Marmelator/image-classification-service
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 29 Wiki Activity

- Adapted response formatting logic for threshold maps passed via env vars.

Browse Source 
- Added test for reading threshold maps and values from env vars.
This commit is contained in:
Matthias Bisping 2022-09-12 13:04:55 +02:00
parent c60e8cd678
commit ad9c5657fe
3 changed files with 91 additions and 51 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
config.yaml
View File

@ -18,10 +18,6 @@ filters:
image_to_page_quotient: # Image size to page size ratio (ratio of geometric means of areas)
min: $MIN_REL_IMAGE_SIZE|0.05 # Minimum permissible
max: $MAX_REL_IMAGE_SIZE|0.75 # Maximum permissible
# Fixme: Temporary solution, delete
customized: # Customized settings per class (RED-5202)
max:
signature: $MAX_REL_SIGNATURE_SIZE|0.4
image_width_to_height_quotient: # Image width to height ratio
min: $MIN_IMAGE_FORMAT|0.1 # Minimum permissible

94
image_prediction/transformer/transformers/response.py
View File

@ -4,11 +4,12 @@ import os
from functools import lru_cache
from operator import itemgetter
from funcy import filter, juxt, first, rest, compose
from image_prediction.config import CONFIG
from image_prediction.exceptions import ParsingError
from image_prediction.transformer.transformer import Transformer
from image_prediction.utils import get_logger
from funcy import filter, juxt, first, rest
logger = get_logger()
@ -19,6 +20,37 @@ class ResponseTransformer(Transformer):
return build_image_info(data)
def get_class_specific_min_image_to_page_quotient(label):
return get_class_specific_quotient("REL_IMAGE_SIZE", label, "min", CONFIG.filters.image_to_page_quotient.min)
def get_class_specific_max_image_to_page_quotient(label):
return get_class_specific_quotient("REL_IMAGE_SIZE", label, "max", CONFIG.filters.image_to_page_quotient.max)
def get_class_specific_min_image_width_to_height_quotient(label):
return get_class_specific_quotient("IMAGE_FORMAT", label, "min", CONFIG.filters.image_width_to_height_quotient.min)
def get_class_specific_max_image_width_to_height_quotient(label):
return get_class_specific_quotient("IMAGE_FORMAT", label, "max", CONFIG.filters.image_width_to_height_quotient.max)
def get_class_specific_min_classification_confidence(label):
return get_class_specific_quotient("CONFIDENCE", label, "min", CONFIG.filters.min_confidence)
def get_class_specific_quotient(prefix, label, bound, fallback_value):
def fallback():
logger.warning(f"Failed to resolve {bound} {prefix.lower().replace('_', '-')} value for class '{label}'.")
return fallback_value
assert bound in ["min", "max"]
threshold_map = parse_env_var(prefix) or {}
return threshold_map.get(label, {}).get(bound) or fallback()
def build_image_info(data: dict) -> dict:
def compute_geometric_quotient():
page_area_sqrt = math.sqrt(abs(page_width * page_height))
@ -29,23 +61,29 @@ def build_image_info(data: dict) -> dict:
"page_width", "page_height", "x1", "x2", "y1", "y2", "width", "height", "alpha"
)(data)
quotient = round(compute_geometric_quotient(), 4)
min_image_to_page_quotient_breached = bool(quotient < CONFIG.filters.image_to_page_quotient.min)
max_image_to_page_quotient_breached = __is_max_image_to_page_quotient_breached(
quotient, data["classification"]["label"]
)
min_image_width_to_height_quotient_breached = bool(
width / height < CONFIG.filters.image_width_to_height_quotient.min
)
max_image_width_to_height_quotient_breached = bool(
width / height > CONFIG.filters.image_width_to_height_quotient.max
)
classification = data["classification"]
label = classification["label"]
representation = data["representation"]
min_confidence_breached = bool(max(classification["probabilities"].values()) < CONFIG.filters.min_confidence)
geometric_quotient = round(compute_geometric_quotient(), 4)
min_image_to_page_quotient_breached = bool(
geometric_quotient < get_class_specific_min_image_to_page_quotient(label)
)
max_image_to_page_quotient_breached = bool(
geometric_quotient > get_class_specific_max_image_to_page_quotient(label)
)
min_image_width_to_height_quotient_breached = bool(
width / height < get_class_specific_min_image_width_to_height_quotient(label)
)
max_image_width_to_height_quotient_breached = bool(
width / height > get_class_specific_max_image_width_to_height_quotient(label)
)
min_confidence_breached = bool(
max(classification["probabilities"].values()) < get_class_specific_min_classification_confidence(label)
)
image_info = {
"classification": classification,
@ -56,7 +94,7 @@ def build_image_info(data: dict) -> dict:
"filters": {
"geometry": {
"imageSize": {
"quotient": quotient,
"quotient": geometric_quotient,
"tooLarge": max_image_to_page_quotient_breached,
"tooSmall": min_image_to_page_quotient_breached,
},
@ -82,28 +120,18 @@ def build_image_info(data: dict) -> dict:
return image_info
def __is_max_image_to_page_quotient_breached(quotient: float, label: str) -> bool:
default_max_quotient = CONFIG.filters.image_to_page_quotient.max
customized_entries = CONFIG.filters.image_to_page_quotient.customized.max
max_quotient = customized_entries.get(label, default_max_quotient)
max_quotient = max_quotient if max_quotient else default_max_quotient
return bool(quotient > max_quotient)
@lru_cache(maxsize=None)
def parse_env_var(prefix, fallback_value):
head, tail = juxt(first, rest)(filter(prefix, os.environ))
if not head or tail:
logger.warning(
f"Found multiple candidates for environment variable with prefix '{prefix}', falling back to default value."
)
return fallback_value
def parse_env_var(prefix):
head, tail = juxt(first, compose(list, rest))(filter(prefix, os.environ))
if not head:
logger.warning(f"Found no environment variable with prefix '{prefix}'.")
elif tail:
logger.warning(f"Found multiple candidates for environment variable with prefix '{prefix}'.")
else:
try:
return parse_env_var_value(os.environ[head])
except ParsingError as err:
logger.warning(f"{err}, falling back to default value.")
return fallback_value
logger.warning(err)
def parse_env_var_value(env_var_value):

44
test/unit_tests/response_transformer_test.py
View File

@ -1,21 +1,37 @@
import json
import os
import pytest
from image_prediction.transformer.transformers.response import __is_max_image_to_page_quotient_breached
from image_prediction.transformer.transformers.response import (
get_class_specific_min_image_to_page_quotient,
get_class_specific_max_image_to_page_quotient,
get_class_specific_max_image_width_to_height_quotient,
get_class_specific_min_image_width_to_height_quotient,
get_class_specific_min_classification_confidence,
)
@pytest.fixture
def expected_is_breached(quotient, label):
if label == "signature" and quotient < 0.4:
return False
elif label == "signature" and quotient >= 0.4:
return True
elif quotient < 0.7:
return False
else:
return True
def label():
return "signature"
@pytest.mark.parametrize("quotient", [0.1, 0.5])
@pytest.mark.parametrize("label", ["logo", "signature"])
def test_customized_per_label_ratio_breach(quotient, label, expected_is_breached):
assert __is_max_image_to_page_quotient_breached(quotient, label) == expected_is_breached
@pytest.fixture
def page_quotient_threshold_map(label):
# TODO: suboptimal, as actual environment is used
os.environ["REL_IMAGE_SIZE_MAP"] = json.dumps({label: {"min": 0.1, "max": 0.2}})
os.environ["IMAGE_FORMAT_MAP"] = json.dumps({label: {"min": 0.5, "max": 0.4}})
os.environ["CONFIDENCE"] = json.dumps({label: {"min": 0.8}})
yield
for env_var in ("REL_IMAGE_SIZE_MAP", "IMAGE_FORMAT_MAP", "CONFIDENCE"):
os.environ.pop(env_var)
# FIXME: Runs correctly in isolation, but fails when other tests are run before
def test_read_environment_vars_for_thresholds(page_quotient_threshold_map, label):
assert get_class_specific_min_image_to_page_quotient(label) == 0.1
assert get_class_specific_max_image_to_page_quotient(label) == 0.2
assert get_class_specific_min_image_width_to_height_quotient(label) == 0.5
assert get_class_specific_max_image_width_to_height_quotient(label) == 0.4
assert get_class_specific_min_classification_confidence(label) == 0.8