Terrance/tensorf (#218)

* added tensorf, updated requirements for windows

* cleaning up imports

* tensorf changes

* initial working version of tensorf

* added callback for feature upsampling

* modifying tests to match changes

* reverting requirements changes

* reverting unnecessary changes

* reverting unnecessary changes

* no more upsampling on step 0

* removed density field, fixed error msg

* modified callbacks and upsampling, reinitializing optimzer in training loop

* deleting tensorf_field file"

* checkpointing progress on restructuring callbacks

* updated callback logic for before and after train capabilities

* fixing upsampling stepping

* merging

* adding tensorf to config, fixing some bugs with new callback

* changed the way callbacks detected steps argument in callback func

* fixing typing

* adding pylint disable no self use to get_training_callbacks

* fixing config

* changing callback in vanilla nerf to allow for no density field

* cleanup

* removing config yaml

Co-authored-by: Ethan Weber <eweb0124@gmail.com>
This commit is contained in:
Terrance WangandEthan Weber authored and GitHub committed 2022-08-31 23:29:13 -07:00
1 parent 6ee6934581
commit 8670f19d2d
13 files changed
+394 -52

No files matched your search

+18 -12
View File
@@ -29,11 +29,14 @@ from torch.cuda.amp.grad_scaler import GradScaler
from torch.nn.parallel import DistributedDataParallel as DDP
from nerfactory.configs import base as cfg
from nerfactory.dataloaders.structs import DatasetInputs
from nerfactory.optimizers.optimizers import Optimizers, setup_optimizers
from nerfactory.pipelines.base import Pipeline
from nerfactory.utils import profiler, writer
from nerfactory.utils.callbacks import Callback
from nerfactory.utils.callbacks import (
TrainingCallback,
TrainingCallbackAttributes,
TrainingCallbackLocation,
)
from nerfactory.utils.decorators import check_main_thread
from nerfactory.utils.writer import EventName, TimeWriter
from nerfactory.viewer.server import viewer_utils
@@ -76,8 +79,6 @@ class Trainer:
if self.device == "cpu":
self.mixed_precision = False
logging.warning("Mixed precision is disabled for CPU training.")
# dataset variables
self.dataset_inputs_train: DatasetInputs
# model variables
self.pipeline: Pipeline
self.optimizers: Optimizers
@@ -89,7 +90,7 @@ class Trainer:
banner_messages = [f"Viewer at: {self.visualizer_state.viewer_url}"]
self.grad_scaler = GradScaler(enabled=self.mixed_precision)
# training callbacks
self.callbacks: List[Callback]
self.callbacks: List[TrainingCallback]
# logging variables
writer.setup_event_writers(
config.logging, max_iter=config.trainer.max_num_iterations, banner_messages=banner_messages
@@ -113,8 +114,9 @@ class Trainer:
)
dist.barrier(device_ids=[self.local_rank])
# TODO(ethan): do this for pipeline, not pipeline.model
self.callbacks = self.pipeline.model.get_training_callbacks()
self.callbacks = self.pipeline.get_training_callbacks(
TrainingCallbackAttributes(optimizers=self.optimizers, grad_scaler=self.grad_scaler, pipeline=self.pipeline)
)
def train(self) -> None:
"""Train the model."""
@@ -126,10 +128,17 @@ class Trainer:
num_iterations = self.config.trainer.max_num_iterations
for step in range(self.start_step, self.start_step + num_iterations):
# Note: if visualizer used, the rendering of the visualizer will be included in the iteration train time
# if the visualizer used, the rendering of the visualizer will be included in the iteration train time
with TimeWriter(writer, EventName.ITER_TRAIN_TIME, step=step) as t:
# training callbacks before the training iteration
for callback in self.callbacks:
callback.run_callback_at_location(
step, location=TrainingCallbackLocation.BEFORE_TRAIN_ITERATION
)
loss_metric_dict = self.train_iteration(step)
with TimeWriter(writer, EventName.ITER_VIS_TIME, step=step) as t:
with TimeWriter(writer, EventName.ITER_VIS_TIME, step=step) as _:
self.visualizer_state.update_scene(step, self.pipeline.model)
train_num_rays_per_batch = self.config.pipeline.dataloader.train_num_rays_per_batch
@@ -219,10 +228,7 @@ class Trainer:
self.grad_scaler.scale(loss).backward() # type: ignore
self.optimizers.optimizer_scaler_step_all(self.grad_scaler)
self.grad_scaler.update()
self.optimizers.scheduler_step_all(step)
for callback in self.callbacks:
callback.after_step(step)
# Merging loss and metrics dict into a single output.
loss_dict["loss"] = loss