def _evaluate(cfg: CN,
dataset: Datasets,
output_folder: Path,
find_mistakes: bool = False,
include_heading: bool = False) -> str:
mistakes = 0
total = 0
folder = URI("data://render") / dataset.value
for img_file in folder.glob("*.png"):
total += 1
img = cv2.imread(str(img_file))
json_file = folder / f"{img_file.stem}.json"
with json_file.open("r") as f:
label = json.load(f)
actual = np.array(label["corners"])
try:
predicted = find_corners(cfg, img)
except Exception:
predicted = None
if predicted is not None:
actual = sort_corner_points(actual)
predicted = sort_corner_points(predicted)
if predicted is None or np.linalg.norm(actual - predicted,
axis=-1).max() > 10.:
mistakes += 1
return mistakes, total
def train_classifier(name: str):
"""Set up CLI interface for training a classifier.
Args:
name (str): the name of the classifier (`"occupancy_classifier"` or `"piece_classifier"`)
"""
configs_dir = URI("config://") / name
def _train(config: str):
cfg = CN.load_yaml_with_base(configs_dir / f"{config}.yaml")
run_dir = URI("runs://") / name / config
# Train the model and save it
train(cfg, run_dir)
# Read available configs
configs = [x.stem for x in configs_dir.glob("*.yaml")
if not x.stem.startswith("_")]
# Set up argument parser
parser = argparse.ArgumentParser(description="Train the network.")
parser.add_argument("--config", help="the configuration to train (default: all)",
type=str, choices=configs, default=None)
args = parser.parse_args()
# Train
if args.config is None:
logger.info("Training all configurations one by one")
for config in configs:
_train(config)
else:
logger.info(f"Training the {args.config} configuration")
_train(args.config)
def create_configs(classifier: str, include_centercrop: bool = False):
"""Create the YAML configuration files for all registered models for a classifier.
Args:
classifier (str): the classifier (either `"occupancy_classifier"` or `"piece_classifier"`)
include_centercrop (bool, optional): whether to create two configs per model, one including center crop and one not. Defaults to False.
"""
config_dir = URI("config://") / classifier
logger.info(f"Removing YAML files from {config_dir}.")
for f in config_dir.glob("*.yaml"):
if not f.name.startswith("_"):
f.unlink()
for name, model in MODELS_REGISTRY[classifier.upper()].items():
for center_crop in ({True, False} if include_centercrop else {False}):
config_file = config_dir / \
(name + ("_centercrop" if center_crop else "") + ".yaml")
logging.info(f"Writing configuration file {config_file}")
size = model.input_size
C = CN()
override_base = f"config://{classifier}/_base_override_{name}.yaml"
if URI(override_base).exists():
C._BASE_ = override_base
else:
suffix = "_pretrained" if model.pretrained else ""
C._BASE_ = f"config://{classifier}/_base{suffix}.yaml"
C.DATASET = CN()
C.DATASET.TRANSFORMS = CN()
C.DATASET.TRANSFORMS.CENTER_CROP = (50, 50) \
if center_crop else None
C.DATASET.TRANSFORMS.RESIZE = size
C.TRAINING = CN()
C.TRAINING.MODEL = CN()
C.TRAINING.MODEL.REGISTRY = classifier.upper()
C.TRAINING.MODEL.NAME = name
with config_file.open("w") as f:
C.dump(stream=f)
import numpy as np
from logging import getLogger
from recap import URI
import argparse
logger = getLogger(__name__)
if __name__ == "__main__":
argparse.ArgumentParser(
description="Split the dataset into train/val/test.").parse_args()
val_split = .03
test_split = .1
render_dir = URI("data://render")
ids = np.array([x.stem for x in render_dir.glob("*.json")])
if len(ids) == 0:
logger.warning(
"No samples found in 'data://render', either you did not download the datset yet or you have already split it."
)
np.random.seed(42)
ids = np.random.permutation(ids)
sample_sizes = (np.array([val_split, test_split]) * len(ids)).astype(
np.int32)
val, test, train = np.split(ids, sample_sizes)
datasets = {"val": val, "test": test, "train": train}
print(
f"{len(ids)} samples will be split into {len(train)} train, {len(val)} val, {len(test)} test."
)
for dataset_name, ids in datasets.items():