennlp/releases/tag/v2.3.1) - 2021-04-20

allennlp/training/learning_rate_schedulers/polynomial_decay.py

@@ -1,6 +1,7 @@
 from overrides import overrides
 import torch
 
+from allennlp.common.checks import ConfigurationError
 from allennlp.training.learning_rate_schedulers.learning_rate_scheduler import LearningRateScheduler
 
 
@@ -41,6 +42,18 @@ def __init__(
     ):
         super().__init__(optimizer, last_epoch)
 
+        # Sanity check here.
+        if num_steps_per_epoch is None:
+            raise ConfigurationError(
+                "'num_steps_per_epoch' is required for this LR scheduler.\n\n"
+                "If you know how many batches per epoch for your training data, you can set this value "
+                "directly in your config. Otherwise you'll need to use compatible settings with your data loader "
+                "so that it can report an accurate number of batches per epoch. "
+                "If you're using the MultiProcessDataLoader, "
+                "this means you either need to set 'batches_per_epoch' "
+                "or leave 'max_instances_in_memory' as None (if your entire dataset can fit into memory)."
+            )
+
         self.power = power
         self.warmup_steps = warmup_steps
         self.total_steps = num_epochs * num_steps_per_epoch