Spaces:
Running
on
Zero
Running
on
Zero
File size: 6,435 Bytes
9b33fca |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 |
# pylint: disable=duplicate-code
"""RetinaNet COCO 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,
LossConnector,
)
from vis4d.engine.loss_module import LossModule
from vis4d.eval.coco import COCODetectEvaluator
from vis4d.model.detect.retinanet import RetinaNet
from vis4d.op.box.encoder import DeltaXYWHBBoxEncoder
from vis4d.op.detect.retinanet import (
RetinaNetHeadLoss,
get_default_anchor_generator,
)
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_VIS,
CONN_BOX_LOSS_2D,
CONN_IMAGES_TEST,
CONN_IMAGES_TRAIN,
)
from vis4d.zoo.base.datasets.coco import (
CONN_COCO_BBOX_EVAL,
get_coco_detection_cfg,
)
def get_config() -> ExperimentConfig:
"""Returns the RetinaNet config dict for the coco detection task.
This is an example that shows how to set up a training experiment for the
COCO detection task.
Note that the high level params are exposed in the config. This allows
to easily change them from the command line.
E.g.:
>>> python -m vis4d.engine.run fit --config vis4d/zoo/retinanet/retinanet_rcnn_coco.py --config.num_epochs 100 --config.params.lr 0.001
Returns:
ExperimentConfig: The configuration
"""
######################################################
## General Config ##
######################################################
config = get_default_cfg(exp_name="retinanet_r50_fpn_coco")
# High level hyper parameters
params = ExperimentParameters()
params.samples_per_gpu = 2
params.workers_per_gpu = 2
params.lr = 0.01
params.num_epochs = 12
params.num_classes = 80
config.params = params
######################################################
## Datasets with augmentations ##
######################################################
data_root = "data/coco"
train_split = "train2017"
test_split = "val2017"
data_backend = class_config(HDF5Backend)
config.data = get_coco_detection_cfg(
data_root=data_root,
train_split=train_split,
test_split=test_split,
data_backend=data_backend,
samples_per_gpu=params.samples_per_gpu,
workers_per_gpu=params.workers_per_gpu,
)
######################################################
## MODEL & LOSS ##
######################################################
config.model = class_config(
RetinaNet,
num_classes=params.num_classes,
# weights="mmdet",
)
box_encoder = class_config(
DeltaXYWHBBoxEncoder,
target_means=(0.0, 0.0, 0.0, 0.0),
target_stds=(1.0, 1.0, 1.0, 1.0),
)
anchor_generator = class_config(get_default_anchor_generator)
retina_loss = class_config(
RetinaNetHeadLoss,
box_encoder=box_encoder,
anchor_generator=anchor_generator,
)
config.loss = class_config(
LossModule,
losses={
"loss": retina_loss,
"connector": class_config(
LossConnector, key_mapping=CONN_BOX_LOSS_2D
),
},
)
######################################################
## 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=[8, 11], gamma=0.1),
),
],
)
]
######################################################
## DATA CONNECTOR ##
######################################################
config.train_data_connector = class_config(
DataConnector,
key_mapping=CONN_IMAGES_TRAIN,
)
config.test_data_connector = class_config(
DataConnector,
key_mapping=CONN_IMAGES_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(
COCODetectEvaluator,
data_root=data_root,
split=test_split,
),
metrics_to_eval=["Det"],
test_connector=class_config(
CallbackConnector,
key_mapping=CONN_COCO_BBOX_EVAL,
),
)
)
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()
|