Replace pluggable components to create a proper Configurable hierarchy.

Summary:
This large diff rewrites a significant portion of Implicitron's config hierarchy. The new hierarchy, and some of the default implementation classes, are as follows:
```
Experiment
    data_source: ImplicitronDataSource
        dataset_map_provider
        data_loader_map_provider
    model_factory: ImplicitronModelFactory
        model: GenericModel
    optimizer_factory: ImplicitronOptimizerFactory
    training_loop: ImplicitronTrainingLoop
        evaluator: ImplicitronEvaluator
```

1) Experiment (used to be ExperimentConfig) is now a top-level Configurable and contains as members mainly (mostly new) high-level factory Configurables.
2) Experiment's job is to run factories, do some accelerate setup and then pass the results to the main training loop.
3) ImplicitronOptimizerFactory and ImplicitronModelFactory are new high-level factories that create the optimizer, scheduler, model, and stats objects.
4) TrainingLoop is a new configurable that runs the main training loop and the inner train-validate step.
5) Evaluator is a new configurable that TrainingLoop uses to run validation/test steps.
6) GenericModel is not the only model choice anymore. Instead, ImplicitronModelBase (by default instantiated with GenericModel) is a member of Experiment and can be easily replaced by a custom implementation by the user.

All the new Configurables are children of ReplaceableBase, and can be easily replaced with custom implementations.

In addition, I added support for the exponential LR schedule, updated the config files and the test, as well as added a config file that reproduces NERF results and a test to run the repro experiment.

Reviewed By: bottler

Differential Revision: D37723227

fbshipit-source-id: b36bee880d6aa53efdd2abfaae4489d8ab1e8a27
This commit is contained in:
Krzysztof Chalupka
2022-07-29 17:32:51 -07:00
committed by Facebook GitHub Bot
parent 6b481595f0
commit 1b0584f7bd
42 changed files with 2045 additions and 1478 deletions

View File

@@ -12,6 +12,7 @@ from hydra import compose, initialize_config_dir
from omegaconf import OmegaConf
from .. import experiment
from .utils import intercept_logs
def interactive_testing_requested() -> bool:
@@ -33,7 +34,10 @@ DEBUG: bool = False
# TODO:
# - add enough files to skateboard_first_5 that this works on RE.
# - share common code with PyTorch3D tests?
# - deal with the temporary output files this test creates
def _parse_float_from_log(line):
return float(line.split()[-1])
class TestExperiment(unittest.TestCase):
@@ -44,15 +48,18 @@ class TestExperiment(unittest.TestCase):
# Test making minimal changes to the dataclass defaults.
if not interactive_testing_requested() or not internal:
return
cfg = OmegaConf.structured(experiment.ExperimentConfig)
cfg.data_source_args.dataset_map_provider_class_type = (
# Manually override config values. Note that this is not necessary out-
# side of the tests!
cfg = OmegaConf.structured(experiment.Experiment)
cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_class_type = (
"JsonIndexDatasetMapProvider"
)
dataset_args = (
cfg.data_source_args.dataset_map_provider_JsonIndexDatasetMapProvider_args
cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_JsonIndexDatasetMapProvider_args
)
dataloader_args = (
cfg.data_source_args.data_loader_map_provider_SequenceDataLoaderMapProvider_args
cfg.data_source_ImplicitronDataSource_args.data_loader_map_provider_SequenceDataLoaderMapProvider_args
)
dataset_args.category = "skateboard"
dataset_args.test_restrict_sequence_id = 0
@@ -62,18 +69,80 @@ class TestExperiment(unittest.TestCase):
dataset_args.dataset_JsonIndexDataset_args.image_width = 80
dataloader_args.dataset_length_train = 1
dataloader_args.dataset_length_val = 1
cfg.solver_args.max_epochs = 2
cfg.training_loop_ImplicitronTrainingLoop_args.max_epochs = 2
cfg.training_loop_ImplicitronTrainingLoop_args.store_checkpoints = False
cfg.optimizer_factory_ImplicitronOptimizerFactory_args.multistep_lr_milestones = [
0,
1,
]
experiment.run_training(cfg)
if DEBUG:
experiment.dump_cfg(cfg)
with intercept_logs(
logger_name="projects.implicitron_trainer.impl.training_loop",
regexp="LR change!",
) as intercepted_logs:
experiment_runner = experiment.Experiment(**cfg)
experiment_runner.run()
# Make sure LR decreased on 0th and 1st epoch 10fold.
self.assertEqual(intercepted_logs[0].split()[-1], "5e-06")
def test_exponential_lr(self):
# Test making minimal changes to the dataclass defaults.
if not interactive_testing_requested():
return
cfg = OmegaConf.structured(experiment.Experiment)
cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_class_type = (
"JsonIndexDatasetMapProvider"
)
dataset_args = (
cfg.data_source_ImplicitronDataSource_args.dataset_map_provider_JsonIndexDatasetMapProvider_args
)
dataloader_args = (
cfg.data_source_ImplicitronDataSource_args.data_loader_map_provider_SequenceDataLoaderMapProvider_args
)
dataset_args.category = "skateboard"
dataset_args.test_restrict_sequence_id = 0
dataset_args.dataset_root = "manifold://co3d/tree/extracted"
dataset_args.dataset_JsonIndexDataset_args.limit_sequences_to = 5
dataset_args.dataset_JsonIndexDataset_args.image_height = 80
dataset_args.dataset_JsonIndexDataset_args.image_width = 80
dataloader_args.dataset_length_train = 1
dataloader_args.dataset_length_val = 1
cfg.training_loop_ImplicitronTrainingLoop_args.max_epochs = 2
cfg.training_loop_ImplicitronTrainingLoop_args.store_checkpoints = False
cfg.optimizer_factory_ImplicitronOptimizerFactory_args.lr_policy = "Exponential"
cfg.optimizer_factory_ImplicitronOptimizerFactory_args.exponential_lr_step_size = (
2
)
if DEBUG:
experiment.dump_cfg(cfg)
with intercept_logs(
logger_name="projects.implicitron_trainer.impl.training_loop",
regexp="LR change!",
) as intercepted_logs:
experiment_runner = experiment.Experiment(**cfg)
experiment_runner.run()
# Make sure we followed the exponential lr schedule with gamma=0.1,
# exponential_lr_step_size=2 -- so after two epochs, should
# decrease lr 10x to 5e-5.
self.assertEqual(intercepted_logs[0].split()[-1], "0.00015811388300841897")
self.assertEqual(intercepted_logs[1].split()[-1], "5e-05")
def test_yaml_contents(self):
cfg = OmegaConf.structured(experiment.ExperimentConfig)
# Check that the default config values, defined by Experiment and its
# members, is what we expect it to be.
cfg = OmegaConf.structured(experiment.Experiment)
yaml = OmegaConf.to_yaml(cfg, sort_keys=False)
if DEBUG:
(DATA_DIR / "experiment.yaml").write_text(yaml)
self.assertEqual(yaml, (DATA_DIR / "experiment.yaml").read_text())
def test_load_configs(self):
# Check that all the pre-prepared configs are valid.
config_files = []
for pattern in ("repro_singleseq*.yaml", "repro_multiseq*.yaml"):
@@ -89,3 +158,17 @@ class TestExperiment(unittest.TestCase):
with self.subTest(file.name):
with initialize_config_dir(config_dir=str(IMPLICITRON_CONFIGS_DIR)):
compose(file.name)
class TestNerfRepro(unittest.TestCase):
@unittest.skip("This test reproduces full NERF training.")
def test_nerf_blender(self):
# Train vanilla NERF.
# Set env vars BLENDER_DATASET_ROOT and BLENDER_SINGLESEQ_CLASS first!
if not interactive_testing_requested():
return
with initialize_config_dir(config_dir=str(IMPLICITRON_CONFIGS_DIR)):
cfg = compose(config_name="repro_singleseq_nerf_blender", overrides=[])
experiment_runner = experiment.Experiment(**cfg)
experiment.dump_cfg(cfg)
experiment_runner.run()