OCT_Segmentation_a1 / src /inference /postprocess.py
farrell236's picture
Deploy native Gradio OCT inference app
8fca640 verified
Raw History Blame Contribute Delete
2.86 kB
"""Prediction-only native-grid filters; no reference arrays or masks are inputs."""
import numpy as np
from scipy.ndimage import gaussian_filter, median_filter
INTEGER_DEVICES = ["topcon_maestro2", "topcon_triton", "zeiss_cirrus"]
def transform(positions, device, height, configuration):
if "by_device" in configuration:
overrides = configuration["by_device"].get(device, {})
configuration = {
**{k: v for k, v in configuration.items() if k != "by_device"},
**overrides,
}
values = np.asarray(positions, dtype=np.float32)
if "boundary_filters" in configuration:
rules = configuration["boundary_filters"]
if (
not isinstance(rules, list)
or len(rules) != values.shape[0]
or not all(isinstance(rule, dict) for rule in rules)
):
raise ValueError(
"One fixed filter configuration is required for every output boundary"
)
common = {
key: value
for key, value in configuration.items()
if key != "boundary_filters"
}
return np.concatenate(
[
transform(values[index : index + 1], device, height, {**common, **rule})
for index, rule in enumerate(rules)
],
axis=0,
)
kind = configuration.get("filter", "none")
if kind == "gaussian":
values = gaussian_filter(
values,
sigma=(
0,
float(configuration.get("sigma_b", 0)),
float(configuration.get("sigma_a", 0)),
),
mode="nearest",
truncate=3,
)
elif kind == "median":
values = median_filter(
values,
size=(
1,
int(configuration.get("size_b", 1)),
int(configuration.get("size_a", 1)),
),
mode="nearest",
)
elif kind != "none":
raise ValueError("Unknown prediction filter")
if device in configuration.get("round_devices", []):
values = np.rint(values)
return np.clip(values, 0, height - 1).astype(np.float32)
CONFIGURATIONS = {
"round_integer_devices": {"round_devices": INTEGER_DEVICES},
"gaussian_b05": {"filter": "gaussian", "sigma_b": 0.5},
"gaussian_b1": {"filter": "gaussian", "sigma_b": 1.0},
"median_b3": {"filter": "median", "size_b": 3},
"gaussian_b05_round": {
"filter": "gaussian",
"sigma_b": 0.5,
"round_devices": INTEGER_DEVICES,
},
"gaussian_b1_round": {
"filter": "gaussian",
"sigma_b": 1.0,
"round_devices": INTEGER_DEVICES,
},
"median_b3_round": {
"filter": "median",
"size_b": 3,
"round_devices": INTEGER_DEVICES,
},
}