support num_sanity_val_steps=-1 (#2246)

* support sanity_val_step=-1

* fix list size

* simplification

* simplify

* add test for num_sanity_val_steps=-1

* update test

* update docs

* extend tests to multiple dataloaders

* changelog

* Update tests/trainer/test_trainer.py

Co-authored-by: Jirka Borovec <[email protected]>

* improve test

* refactor the sanity check decision

* fix merge

* Update trainer.py

Co-authored-by: Jirka Borovec <[email protected]>
Co-authored-by: William Falcon <[email protected]>

Author: 3 people

CHANGELOG.md

@@ -11,6 +11,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Added BLEU metrics ([#2535](https://github.com/PyTorchLightning/pytorch-lightning/pull/2535))
 
+- Added support for `Trainer(num_sanity_val_steps=-1)` to check all validation data before training ([#2246](https://github.com/PyTorchLightning/pytorch-lightning/pull/2246))
+
 ### Changed
 
 

docs/source/debugging.rst

@@ -129,4 +129,4 @@ argument of :class:`~pytorch_lightning.trainer.trainer.Trainer`)
 .. testcode::
 
     # DEFAULT
-    trainer = Trainer(num_sanity_val_steps=5)
+    trainer = Trainer(num_sanity_val_steps=2)

pytorch_lightning/callbacks/progress.py

@@ -102,7 +102,7 @@ def total_val_batches(self) -> int:
         total_val_batches = 0
         if trainer.fast_dev_run and trainer.val_dataloaders is not None:
             total_val_batches = len(trainer.val_dataloaders)
-        elif not self.trainer.disable_validation:
+        elif self.trainer.enable_validation:
             is_val_epoch = (trainer.current_epoch + 1) % trainer.check_val_every_n_epoch == 0
             total_val_batches = sum(trainer.num_val_batches) if is_val_epoch else 0
         return total_val_batches
@@ -302,7 +302,7 @@ def init_test_tqdm(self) -> tqdm:
     def on_sanity_check_start(self, trainer, pl_module):
         super().on_sanity_check_start(trainer, pl_module)
         self.val_progress_bar = self.init_sanity_tqdm()
-        self.val_progress_bar.total = trainer.num_sanity_val_steps * len(trainer.val_dataloaders)
+        self.val_progress_bar.total = convert_inf(trainer.num_sanity_val_steps * len(trainer.val_dataloaders))
         self.main_progress_bar = tqdm(disable=True)  # dummy progress bar
 
     def on_sanity_check_end(self, trainer, pl_module):

pytorch_lightning/trainer/__init__.py

@@ -603,16 +603,19 @@ def on_train_end(self, trainer, pl_module):
 
 Sanity check runs n batches of val before starting the training routine.
 This catches any bugs in your validation without having to wait for the first validation check.
-The Trainer uses 5 steps by default. Turn it off or modify it here.
+The Trainer uses 2 steps by default. Turn it off or modify it here.
 
 .. testcode::
 
     # default used by the Trainer
-    trainer = Trainer(num_sanity_val_steps=5)
+    trainer = Trainer(num_sanity_val_steps=2)
 
     # turn it off
     trainer = Trainer(num_sanity_val_steps=0)
 
+    # check all validation data
+    trainer = Trainer(num_sanity_val_steps=-1)
+
 num_tpu_cores
 ^^^^^^^^^^^^^
 .. warning:: .. deprecated:: 0.7.6

pytorch_lightning/trainer/evaluation_loop.py

@@ -72,15 +72,16 @@
 -----------------------------------------
 
 Lightning runs a few steps of validation in the beginning of training.
- This avoids crashing in the validation loop sometime deep into a lengthy training loop.
+This avoids crashing in the validation loop sometime deep into a lengthy training loop.
 
 .. code-block:: python
 
     # DEFAULT
-    trainer = Trainer(num_sanity_val_steps=5)
+    trainer = Trainer(num_sanity_val_steps=2)
 
 
-You can use `Trainer(num_sanity_val_steps=0)` to skip the sanity check.
+You can use `Trainer(num_sanity_val_steps=0)` to skip the sanity check or `Trainer(num_sanity_val_steps=-1)`
+to check all the validation data.
 
 # Testing loop
 

pytorch_lightning/trainer/trainer.py

@@ -336,7 +336,8 @@ def __init__(
 
             amp_level: The optimization level to use (O1, O2, etc...).
 
-            num_sanity_val_steps: Sanity check runs n batches of val before starting the training routine.
+            num_sanity_val_steps: Sanity check runs n validation batches before starting the training routine.
+                Set it to `-1` to run all batches in all validation dataloaders. Default: 2
 
             truncated_bptt_steps: Truncated back prop breaks performs backprop every k steps of
 
@@ -408,7 +409,6 @@ def __init__(
         # training state
         self.model = None
         self.testing = False
-        self.disable_validation = False
         self.prepare_data_per_node = prepare_data_per_node
         self.lr_schedulers = []
         self.optimizers = None
@@ -488,7 +488,7 @@ def __init__(
         self.max_steps = max_steps
         self.min_steps = min_steps
 
-        self.num_sanity_val_steps = num_sanity_val_steps
+        self.num_sanity_val_steps = float("inf") if num_sanity_val_steps == -1 else num_sanity_val_steps
         # Backward compatibility, TODO: remove in v0.9.0
         if print_nan_grads:
             rank_zero_warn("Argument `print_nan_grads` has no effect and will be removed in v0.9.0."
@@ -883,6 +883,17 @@ def progress_bar_dict(self) -> dict:
         ref_model = self.model if not self.data_parallel else self.model.module
         return dict(**ref_model.get_progress_bar_dict(), **self.progress_bar_metrics)
 
+    @property
+    def disable_validation(self) -> bool:
+        """ Check if validation is disabled during training. """
+        return not self.enable_validation
+
+    @property
+    def enable_validation(self) -> bool:
+        """ Check if we should run validation during training. """
+        val_loop_enabled = (self.is_overridden('validation_step') and self.limit_val_batches > 0)
+        return val_loop_enabled or self.fast_dev_run
+
     # -----------------------------
     # MODEL TRAINING
     # -----------------------------
@@ -1186,10 +1197,6 @@ def run_pretrain_routine(self, model: LightningModule):
 
             return eval_loop_results
 
-        # check if we should run validation during training
-        self.disable_validation = not (self.is_overridden('validation_step') and self.limit_val_batches > 0) \
-            and not self.fast_dev_run
-
         # run a few val batches before training starts
         self._run_sanity_check(ref_model, model)
 
@@ -1204,9 +1211,12 @@ def run_pretrain_routine(self, model: LightningModule):
         self.train()
 
     def _run_sanity_check(self, ref_model, model):
+        should_sanity_check = self.is_overridden('validation_step') and self.num_sanity_val_steps > 0 \
+            and self.limit_val_batches > 0
+
         # run tiny validation (if validation defined)
         # to make sure program won't crash during val
-        if not self.disable_validation and self.num_sanity_val_steps > 0:
+        if should_sanity_check:
             self.reset_val_dataloader(ref_model)
 
             # hook and callback

pytorch_lightning/trainer/training_loop.py

@@ -651,7 +651,7 @@ def should_check_val(self, batch_idx, is_last_batch):
         # decide if we should run validation
         is_val_check_batch = (batch_idx + 1) % self.val_check_batch == 0
         can_check_epoch = (self.current_epoch + 1) % self.check_val_every_n_epoch == 0
-        can_check_val = not self.disable_validation and can_check_epoch
+        can_check_val = self.enable_validation and can_check_epoch
         should_check_val = is_val_check_batch or self.should_stop
         is_last_batch_for_infinite_dataset = (is_last_batch and self.val_check_batch == float('inf'))
         should_check_val = can_check_val and (should_check_val or is_last_batch_for_infinite_dataset)

tests/trainer/test_trainer.py

@@ -6,6 +6,7 @@
 import types
 from argparse import Namespace
 from pathlib import Path
+from unittest.mock import patch
 
 import cloudpickle
 import pytest
@@ -807,6 +808,34 @@ def test_tpu_choice(tmpdir, tpu_cores, expected_tpu_id, error_expected):
         assert trainer.tpu_id == expected_tpu_id
 
 
+@pytest.mark.parametrize(['limit_val_batches'], [
+    pytest.param(0.0),  # this should run no sanity checks
+    pytest.param(1),
+    pytest.param(1.0),
+    pytest.param(0.3),
+])
+def test_num_sanity_val_steps(tmpdir, limit_val_batches):
+    """
+    Test that num_sanity_val_steps=-1 runs through all validation data once.
+    Makes sure this setting is independent of limit_val_batches.
+    """
+    model = EvalModelTemplate()
+    model.validation_step = model.validation_step__multiple_dataloaders
+    model.validation_epoch_end = model.validation_epoch_end__multiple_dataloaders
+    trainer = Trainer(
+        default_root_dir=tmpdir,
+        num_sanity_val_steps=-1,
+        limit_val_batches=limit_val_batches,  # should have no influence
+        max_steps=1,
+    )
+    assert trainer.num_sanity_val_steps == float('inf')
+    val_dataloaders = model.val_dataloader__multiple()
+
+    with patch.object(trainer, 'evaluation_forward', wraps=trainer.evaluation_forward) as mocked:
+        trainer.fit(model, val_dataloaders=val_dataloaders)
+        assert mocked.call_count == sum(len(dl) * (limit_val_batches > 0) for dl in val_dataloaders)
+
+
 @pytest.mark.parametrize("trainer_kwargs,expected", [
     pytest.param(
         dict(distributed_backend=None, gpus=None),