# pylint: disable=duplicate-code
"""Faster RCNN BDD100K training example."""
from __future__ import annotations
from torch.optim.lr_scheduler import LinearLR, MultiStepLR
from torch.optim.sgd import SGD
from vis4d.config import class_config
from vis4d.config.typing import ExperimentConfig, ExperimentParameters
from vis4d.data.io.hdf5 import HDF5Backend
from vis4d.engine.callbacks import EvaluatorCallback, VisualizerCallback
from vis4d.engine.connectors import CallbackConnector, DataConnector
from vis4d.eval.bdd100k import BDD100KDetectEvaluator
from vis4d.op.base import ResNet
from vis4d.vis.image import BoundingBoxVisualizer
from vis4d.zoo.base import (
get_default_callbacks_cfg,
get_default_cfg,
get_default_pl_trainer_cfg,
get_lr_scheduler_cfg,
get_optimizer_cfg,
)
from vis4d.zoo.base.data_connectors import (
CONN_BBOX_2D_TEST,
CONN_BBOX_2D_TRAIN,
CONN_BBOX_2D_VIS,
)
from vis4d.zoo.base.datasets.bdd100k import (
CONN_BDD100K_DET_EVAL,
get_bdd100k_detection_config,
)
from vis4d.zoo.base.models.faster_rcnn import get_faster_rcnn_cfg
[docs]
def get_config() -> ExperimentConfig:
"""Returns the Faster-RCNN config dict for the BDD100K detection task.
Returns:
ExperimentConfig: The configuration
"""
######################################################
## General Config ##
######################################################
config = get_default_cfg(exp_name="faster_rcnn_r50_3x_bdd100k")
# High level hyper parameters
params = ExperimentParameters()
params.samples_per_gpu = 2
params.workers_per_gpu = 2
params.lr = 0.02
params.num_epochs = 36
params.num_classes = 10
config.params = params
######################################################
## Datasets with augmentations ##
######################################################
data_root = "data/bdd100k/images/100k"
train_split = "train"
test_split = "val"
data_backend = class_config(HDF5Backend)
config.data = get_bdd100k_detection_config(
data_root=data_root,
train_split=train_split,
test_split=test_split,
data_backend=data_backend,
multi_scale=True,
samples_per_gpu=params.samples_per_gpu,
workers_per_gpu=params.workers_per_gpu,
)
######################################################
## MODEL & LOSS ##
######################################################
basemodel = class_config(
ResNet, resnet_name="resnet50", pretrained=True, trainable_layers=3
)
config.model, config.loss = get_faster_rcnn_cfg(
num_classes=params.num_classes, basemodel=basemodel
)
######################################################
## OPTIMIZERS ##
######################################################
config.optimizers = [
get_optimizer_cfg(
optimizer=class_config(
SGD, lr=params.lr, momentum=0.9, weight_decay=0.0001
),
lr_schedulers=[
get_lr_scheduler_cfg(
class_config(
LinearLR, start_factor=0.001, total_iters=500
),
end=500,
epoch_based=False,
),
get_lr_scheduler_cfg(
class_config(MultiStepLR, milestones=[24, 33], gamma=0.1),
),
],
)
]
######################################################
## DATA CONNECTOR ##
######################################################
config.train_data_connector = class_config(
DataConnector, key_mapping=CONN_BBOX_2D_TRAIN
)
config.test_data_connector = class_config(
DataConnector, key_mapping=CONN_BBOX_2D_TEST
)
######################################################
## CALLBACKS ##
######################################################
# Logger
callbacks = get_default_callbacks_cfg()
# Visualizer
callbacks.append(
class_config(
VisualizerCallback,
visualizer=class_config(BoundingBoxVisualizer, vis_freq=100),
output_dir=config.output_dir,
test_connector=class_config(
CallbackConnector, key_mapping=CONN_BBOX_2D_VIS
),
)
)
# Evaluator
callbacks.append(
class_config(
EvaluatorCallback,
evaluator=class_config(
BDD100KDetectEvaluator,
annotation_path="data/bdd100k/labels/det_20/det_val.json",
config_path="det",
),
test_connector=class_config(
CallbackConnector, key_mapping=CONN_BDD100K_DET_EVAL
),
metrics_to_eval=[BDD100KDetectEvaluator.METRICS_DET],
)
)
config.callbacks = callbacks
######################################################
## PL CLI ##
######################################################
# PL Trainer args
pl_trainer = get_default_pl_trainer_cfg(config)
pl_trainer.max_epochs = params.num_epochs
config.pl_trainer = pl_trainer
return config.value_mode()