feat: RED-10765: ignore perceptual hash for image deduplication and prefer to keep the ones with allPassed set to True

feat: RED-10765: filter out classifications for 'duplicate' images present in the document

Closes RED-10765

See merge request redactmanager/image-classification-service!23

.coveragerc

@@ -0,0 +1,63 @@
+# .coveragerc to control coverage.py
+[run]
+branch = True
+parallel = True
+command_line = -m pytest
+concurrency = multiprocessing
+omit =
+    */site-packages/*
+    */distutils/*
+    */test/*
+    */__init__.py
+    */setup.py
+	*/venv/*
+	*/env/*
+	*/build_venv/*
+	*/build_env/*
+	*/utils/banner.py
+	*/utils/logger.py
+	*/src/*
+source =
+    image_prediction
+relative_files = True
+data_file = .coverage
+
+[report]
+# Regexes for lines to exclude from consideration
+exclude_lines =
+    # Have to re-enable the standard pragma
+    pragma: no cover
+
+    # Don't complain about missing debug-only code:
+    def __repr__
+    if self\.debug
+
+    # Don't complain if tests don't hit defensive assertion code:
+    raise AssertionError
+    raise NotImplementedError
+
+    # Don't complain if non-runnable code isn't run:
+    if 0:
+    if __name__ == .__main__.:
+omit =
+    */site-packages/*
+    */distutils/*
+    */test/*
+    */__init__.py
+    */setup.py
+	*/venv/*
+	*/env/*
+	*/build_venv/*
+	*/build_env/*
+	*/utils/banner.py
+	*/utils/logger.py
+	*/src/*
+	*/pdf_annotation.py
+
+ignore_errors = True
+
+[html]
+directory = reports
+
+[xml]
+output = reports/coverage.xml

.dvc/config

@@ -1,5 +1,8 @@
 [core]
-    remote = vector
+    remote = azure_remote
+    autostage = true
 ['remote "vector"']
-    url = ssh://vector.iqser.com/research/image_service/
+    url = ssh://vector.iqser.com/research/image-prediction/
     port = 22
+['remote "azure_remote"']
+    url = azure://image-classification-dvc/

.gitignore

@@ -1,7 +1,8 @@
 .vscode/
 *.h5
-/venv/
+*venv
 .idea/
+src/data
 
 !.gitignore
 *.project
@@ -32,6 +33,9 @@
 **/classpath-data.json
 **/dependencies-and-licenses-overview.txt
 
+.coverage
+.coverage\.*\.*
+
 
 *__pycache__
 *.egg-info*
@@ -44,7 +48,6 @@
 *misc
 
 /coverage_html_report/
-.coverage
 
 # Created by https://www.toptal.com/developers/gitignore/api/linux,pycharm
 # Edit at https://www.toptal.com/developers/gitignore?templates=linux,pycharm
@@ -170,6 +173,4 @@ fabric.properties
 # https://plugins.jetbrains.com/plugin/12206-codestream
 .idea/codestream.xml
 
-# End of https://www.toptal.com/developers/gitignore/api/linux,pycharm
-/image_prediction/data/mlruns/
-/data/mlruns/
+# End of https://www.toptal.com/developers/gitignore/api/linux,pycharm

.gitlab-ci.yml

@@ -0,0 +1,51 @@
+include:
+  - project: "Gitlab/gitlab"
+    ref: main
+    file: "/ci-templates/research/dvc.gitlab-ci.yml"
+  - project: "Gitlab/gitlab"
+    ref: main
+    file: "/ci-templates/research/versioning-build-test-release.gitlab-ci.yml"
+
+variables:
+  NEXUS_PROJECT_DIR: red
+  IMAGENAME: "${CI_PROJECT_NAME}"
+  INTEGRATION_TEST_FILE: "${CI_PROJECT_ID}.pdf"
+  FF_USE_FASTZIP: "true" # enable fastzip - a faster zip implementation that also supports level configuration.
+  ARTIFACT_COMPRESSION_LEVEL: default # can also be set to fastest, fast, slow and slowest. If just enabling fastzip is not enough try setting this to fastest or fast.
+  CACHE_COMPRESSION_LEVEL: default # same as above, but for caches
+  # TRANSFER_METER_FREQUENCY: 5s # will display transfer progress every 5 seconds for artifacts and remote caches. For debugging purposes.
+
+stages:
+  - data
+  - setup
+  - tests
+  - sonarqube
+  - versioning
+  - build
+  - integration-tests
+  - release
+
+docker-build:
+  extends: .docker-build
+  needs:
+    - job: dvc-pull
+      artifacts: true
+    - !reference [.needs-versioning, needs] # leave this line as is
+  
+###################
+# INTEGRATION TESTS
+trigger-integration-tests:
+  extends: .integration-tests
+  # ADD THE MODEL BUILD WHICH SHOULD TRIGGER THE INTEGRATION TESTS
+  # needs:
+  #   - job: docker-build::model_name
+  #     artifacts: true
+  rules:
+    - when: never
+
+#########
+# RELEASE
+release:
+  extends: .release
+  needs:
+    - !reference [.needs-versioning, needs] # leave this line as is

.gitmodules

@@ -1,3 +0,0 @@
-[submodule "incl/redai_image"]
-	path = incl/redai_image
-	url = ssh://git@git.iqser.com:2222/rr/redai_image.git

.python-version

@@ -0,0 +1 @@
+3.10

Dockerfile

@@ -1,25 +1,73 @@
-ARG BASE_ROOT="nexus.iqser.com:5001/red/"
-ARG VERSION_TAG="latest"
+FROM python:3.10-slim AS builder
 
-FROM ${BASE_ROOT}image-prediction-base:${VERSION_TAG}
+ARG GITLAB_USER
+ARG GITLAB_ACCESS_TOKEN
 
-WORKDIR /app/service
+ARG PYPI_REGISTRY_RESEARCH=https://gitlab.knecon.com/api/v4/groups/19/-/packages/pypi
+ARG POETRY_SOURCE_REF_RESEARCH=gitlab-research
 
-COPY src src
-COPY data data
-COPY image_prediction image_prediction
-COPY incl/redai_image/redai incl/redai_image/redai
-COPY setup.py setup.py
-COPY requirements.txt requirements.txt
-COPY config.yaml config.yaml
+ARG PYPI_REGISTRY_RED=https://gitlab.knecon.com/api/v4/groups/12/-/packages/pypi
+ARG POETRY_SOURCE_REF_RED=gitlab-red
 
-# Install dependencies differing from base image.
-RUN python3 -m pip install -r requirements.txt
+ARG PYPI_REGISTRY_FFORESIGHT=https://gitlab.knecon.com/api/v4/groups/269/-/packages/pypi
+ARG POETRY_SOURCE_REF_FFORESIGHT=gitlab-fforesight
 
-RUN python3 -m pip install -e .
-RUN python3 -m pip install -e incl/redai_image/redai
+ARG VERSION=dev
+
+LABEL maintainer="Research <research@knecon.com>"
+LABEL version="${VERSION}"
+
+WORKDIR /app
+
+###########
+# ENV SETUP
+ENV PYTHONDONTWRITEBYTECODE=true
+ENV PYTHONUNBUFFERED=true
+ENV POETRY_HOME=/opt/poetry
+ENV PATH="$POETRY_HOME/bin:$PATH"
+
+RUN apt-get update && \
+    apt-get install -y curl git bash build-essential libffi-dev libssl-dev && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
+
+RUN curl -sSL https://install.python-poetry.org | python3 -
+RUN poetry --version
+
+COPY pyproject.toml poetry.lock ./
+
+RUN poetry config virtualenvs.create true && \
+    poetry config virtualenvs.in-project true && \
+    poetry config installer.max-workers 10 && \
+    poetry config repositories.${POETRY_SOURCE_REF_RESEARCH} ${PYPI_REGISTRY_RESEARCH} && \
+    poetry config http-basic.${POETRY_SOURCE_REF_RESEARCH} ${GITLAB_USER} ${GITLAB_ACCESS_TOKEN} && \
+    poetry config repositories.${POETRY_SOURCE_REF_RED} ${PYPI_REGISTRY_RED} && \
+    poetry config http-basic.${POETRY_SOURCE_REF_RED} ${GITLAB_USER} ${GITLAB_ACCESS_TOKEN} && \
+    poetry config repositories.${POETRY_SOURCE_REF_FFORESIGHT} ${PYPI_REGISTRY_FFORESIGHT} && \
+    poetry config http-basic.${POETRY_SOURCE_REF_FFORESIGHT} ${GITLAB_USER} ${GITLAB_ACCESS_TOKEN} && \
+    poetry install --without=dev -vv --no-interaction --no-root
+
+###############
+# WORKING IMAGE
+FROM python:3.10-slim
+
+WORKDIR /app
+
+# COPY SOURCE CODE FROM BUILDER IMAGE
+COPY --from=builder /app /app
+# COPY BILL OF MATERIALS (BOM)
+COPY bom.json /bom.json
+
+ENV PATH="/app/.venv/bin:$PATH"
+
+###################
+# COPY SOURCE CODE
+COPY ./src ./src
+COPY ./config ./config
+COPY ./data ./data
+COPY banner.txt ./
 
 EXPOSE 5000
 EXPOSE 8080
 
-CMD ["python3", "src/serve.py"]
+CMD [ "python", "src/serve.py"]

Dockerfile_base

@@ -1,25 +0,0 @@
-FROM python:3.8 as builder1
-
-# Use a virtual environment.
-RUN python -m venv /app/venv
-ENV PATH="/app/venv/bin:$PATH"
-
-# Upgrade pip.
-RUN python -m pip install --upgrade pip
-
-# Make a directory for the service files and copy the service repo into the container.
-WORKDIR /app/service
-COPY ./requirements.txt ./requirements.txt
-
-# Install dependencies.
-RUN python3 -m pip install -r requirements.txt
-
-# Make a new container and copy all relevant files over to filter out temporary files
-# produced during setup to reduce the final container's size.
-FROM python:3.8
-
-WORKDIR /app/
-COPY --from=builder1  /app .
-ENV PATH="/app/venv/bin:$PATH"
-
-WORKDIR /app/service

Dockerfile_tests

@@ -0,0 +1,43 @@
+FROM python:3.10
+
+ARG USERNAME
+ARG TOKEN
+ARG PYPI_REGISTRY_RESEARCH=https://gitlab.knecon.com/api/v4/groups/19/-/packages/pypi
+ARG POETRY_SOURCE_REF_RESEARCH=gitlab-research
+ARG PYPI_REGISTRY_RED=https://gitlab.knecon.com/api/v4/groups/12/-/packages/pypi
+ARG POETRY_SOURCE_REF_RED=gitlab-red
+ARG VERSION=dev
+
+LABEL maintainer="Research <research@knecon.com>"
+LABEL version="${VERSION}"
+
+WORKDIR /app
+
+ENV PYTHONUNBUFFERED=true
+ENV POETRY_HOME=/opt/poetry
+ENV PATH="$POETRY_HOME/bin:$PATH"
+
+RUN curl -sSL https://install.python-poetry.org | python3 -
+
+COPY ./data ./data
+COPY ./test ./test
+COPY ./config ./config
+COPY ./src ./src
+COPY pyproject.toml poetry.lock banner.txt config.yaml./
+
+RUN poetry config virtualenvs.create false && \
+    poetry config installer.max-workers 10 && \
+    poetry config repositories.${POETRY_SOURCE_REF_RESEARCH} ${PYPI_REGISTRY_RESEARCH} && \
+    poetry config http-basic.${POETRY_SOURCE_REF_RESEARCH} ${USERNAME} ${TOKEN} && \
+    poetry config repositories.${POETRY_SOURCE_REF_RED} ${PYPI_REGISTRY_RED} && \
+    poetry config http-basic.${POETRY_SOURCE_REF_RED} ${USERNAME} ${TOKEN} && \
+    poetry install --without=dev -vv --no-interaction --no-root
+
+EXPOSE 5000
+EXPOSE 8080
+
+RUN apt update --yes
+RUN apt install vim --yes
+RUN apt install poppler-utils --yes
+
+CMD coverage run -m pytest test/ --tb=native -q -s -vvv -x && coverage combine && coverage report -m && coverage xml

README.md

@@ -1,25 +1,143 @@
-### Building
+### Setup
 
 Build base image
 ```bash
-setup/docker.sh
-```
-
-Build head image
-```bash
-docker build -f Dockerfile -t image-prediction . --build-arg BASE_ROOT=""
+docker build -t image-classification-image --progress=plain --no-cache \
+    -f Dockerfile \
+    --build-arg USERNAME=$GITLAB_USER \
+    --build-arg TOKEN=$GITLAB_ACCESS_TOKEN \
+    .
 ```
 
 ### Usage
 
+#### Without Docker
+
+
+```bash
+py scripts/run_pipeline.py /path/to/a/pdf
+```
+
+#### With Docker
+
 Shell 1
 
 ```bash
-docker run --rm --net=host --rm image-prediction
+docker run --rm --net=host image-prediction
 ```
 
 Shell 2
 
 ```bash
-python scripts/pyinfra_mock.py --pdf_path /path/to/a/pdf
+python scripts/pyinfra_mock.py /path/to/a/pdf
 ```
+
+### Tests
+
+Run for example this command to execute all tests and get a coverage report:
+
+```bash
+coverage run -m pytest test --tb=native -q -s -vvv -x && coverage combine && coverage report -m
+```
+
+After having built the service container as specified above, you can also run tests in a container as follows:
+
+```bash
+./run_tests.sh
+```
+
+### Message Body Formats
+
+
+#### Request Format
+
+The request messages need to provide the fields `"dossierId"` and `"fileId"`. A request should look like this:
+
+```json
+{
+    "dossierId": "<string identifier>",
+    "fileId": "<string identifier>"
+}
+```
+
+Any additional keys are ignored.
+
+
+#### Response Format
+
+Response bodies contain information about the identified class of the image, the confidence of the classification, the
+position and size of the image as well as the results of additional convenience filters which can be configured through
+environment variables. A response body looks like this:
+
+```json
+{
+  "dossierId": "debug",
+  "fileId": "13ffa9851740c8d20c4c7d1706d72f2a",
+  "data": [...]
+}
+```
+
+An image metadata record (entry in `"data"` field of a response body) looks like this:
+
+```json
+{
+  "classification": {
+    "label": "logo",
+    "probabilities": {
+      "logo": 1.0,
+      "signature": 1.1599173226749333e-17,
+      "other": 2.994595513398207e-23,
+      "formula": 4.352109377281029e-31
+    }
+  },
+  "position": {
+    "x1": 475.95,
+    "x2": 533.4,
+    "y1": 796.47,
+    "y2": 827.62,
+    "pageNumber": 6
+  },
+  "geometry": {
+    "width": 57.44999999999999,
+    "height": 31.149999999999977
+  },
+  "alpha": false,
+  "filters": {
+    "geometry": {
+      "imageSize": {
+        "quotient": 0.05975350599135938,
+        "tooLarge": false,
+        "tooSmall": false
+      },
+      "imageFormat": {
+        "quotient": 1.8443017656500813,
+        "tooTall": false,
+        "tooWide": false
+      }
+    },
+    "probability": {
+      "unconfident": false
+    },
+    "allPassed": true
+  }
+}
+```
+
+
+## Configuration
+
+A configuration file is located under `config.yaml`. All relevant variables can be configured via
+exporting environment variables.
+
+| __Environment Variable__           | Default                            | Description                                                                            |
+|------------------------------------|------------------------------------|----------------------------------------------------------------------------------------|
+| __LOGGING_LEVEL_ROOT__             | "INFO"                             | Logging level for log file messages                                                    |
+| __VERBOSE__                        | *true*                             | Service prints document processing progress to stdout                                  |
+| __BATCH_SIZE__                     | 16                                 | Number of images in memory simultaneously per service instance                         |
+| __RUN_ID__                         | "fabfb1f192c745369b88cab34471aba7" | The ID of the mlflow run to load the image classifier from                             |
+| __MIN_REL_IMAGE_SIZE__             | 0.05                               | Minimally permissible image size to page size ratio                                    |
+| __MAX_REL_IMAGE_SIZE__             | 0.75                               | Maximally permissible image size to page size ratio                                    |
+| __MIN_IMAGE_FORMAT__               | 0.1                                | Minimally permissible image width to height ratio                                      |
+| __MAX_IMAGE_FORMAT__               | 10                                 | Maximally permissible image width to height ratio                                      |
+
+See also: https://git.iqser.com/projects/RED/repos/helm/browse/redaction/templates/image-service-v2

bamboo-specs/pom.xml

@@ -1,40 +0,0 @@
-<project xmlns="http://maven.apache.org/POM/4.0.0" xmlns:xsi="http://www.w3.org/2001/XMLSchema-instance"
-         xsi:schemaLocation="http://maven.apache.org/POM/4.0.0 http://maven.apache.org/maven-v4_0_0.xsd">
-  <modelVersion>4.0.0</modelVersion>
-
-  <parent>
-    <groupId>com.atlassian.bamboo</groupId>
-    <artifactId>bamboo-specs-parent</artifactId>
-    <version>7.1.2</version>
-    <relativePath/>
-  </parent>
-
-  <artifactId>bamboo-specs</artifactId>
-  <version>1.0.0-SNAPSHOT</version>
-  <packaging>jar</packaging>
-
-  <properties>
-    <sonar.skip>true</sonar.skip>
-  </properties>
-
-  <dependencies>
-    <dependency>
-      <groupId>com.atlassian.bamboo</groupId>
-      <artifactId>bamboo-specs-api</artifactId>
-    </dependency>
-    <dependency>
-      <groupId>com.atlassian.bamboo</groupId>
-      <artifactId>bamboo-specs</artifactId>
-    </dependency>
-
-    <!-- Test dependencies -->
-    <dependency>
-      <groupId>junit</groupId>
-      <artifactId>junit</artifactId>
-      <scope>test</scope>
-    </dependency>
-  </dependencies>
-
-  <!-- run 'mvn test' to perform offline validation of the plan -->
-  <!-- run 'mvn -Ppublish-specs' to upload the plan to your Bamboo server -->
-</project>

bamboo-specs/src/main/java/buildjob/PlanSpec.java

@@ -1,182 +0,0 @@
-package buildjob;
-
-import com.atlassian.bamboo.specs.api.BambooSpec;
-import com.atlassian.bamboo.specs.api.builders.BambooKey;
-import com.atlassian.bamboo.specs.api.builders.docker.DockerConfiguration;
-import com.atlassian.bamboo.specs.api.builders.permission.PermissionType;
-import com.atlassian.bamboo.specs.api.builders.permission.Permissions;
-import com.atlassian.bamboo.specs.api.builders.permission.PlanPermissions;
-import com.atlassian.bamboo.specs.api.builders.plan.Job;
-import com.atlassian.bamboo.specs.api.builders.plan.Plan;
-import com.atlassian.bamboo.specs.api.builders.plan.PlanIdentifier;
-import com.atlassian.bamboo.specs.api.builders.plan.Stage;
-import com.atlassian.bamboo.specs.api.builders.plan.branches.BranchCleanup;
-import com.atlassian.bamboo.specs.api.builders.plan.branches.PlanBranchManagement;
-import com.atlassian.bamboo.specs.api.builders.project.Project;
-import com.atlassian.bamboo.specs.builders.task.CheckoutItem;
-import com.atlassian.bamboo.specs.builders.task.InjectVariablesTask;
-import com.atlassian.bamboo.specs.builders.task.ScriptTask;
-import com.atlassian.bamboo.specs.builders.task.VcsCheckoutTask;
-import com.atlassian.bamboo.specs.builders.task.CleanWorkingDirectoryTask;
-import com.atlassian.bamboo.specs.builders.task.VcsTagTask;
-import com.atlassian.bamboo.specs.builders.trigger.BitbucketServerTrigger;
-import com.atlassian.bamboo.specs.model.task.InjectVariablesScope;
-import com.atlassian.bamboo.specs.api.builders.Variable;
-import com.atlassian.bamboo.specs.util.BambooServer;
-import com.atlassian.bamboo.specs.builders.task.ScriptTask;
-import com.atlassian.bamboo.specs.model.task.ScriptTaskProperties.Location;
-
-/**
- * Plan configuration for Bamboo.
- * Learn more on: <a href="https://confluence.atlassian.com/display/BAMBOO/Bamboo+Specs">https://confluence.atlassian.com/display/BAMBOO/Bamboo+Specs</a>
- */
-@BambooSpec
-public class PlanSpec {
-
-    private static final String SERVICE_NAME = "image-prediction";
-    private static final String SERVICE_NAME_BASE = "image-prediction-base";
-
-    private static final String SERVICE_KEY = SERVICE_NAME.toUpperCase().replaceAll("-","").replaceAll("_","");
-
-    /**
-     * Run main to publish plan on Bamboo
-     */
-    public static void main(final String[] args) throws Exception {
-        //By default credentials are read from the '.credentials' file.
-        BambooServer bambooServer = new BambooServer("http://localhost:8085");
-
-        Plan plan = new PlanSpec().createDockerBuildPlan();
-        bambooServer.publish(plan);
-        PlanPermissions planPermission = new PlanSpec().createPlanPermission(plan.getIdentifier());
-        bambooServer.publish(planPermission);
-    }
-
-    private PlanPermissions createPlanPermission(PlanIdentifier planIdentifier) {
-        Permissions permission = new Permissions()
-                .userPermissions("atlbamboo", PermissionType.EDIT, PermissionType.VIEW, PermissionType.ADMIN, PermissionType.CLONE, PermissionType.BUILD)
-                .groupPermissions("research", PermissionType.EDIT, PermissionType.VIEW, PermissionType.CLONE, PermissionType.BUILD)
-                .groupPermissions("Development", PermissionType.EDIT, PermissionType.VIEW, PermissionType.CLONE, PermissionType.BUILD)
-                .groupPermissions("QA", PermissionType.EDIT, PermissionType.VIEW, PermissionType.CLONE, PermissionType.BUILD)
-                .loggedInUserPermissions(PermissionType.VIEW)
-                .anonymousUserPermissionView();
-        return new PlanPermissions(planIdentifier.getProjectKey(), planIdentifier.getPlanKey()).permissions(permission);
-    }
-
-    private Project project() {
-        return new Project()
-                .name("RED")
-                .key(new BambooKey("RED"));
-    }
-
-    public Plan createDockerBuildPlan() {
-    return new Plan(
-            project(),
-            SERVICE_NAME, new BambooKey(SERVICE_KEY))
-            .description("Docker build for image-prediction.")
-            // .variables()
-            .stages(new Stage("Build Stage")
-              .jobs(
-                new Job("Build Job", new BambooKey("BUILD"))
-                  .tasks(
-                    new CleanWorkingDirectoryTask()
-                        .description("Clean working directory.")
-                        .enabled(true),
-                    new VcsCheckoutTask()
-                        .description("Checkout default repository.")
-                        .checkoutItems(new CheckoutItem().defaultRepository()),
-                    new VcsCheckoutTask()
-                        .description("Checkout redai_image research repository.")
-                        .checkoutItems(new CheckoutItem().repository("RR / redai_image").path("redai_image")),
-                    new ScriptTask()
-                        .description("Set config and keys.")
-                        .inlineBody("mkdir -p ~/.ssh\n" +
-                                    "echo \"${bamboo.bamboo_agent_ssh}\" | base64 -d >> ~/.ssh/id_rsa\n" +
-                                    "echo \"host vector.iqser.com\" > ~/.ssh/config\n" +
-                                    "echo \"    user bamboo-agent\" >> ~/.ssh/config\n" +
-                                    "chmod 600 ~/.ssh/config ~/.ssh/id_rsa"),
-                    new ScriptTask()
-                        .description("Build Docker container.")
-                        .location(Location.FILE)
-                        .fileFromPath("bamboo-specs/src/main/resources/scripts/docker-build.sh")
-                        .argument(SERVICE_NAME + " " + SERVICE_NAME_BASE))
-                  .dockerConfiguration(
-                      new DockerConfiguration()
-                        .image("nexus.iqser.com:5001/infra/release_build:4.2.0")
-                        .volume("/var/run/docker.sock", "/var/run/docker.sock")),
-                new Job("Sonar Job", new BambooKey("SONAR"))
-                  .tasks(
-                    new CleanWorkingDirectoryTask()
-                        .description("Clean working directory.")
-                        .enabled(true),
-                    new VcsCheckoutTask()
-                        .description("Checkout default repository.")
-                        .checkoutItems(new CheckoutItem().defaultRepository()),
-                    new VcsCheckoutTask()
-                        .description("Checkout redai_image repository.")
-                        .checkoutItems(new CheckoutItem().repository("RR / redai_image").path("redai_image")),
-                    new ScriptTask()
-                        .description("Set config and keys.")
-                        .inlineBody("mkdir -p ~/.ssh\n" +
-                                    "echo \"${bamboo.bamboo_agent_ssh}\" | base64 -d >> ~/.ssh/id_rsa\n" +
-                                    "echo \"host vector.iqser.com\" > ~/.ssh/config\n" +
-                                    "echo \"    user bamboo-agent\" >> ~/.ssh/config\n" +
-                                    "chmod 600 ~/.ssh/config ~/.ssh/id_rsa"),
-                    new ScriptTask()
-                        .description("Run Sonarqube scan.")
-                        .location(Location.FILE)
-                        .fileFromPath("bamboo-specs/src/main/resources/scripts/sonar-scan.sh")
-                        .argument(SERVICE_NAME))
-                  .dockerConfiguration(
-                      new DockerConfiguration()
-                        .image("nexus.iqser.com:5001/infra/release_build:4.2.0")
-                        .volume("/var/run/docker.sock", "/var/run/docker.sock"))),
-            new Stage("Licence Stage")
-              .jobs(
-                new Job("Git Tag Job", new BambooKey("GITTAG"))
-                  .tasks(
-                    new VcsCheckoutTask()
-                        .description("Checkout default repository.")
-                        .checkoutItems(new CheckoutItem().defaultRepository()),
-                    new ScriptTask()
-                        .description("Build git tag.")
-                        .location(Location.FILE)
-                        .fileFromPath("bamboo-specs/src/main/resources/scripts/git-tag.sh"),
-                    new InjectVariablesTask()
-                        .description("Inject git tag.")
-                        .path("git.tag")
-                        .namespace("g")
-                        .scope(InjectVariablesScope.LOCAL),
-                    new VcsTagTask()
-                        .description("${bamboo.g.gitTag}")
-                        .tagName("${bamboo.g.gitTag}")
-                        .defaultRepository())
-                .dockerConfiguration(
-                    new DockerConfiguration()
-                        .image("nexus.iqser.com:5001/infra/release_build:4.4.1")),
-                new Job("Licence Job", new BambooKey("LICENCE"))
-                  .enabled(false)
-                  .tasks(
-                    new VcsCheckoutTask()
-                        .description("Checkout default repository.")
-                        .checkoutItems(new CheckoutItem().defaultRepository()),
-                    new ScriptTask()
-                        .description("Build licence.")
-                        .location(Location.FILE)
-                        .fileFromPath("bamboo-specs/src/main/resources/scripts/create-licence.sh"))
-                  .dockerConfiguration(
-                    new DockerConfiguration()
-                        .image("nexus.iqser.com:5001/infra/maven:3.6.2-jdk-13-3.0.0")
-                        .volume("/etc/maven/settings.xml", "/usr/share/maven/ref/settings.xml")
-                        .volume("/var/run/docker.sock", "/var/run/docker.sock"))))
-            .linkedRepositories("RR / " + SERVICE_NAME)
-            .linkedRepositories("RR / redai_image")
-            .triggers(new BitbucketServerTrigger())
-            .planBranchManagement(new PlanBranchManagement()
-              .createForVcsBranch()
-              .delete(new BranchCleanup()
-                  .whenInactiveInRepositoryAfterDays(14))
-              .notificationForCommitters());
-    }
-
-
-}

bamboo-specs/src/main/resources/scripts/create-licence.sh

@@ -1,19 +0,0 @@
-#!/bin/bash
-set -e
-
-if [[ \"${bamboo_version_tag}\" != \"dev\" ]]
-then
-    ${bamboo_capability_system_builder_mvn3_Maven_3}/bin/mvn \
-                    -f ${bamboo_build_working_directory}/pom.xml \
-                    versions:set  \
-                    -DnewVersion=${bamboo_version_tag}
-
-    ${bamboo_capability_system_builder_mvn3_Maven_3}/bin/mvn \
-                    -f ${bamboo_build_working_directory}/pom.xml \
-                    -B clean deploy \
-                    -e -DdeployAtEnd=true \
-                    -Dmaven.wagon.http.ssl.insecure=true \
-                    -Dmaven.wagon.http.ssl.allowall=true \
-                    -Dmaven.wagon.http.ssl.ignore.validity.dates=true \
-                    -DaltDeploymentRepository=iqser_release::default::https://nexus.iqser.com/repository/gin4-platform-releases
-fi

bamboo-specs/src/main/resources/scripts/docker-build.sh

@@ -1,19 +0,0 @@
-#!/bin/bash
-set -e
-
-SERVICE_NAME=$1
-SERVICE_NAME_BASE=$2
-
-python3 -m venv build_venv
-source build_venv/bin/activate
-python3 -m pip install --upgrade pip
-
-pip install dvc
-pip install 'dvc[ssh]'
-dvc pull
-
-echo "index-url = https://${bamboo_nexus_user}:${bamboo_nexus_password}@nexus.iqser.com/repository/python-combind/simple" >> pip.conf
-docker build -f Dockerfile_base  -t nexus.iqser.com:5001/red/$SERVICE_NAME_BASE:${bamboo_version_tag} .
-docker build -f Dockerfile  -t nexus.iqser.com:5001/red/$SERVICE_NAME:${bamboo_version_tag} --build-arg VERSION_TAG=${bamboo_version_tag} .
-echo "${bamboo_nexus_password}" | docker login --username "${bamboo_nexus_user}" --password-stdin nexus.iqser.com:5001
-docker push nexus.iqser.com:5001/red/$SERVICE_NAME:${bamboo_version_tag}

bamboo-specs/src/main/resources/scripts/git-tag.sh

@@ -1,9 +0,0 @@
-#!/bin/bash
-set -e
-
-if [[ "${bamboo_version_tag}" = "dev" ]]
-then
-    echo "gitTag=${bamboo_planRepository_1_branch}_${bamboo_buildNumber}" > git.tag
-else
-    echo "gitTag=${bamboo_version_tag}" > git.tag
-fi

bamboo-specs/src/main/resources/scripts/sonar-scan.sh

@@ -1,51 +0,0 @@
-#!/bin/bash
-set -e
-
-export JAVA_HOME=/usr/bin/sonar-scanner/jre
-
-python3 -m venv build_venv
-source build_venv/bin/activate
-python3 -m pip install --upgrade pip
-
-echo "dev setup for unit test and coverage 💖"
-
-pip install -e .
-pip install -r requirements.txt
-
-SERVICE_NAME=$1
-
-echo "dependency-check:aggregate"
-mkdir -p reports
-dependency-check --enableExperimental -f JSON -f HTML -f XML \
-  --disableAssembly -s . -o reports --project $SERVICE_NAME --exclude ".git/**" --exclude "venv/**" \
-  --exclude "build_venv/**" --exclude "**/__pycache__/**" --exclude "bamboo-specs/**"
-
-if [[ -z "${bamboo_repository_pr_key}" ]]
-then
-    echo "Sonar Scan for branch: ${bamboo_planRepository_1_branch}"
-    /usr/bin/sonar-scanner/bin/sonar-scanner \
-      -Dsonar.projectKey=RED_$SERVICE_NAME \
-      -Dsonar.sources=image_prediction \
-      -Dsonar.host.url=https://sonarqube.iqser.com \
-      -Dsonar.login=${bamboo_sonarqube_api_token_secret} \
-      -Dsonar.branch.name=${bamboo_planRepository_1_branch} \
-      -Dsonar.dependencyCheck.jsonReportPath=reports/dependency-check-report.json \
-      -Dsonar.dependencyCheck.xmlReportPath=reports/dependency-check-report.xml \
-      -Dsonar.dependencyCheck.htmlReportPath=reports/dependency-check-report.html \
-      -Dsonar.python.coverage.reportPaths=reports/coverage.xml
- 
-else
-    echo "Sonar Scan for PR with key1: ${bamboo_repository_pr_key}"
-    /usr/bin/sonar-scanner/bin/sonar-scanner \
-      -Dsonar.projectKey=RED_$SERVICE_NAME \
-      -Dsonar.sources=image_prediction \
-      -Dsonar.host.url=https://sonarqube.iqser.com \
-      -Dsonar.login=${bamboo_sonarqube_api_token_secret} \
-      -Dsonar.pullrequest.key=${bamboo_repository_pr_key} \
-      -Dsonar.pullrequest.branch=${bamboo_repository_pr_sourceBranch} \
-      -Dsonar.pullrequest.base=${bamboo_repository_pr_targetBranch} \
-      -Dsonar.dependencyCheck.jsonReportPath=reports/dependency-check-report.json \
-      -Dsonar.dependencyCheck.xmlReportPath=reports/dependency-check-report.xml \
-      -Dsonar.dependencyCheck.htmlReportPath=reports/dependency-check-report.html \
-      -Dsonar.python.coverage.reportPaths=reports/coverage.xml
-fi

bamboo-specs/src/test/java/buildjob/PlanSpecTest.java

@@ -1,16 +0,0 @@
-package buildjob;
-
-
-import com.atlassian.bamboo.specs.api.builders.plan.Plan;
-import com.atlassian.bamboo.specs.api.exceptions.PropertiesValidationException;
-import com.atlassian.bamboo.specs.api.util.EntityPropertiesBuilders;
-import org.junit.Test;
-
-public class PlanSpecTest {
-    @Test
-    public void checkYourPlanOffline() throws PropertiesValidationException {
-        Plan plan = new PlanSpec().createDockerBuildPlan();
-
-        EntityPropertiesBuilders.build(plan);
-    }
-}

banner.txt

@@ -0,0 +1,11 @@
++----------------------------------------------------+
+|                                  ___               |
+|                               __/_  `.  .-"""-.    |
+|_._     _,-'""`-._             \_,` | \-'  /   )`-')|
+|(,-.`._,'(       |\`-/|         "") `"`    \  ((`"` |
+|    `-.-' \ )-`( , o o)        ___Y  ,    .'7 /|    |
+|          `-    \`_`"'-       (_,___/...-` (_/_/    |
+|                                                    |
++----------------------------------------------------+
+|            Image Classification Service            |
++----------------------------------------------------+

config.yaml

@@ -1,28 +0,0 @@
-webserver:
-  host: $SERVER_HOST|"127.0.0.1"  # webserver address
-  port: $SERVER_PORT|5000  # webserver port
-  mode: $SERVER_MODE|production  # webserver mode: {development, production}
-
-service:
-  logging_level: $LOGGING_LEVEL_ROOT|DEBUG  # Logging level for service logger
-  batch_size: $BATCH_SIZE|32  # Number of images in memory simultaneously
-  verbose: $VERBOSE|True  # Service prints document processing progress to stdout
-  run_id: $RUN_ID|fabfb1f192c745369b88cab34471aba7  # The ID of the mlflow run to load the model from
-
-
-# These variables control filters that are applied to either images, image metadata or model predictions. The filter
-# result values are reported in the service responses. For convenience the response to a request contains a
-# "filters.allPassed" field, which is set to false if any of the filters returned values did not meet its specified
-# required value.
-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
-
-  image_width_to_height_quotient:  # Image width to height ratio
-    min: $MIN_IMAGE_FORMAT|0.1  # Minimum permissible
-    max: $MAX_IMAGE_FORMAT|10  # Maximum permissible
-
-  min_confidence: $MIN_CONFIDENCE|0.5  # Minimum permissible prediction confidence
-

config/pyinfra.toml

@@ -0,0 +1,68 @@
+
+[asyncio]
+max_concurrent_tasks = 10
+
+[dynamic_tenant_queues]
+enabled = true
+
+[metrics.prometheus]
+enabled = true
+prefix = "redactmanager_image_service"
+
+[tracing]
+enabled = true
+# possible values "opentelemetry" | "azure_monitor" (Excpects APPLICATIONINSIGHTS_CONNECTION_STRING environment variable.)
+type = "azure_monitor" 
+
+[tracing.opentelemetry]
+endpoint = "http://otel-collector-opentelemetry-collector.otel-collector:4318/v1/traces"
+service_name = "redactmanager_image_service"
+exporter = "otlp"
+
+[webserver]
+host = "0.0.0.0"
+port = 8080
+
+[rabbitmq]
+host = "localhost"
+port = 5672
+username = ""
+password = ""
+heartbeat = 60
+# Has to be a divider of heartbeat, and shouldn't be too big, since only in these intervals queue interactions happen (like receiving new messages)
+# This is also the minimum time the service needs to process a message
+connection_sleep = 5
+input_queue = "request_queue"
+output_queue = "response_queue"
+dead_letter_queue = "dead_letter_queue"
+
+tenant_event_queue_suffix = "_tenant_event_queue"
+tenant_event_dlq_suffix = "_tenant_events_dlq"
+tenant_exchange_name = "tenants-exchange"
+queue_expiration_time = 300000  # 5 minutes in milliseconds
+
+service_request_queue_prefix = "image_request_queue"
+service_request_exchange_name = "image_request_exchange"
+service_response_exchange_name = "image_response_exchange"
+service_dlq_name = "image_dlq"
+
+[storage]
+backend = "s3"
+
+[storage.s3]
+bucket = "redaction"
+endpoint = "http://127.0.0.1:9000"
+key = ""
+secret = ""
+region = "eu-central-1"
+
+[storage.azure]
+container = "redaction"
+connection_string = ""
+
+[storage.tenant_server]
+public_key = ""
+endpoint =  "http://tenant-user-management:8081/internal-api/tenants"
+
+[kubernetes]
+pod_name = "test_pod"

config/settings.toml

@@ -0,0 +1,42 @@
+[logging]
+level = "INFO"
+
+[service]
+# Print document processing progress to stdout
+verbose = false
+batch_size = 6
+image_stiching_tolerance = 1  # in pixels
+mlflow_run_id = "fabfb1f192c745369b88cab34471aba7"
+
+# These variables control filters that are applied to either images, image metadata or service_estimator predictions.
+# The filter result values are reported in the service responses. For convenience the response to a request contains a
+# "filters.allPassed" field, which is set to false if any of the values returned by the filters did not meet its
+# specified required value.
+[filters.confidence]
+# Minimum permissible prediction confidence
+min = 0.5
+
+# Image size to page size ratio (ratio of geometric means of areas)
+[filters.image_to_page_quotient]
+min = 0.05
+max = 0.75
+
+[filters.is_scanned_page]
+# Minimum permissible image to page ratio tolerance for a page to be considered scanned.
+# This is only used for filtering small images on scanned pages and is applied before processing the image, therefore
+# superseding the image_to_page_quotient filter that only applies a tag to the image after processing.
+tolerance = 0
+
+# Image width to height ratio
+[filters.image_width_to_height_quotient]
+min = 0.1
+max = 10
+
+# put class specific filters here ['signature', 'formula', 'logo']
+[filters.overrides.signature.image_to_page_quotient]
+max = 0.4
+
+[filters.overrides.logo.image_to_page_quotient]
+min = 0.06
+
+

data/.gitignore

@@ -0,0 +1 @@
+/mlruns

data/base_weights.h5.dvc

@@ -1,4 +0,0 @@
-outs:
-- md5: 6d0186c1f25e889d531788f168fa6cf0
-  size: 16727296
-  path: base_weights.h5

data/mlruns.dvc

@@ -1,5 +1,5 @@
 outs:
-- md5: d1c708270bab6fcd344d4a8b05d1103d.dir
-  size: 150225383
-  nfiles: 178
+- md5: ad061d607f615afc149643f62dbf37cc.dir
+  size: 166952700
+  nfiles: 179
   path: mlruns

doc/tests.drawio

@@ -0,0 +1 @@
+<mxfile host="app.diagrams.net" modified="2022-03-17T15:35:10.371Z" agent="5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/99.0.4844.51 Safari/537.36" etag="b-CbBXg6FXQ9T3Px-oLc" version="17.1.1" type="device"><diagram id="tS3WR_Pr6QhNVK3FqSUP" name="Page-1">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</diagram></mxfile>

image_prediction/config.py

@@ -1,40 +0,0 @@
-"""Implements a config object with dot-indexing syntax."""
-
-
-from envyaml import EnvYAML
-
-from image_prediction.locations import CONFIG_FILE
-
-
-def _get_item_and_maybe_make_dotindexable(container, item):
-    ret = container[item]
-    return DotIndexable(ret) if isinstance(ret, dict) else ret
-
-
-class DotIndexable:
-    def __init__(self, x):
-        self.x = x
-
-    def __getattr__(self, item):
-        return _get_item_and_maybe_make_dotindexable(self.x, item)
-
-    def __setitem__(self, key, value):
-        self.x[key] = value
-
-    def __repr__(self):
-        return self.x.__repr__()
-
-
-class Config:
-    def __init__(self, config_path):
-        self.__config = EnvYAML(config_path)
-
-    def __getattr__(self, item):
-        if item in self.__config:
-            return _get_item_and_maybe_make_dotindexable(self.__config, item)
-
-    def __getitem__(self, item):
-        return self.__getattr__(item)
-
-
-CONFIG = Config(CONFIG_FILE)

image_prediction/locations.py

@@ -1,14 +0,0 @@
-from os import path
-
-MODULE_DIR = path.dirname(path.abspath(__file__))
-PACKAGE_ROOT_DIR = path.dirname(MODULE_DIR)
-REPO_ROOT_DIR = path.dirname(path.dirname(PACKAGE_ROOT_DIR))
-
-DOCKER_COMPOSE_FILE = path.join(REPO_ROOT_DIR, "docker-compose.yaml")
-
-CONFIG_FILE = path.join(PACKAGE_ROOT_DIR, "config.yaml")
-LOG_FILE = "/tmp/log.log"
-
-DATA_DIR = path.join(PACKAGE_ROOT_DIR, "data")
-MLRUNS_DIR = path.join(DATA_DIR, "mlruns")
-BASE_WEIGHTS = path.join(DATA_DIR, "base_weights.h5")

image_prediction/predictor.py

@@ -1,116 +0,0 @@
-import logging
-from itertools import chain
-from operator import itemgetter
-from typing import List, Dict, Iterable
-
-import numpy as np
-
-from image_prediction.config import CONFIG
-from image_prediction.locations import MLRUNS_DIR, BASE_WEIGHTS
-from incl.redai_image.redai.redai.backend.model.model_handle import ModelHandle
-from incl.redai_image.redai.redai.backend.pdf.image_extraction import extract_and_stitch
-from incl.redai_image.redai.redai.utils.mlflow_reader import MlflowModelReader
-from incl.redai_image.redai.redai.utils.shared import chunk_iterable
-
-
-class Predictor:
-    """`ModelHandle` wrapper. Forwards to wrapped model handle for prediction and produces structured output that is
-    interpretable independently of the wrapped model (e.g. with regard to a .classes_ attribute).
-    """
-
-    def __init__(self, model_handle: ModelHandle = None):
-        """Initializes a ServiceEstimator.
-
-        Args:
-            model_handle: ModelHandle object to forward to for prediction. By default, a model handle is loaded from the
-                mlflow database via CONFIG.service.run_id.
-        """
-        try:
-            if model_handle is None:
-                reader = MlflowModelReader(run_id=CONFIG.service.run_id, mlruns_dir=MLRUNS_DIR)
-                self.model_handle = reader.get_model_handle(BASE_WEIGHTS)
-            else:
-                self.model_handle = model_handle
-
-            self.classes = self.model_handle.model.classes_
-            self.classes_readable = np.array(self.model_handle.classes)
-            self.classes_readable_aligned = self.classes_readable[self.classes[list(range(len(self.classes)))]]
-        except Exception as e:
-            logging.info(f"Service estimator initialization failed: {e}")
-
-    def __make_predictions_human_readable(self, probs: np.ndarray) -> List[Dict[str, float]]:
-        """Translates an n x m matrix of probabilities over classes into an n-element list of mappings from classes to
-        probabilities.
-
-        Args:
-            probs: probability matrix (items x classes)
-
-        Returns:
-            list of mappings from classes to probabilities.
-        """
-        classes = np.argmax(probs, axis=1)
-        classes = self.classes[classes]
-        classes_readable = [self.model_handle.classes[c] for c in classes]
-        return classes_readable
-
-    def predict(self, images: List, probabilities: bool = False, **kwargs):
-        """Gathers predictions for list of images. Assigns each image a class and optionally a probability distribution
-        over all classes.
-
-        Args:
-            images (List[PIL.Image]) : Images to gather predictions for.
-            probabilities: Whether to return dictionaries of the following form instead of strings:
-                {
-                    "class": predicted class,
-                    "probabilities": {
-                        "class 1" : class 1 probability,
-                        "class 2" : class 2 probability,
-                         ...
-                    }
-                }
-
-        Returns:
-            By default the return value is a list of classes (meaningful class name strings). Alternatively a list of
-            dictionaries with an additional probability field for estimated class probabilities per image can be
-            returned.
-        """
-        X = self.model_handle.prep_images(list(images))
-
-        probs_per_item = self.model_handle.model.predict_proba(X, **kwargs).astype(float)
-        classes = self.__make_predictions_human_readable(probs_per_item)
-
-        class2prob_per_item = [dict(zip(self.classes_readable_aligned, probs)) for probs in probs_per_item]
-        class2prob_per_item = [
-            dict(sorted(c2p.items(), key=itemgetter(1), reverse=True)) for c2p in class2prob_per_item
-        ]
-
-        predictions = [{"class": c, "probabilities": c2p} for c, c2p in zip(classes, class2prob_per_item)]
-
-        return predictions if probabilities else classes
-
-
-def extract_image_metadata_pairs(pdf_path: str, **kwargs):
-    def image_is_large_enough(metadata: dict):
-        x1, x2, y1, y2 = itemgetter("x1", "x2", "y1", "y2")(metadata)
-
-        return abs(x1 - x2) > 2 and abs(y1 - y2) > 2
-
-    yield from extract_and_stitch(pdf_path, convert_to_rgb=True, filter_fn=image_is_large_enough, **kwargs)
-
-
-def classify_images(predictor, image_metadata_pairs: Iterable, batch_size: int = CONFIG.service.batch_size):
-    def process_chunk(chunk):
-        images, metadata = zip(*chunk)
-        predictions = predictor.predict(images, probabilities=True)
-        return predictions, metadata
-
-    def predict(image_metadata_pair_generator):
-        chunks = chunk_iterable(image_metadata_pair_generator, n=batch_size)
-        return map(chain.from_iterable, zip(*map(process_chunk, chunks)))
-
-    try:
-        predictions, metadata = predict(image_metadata_pairs)
-        return predictions, metadata
-
-    except ValueError:
-        return [], []

image_prediction/response.py

@@ -1,71 +0,0 @@
-"""Defines functions for constructing service responses."""
-
-
-from itertools import starmap
-from operator import itemgetter
-
-import numpy as np
-
-from image_prediction.config import CONFIG
-
-
-def build_response(predictions: list, metadata: list) -> list:
-    return list(starmap(build_image_info, zip(predictions, metadata)))
-
-
-def build_image_info(prediction: dict, metadata: dict) -> dict:
-    def compute_geometric_quotient():
-        page_area_sqrt = np.sqrt(abs(page_width * page_height))
-        image_area_sqrt = np.sqrt(abs(x2 - x1) * abs(y2 - y1))
-        return image_area_sqrt / page_area_sqrt
-
-    page_width, page_height, x1, x2, y1, y2, width, height = itemgetter(
-        "page_width", "page_height", "x1", "x2", "y1", "y2", "width", "height"
-    )(metadata)
-
-    quotient = compute_geometric_quotient()
-
-    min_image_to_page_quotient_breached = bool(quotient < CONFIG.filters.image_to_page_quotient.min)
-    max_image_to_page_quotient_breached = bool(quotient > CONFIG.filters.image_to_page_quotient.max)
-    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
-    )
-
-    min_confidence_breached = bool(max(prediction["probabilities"].values()) < CONFIG.filters.min_confidence)
-    prediction["label"] = prediction.pop("class")  # "class" as field name causes problem for Java objectmapper
-    prediction["probabilities"] = {klass: np.round(prob, 6) for klass, prob in prediction["probabilities"].items()}
-
-    image_info = {
-        "classification": prediction,
-        "position": {"x1": x1, "x2": x2, "y1": y1, "y2": y2, "pageNumber": metadata["page_idx"] + 1},
-        "geometry": {"width": width, "height": height},
-        "filters": {
-            "geometry": {
-                "imageSize": {
-                    "quotient": quotient,
-                    "tooLarge": max_image_to_page_quotient_breached,
-                    "tooSmall": min_image_to_page_quotient_breached,
-                },
-                "imageFormat": {
-                    "quotient": width / height,
-                    "tooTall": min_image_width_to_height_quotient_breached,
-                    "tooWide": max_image_width_to_height_quotient_breached,
-                },
-            },
-            "probability": {"unconfident": min_confidence_breached},
-            "allPassed": not any(
-                [
-                    max_image_to_page_quotient_breached,
-                    min_image_to_page_quotient_breached,
-                    min_image_width_to_height_quotient_breached,
-                    max_image_width_to_height_quotient_breached,
-                    min_confidence_breached,
-                ]
-            ),
-        },
-    }
-
-    return image_info

incl/redai_image

@@ -1 +0,0 @@
-Subproject commit 4c3b26d7673457aaa99e0663dad6950cd36da967

pyproject.toml

@@ -0,0 +1,73 @@
+[tool.poetry]
+name = "image-classification-service"
+version = "2.17.0"
+description = ""
+authors = ["Team Research <research@knecon.com>"]
+readme = "README.md"
+packages = [{ include = "image_prediction", from = "src" }]
+
+[tool.poetry.dependencies]
+python = ">=3.10,<3.11"
+# FIXME: This should be recent pyinfra, but the recent protobuf packages are not compatible with tensorflow 2.9.0, also
+#  see RED-9948.
+pyinfra = { version = "3.4.2", source = "gitlab-research" }
+kn-utils = { version = ">=0.4.0", source = "gitlab-research" }
+dvc = "^2.34.0"
+dvc-ssh = "^2.20.0"
+dvc-azure = "^2.21.2"
+Flask = "^2.1.1"
+requests = "^2.27.1"
+iteration-utilities = "^0.11.0"
+waitress = "^2.1.1"
+envyaml = "^1.10.211231"
+dependency-check = "^0.6.0"
+mlflow = "^1.24.0"
+numpy = "^1.22.3"
+tqdm = "^4.64.0"
+pandas = "^1.4.2"
+# FIXME: Our current model significantly changes the prediction behaviour when using newer tensorflow (/ protobuf)
+#  versions which is introduuced by pyinfra updates using newer protobuf versions, see RED-9948.
+tensorflow = "2.9.0"
+protobuf = "^3.20"
+pytest = "^7.1.0"
+funcy = "^2"
+PyMuPDF = "^1.19.6"
+fpdf = "^1.7.2"
+coverage = "^6.3.2"
+Pillow = "^9.1.0"
+pdf2image = "^1.16.0"
+frozendict = "^2.3.0"
+fsspec = "^2022.11.0"
+PyMonad = "^2.4.0"
+pdfnetpython3 = "9.4.2"
+loguru = "^0.7.0"
+cyclonedx-bom = "^4.5.0"
+
+[tool.poetry.group.dev.dependencies]
+pytest = "^7.0.1"
+pymonad = "^2.4.0"
+pylint = "^2.17.4"
+ipykernel = "^6.23.2"
+
+[tool.pytest.ini_options]
+testpaths = ["test"]
+addopts = "--ignore=data"
+filterwarnings = ["ignore:.*:DeprecationWarning"]
+
+[[tool.poetry.source]]
+name = "PyPI"
+priority = "primary"
+
+[[tool.poetry.source]]
+name = "gitlab-research"
+url = "https://gitlab.knecon.com/api/v4/groups/19/-/packages/pypi/simple"
+priority = "explicit"
+
+[[tool.poetry.source]]
+name = "gitlab-red"
+url = "https://gitlab.knecon.com/api/v4/groups/12/-/packages/pypi/simple"
+priority = "explicit"
+
+[build-system]
+requires = ["poetry-core"]
+build-backend = "poetry.core.masonry.api"

requirements.txt

@@ -1,21 +0,0 @@
-Flask==2.0.2
-requests==2.27.1
-iteration-utilities==0.11.0
-dvc==2.9.3
-dvc[ssh]
-frozendict==2.3.0
-waitress==2.0.0
-envyaml~=1.8.210417
-dependency-check==0.6.*
-envyaml~=1.8.210417
-mlflow~=1.20.2
-numpy~=1.19.3
-PDFNetPython3~=9.1.0
-tqdm~=4.62.2
-pandas~=1.3.1
-mlflow~=1.20.2
-tensorflow~=2.5.0
-PDFNetPython3~=9.1.0
-Pillow~=8.3.2
-PyYAML~=5.4.1
-scikit_learn~=0.24.2

scripts/debug/debug.py

@@ -0,0 +1,46 @@
+"""Script to debug RED-9948. The predictions unexpectedly changed for some images, and we need to understand why."""
+
+import json
+import random
+from pathlib import Path
+
+import numpy as np
+import tensorflow as tf
+from kn_utils.logging import logger
+
+from image_prediction.config import CONFIG
+from image_prediction.pipeline import load_pipeline
+
+
+def process_pdf(pipeline, pdf_path, page_range=None):
+    with open(pdf_path, "rb") as f:
+        logger.info(f"Processing {pdf_path}")
+        predictions = list(pipeline(f.read(), page_range=page_range))
+
+    return predictions
+
+
+def ensure_seeds():
+    seed = 42
+    np.random.seed(seed)
+    random.seed(seed)
+    tf.random.set_seed(seed)
+
+
+def debug_info():
+    devices = tf.config.list_physical_devices()
+    print("Available devices:", devices)
+
+
+if __name__ == "__main__":
+    # For in container debugging, copy the file and adjust the path.
+    debug_file_path = Path(__file__).parents[2] / "test" / "data" / "RED-9948" / "SYNGENTA_EFSA_sanitisation_GFL_v2"
+    ensure_seeds()
+    debug_info()
+
+    pipeline = load_pipeline(verbose=True, batch_size=CONFIG.service.batch_size)
+    predictions = process_pdf(pipeline, debug_file_path)
+    # This is the image that has the wrong prediction mentioned in RED-9948. The predictions should inconclusive, and
+    # the flag all passed should be false.
+    predictions = [x for x in predictions if x["representation"] == "FA30F080F0C031CE17E8CF237"]
+    print(json.dumps(predictions, indent=2))

scripts/devenvsetup.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+python_version=$1
+gitlab_user=$2
+gitlab_personal_access_token=$3
+
+# cookiecutter https://gitlab.knecon.com/knecon/research/template-python-project.git --checkout master
+# latest_dir=$(ls -td -- */ | head -n 1)  # should be the dir cookiecutter just created
+
+# cd $latest_dir
+
+pyenv install $python_version
+pyenv local $python_version
+pyenv shell $python_version
+
+pip install --upgrade pip
+pip install poetry
+
+poetry config installer.max-workers 10
+# research package registry
+poetry config repositories.gitlab-research https://gitlab.knecon.com/api/v4/groups/19/-/packages/pypi
+poetry config http-basic.gitlab-research ${gitlab_user} ${gitlab_personal_access_token}
+# redactmanager package registry
+poetry config repositories.gitlab-red https://gitlab.knecon.com/api/v4/groups/12/-/packages/pypi
+poetry config http-basic.gitlab-red ${gitlab_user} ${gitlab_personal_access_token}
+
+poetry env use $(pyenv which python)
+poetry install --with=dev
+poetry update
+
+source .venv/bin/activate

scripts/docker_build_run.sh

@@ -0,0 +1,6 @@
+docker build -t --platform linux/amd64 image-clsasification-service:$(poetry version -s)-dev \                                                               
+    -f Dockerfile \
+    --build-arg GITLAB_USER=$GITLAB_USER \
+    --build-arg GITLAB_ACCESS_TOKEN=$GITLAB_ACCESS_TOKEN \
+    . && \
+docker run -it --rm image-clsasification-service:$(poetry version -s)-dev

scripts/docker_tag_push.sh

@@ -0,0 +1,3 @@
+docker tag image-clsasification-service:$(poetry version -s)-dev $NEXUS_REGISTRY/red/image-clsasification-service:$(poetry version -s)-dev
+
+docker push $NEXUS_REGISTRY/red/image-clsasification-service:$(poetry version -s)-dev

scripts/k8s_startup_probe.py

@@ -0,0 +1,6 @@
+from pyinfra.k8s_probes import startup
+from loguru import logger
+
+if __name__ == "__main__":
+    logger.debug("running health check")
+    startup.run_checks()

scripts/keras_MnWE.py

@@ -0,0 +1,58 @@
+import multiprocessing
+
+import numpy as np
+from tensorflow import keras
+from tensorflow.keras import layers
+
+
+def process(predict_fn_wrapper):
+    # We observed memory doesn't get properly deallocated unless we do this:
+    manager = multiprocessing.Manager()
+    return_dict = manager.dict()
+    p = multiprocessing.Process(
+        target=predict_fn_wrapper,
+        args=(return_dict,),
+    )
+    p.start()
+    p.join()
+    try:
+        return dict(return_dict)["result"]
+    except KeyError:
+        pass
+
+
+def make_model():
+    inputs = keras.Input(shape=(784,))
+    dense = layers.Dense(64, activation="relu")
+    x = dense(inputs)
+    outputs = layers.Dense(10)(x)
+    model = keras.Model(inputs=inputs, outputs=outputs, name="mnist_model")
+    model.compile(
+        loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
+        optimizer=keras.optimizers.RMSprop(),
+        metrics=["accuracy"],
+    )
+    return model
+
+
+def make_predict_fn():
+    # Keras bug: doesn't work in outer scope
+    model = make_model()
+
+    def predict(*args):
+        # service_estimator = make_model()
+        return model.predict(np.random.random(size=(1, 784)))
+
+    return predict
+
+
+def make_predict_fn_wrapper(predict_fn):
+    def predict_fn_wrapper(return_dict):
+        return_dict["result"] = predict_fn()
+
+    return predict_fn_wrapper
+
+
+if __name__ == "__main__":
+    predict_fn = make_predict_fn()
+    print(process(make_predict_fn_wrapper(predict_fn)))

scripts/pyinfra_mock.py

@@ -6,7 +6,7 @@ import requests
 
 def parse_args():
     parser = argparse.ArgumentParser()
-    parser.add_argument("--pdf_path", required=True)
+    parser.add_argument("pdf_path")
     args = parser.parse_args()
 
     return args

scripts/run_pipeline.py

@@ -0,0 +1,58 @@
+import argparse
+import json
+import os
+from glob import glob
+
+from image_prediction.config import CONFIG
+from image_prediction.pipeline import load_pipeline
+from image_prediction.utils import get_logger
+from image_prediction.utils.pdf_annotation import annotate_pdf
+
+logger = get_logger()
+
+logger.setLevel("DEBUG")
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+
+    parser.add_argument("input", help="pdf file or directory")
+    parser.add_argument("--print", "-p", help="print output to terminal", action="store_true", default=False)
+    parser.add_argument("--page_interval", "-i", help="page interval [i, j), min index = 0", nargs=2, type=int)
+
+    args = parser.parse_args()
+
+    return args
+
+
+def process_pdf(pipeline, pdf_path, page_range=None):
+    with open(pdf_path, "rb") as f:
+        logger.info(f"Processing {pdf_path}")
+        predictions = list(pipeline(f.read(), page_range=page_range))
+
+    annotate_pdf(
+        pdf_path, predictions, os.path.join("/tmp", os.path.basename(pdf_path.replace(".pdf", "_annotated.pdf")))
+    )
+
+    return predictions
+
+
+def main(args):
+    pipeline = load_pipeline(verbose=CONFIG.service.verbose, batch_size=CONFIG.service.batch_size, tolerance=CONFIG.service.image_stiching_tolerance)
+
+    if os.path.isfile(args.input):
+        pdf_paths = [args.input]
+    else:
+        pdf_paths = glob(os.path.join(args.input, "*.pdf"))
+    page_range = range(*args.page_interval) if args.page_interval else None
+
+    for pdf_path in pdf_paths:
+        predictions = process_pdf(pipeline, pdf_path, page_range=page_range)
+        if args.print:
+            print(pdf_path)
+            print(json.dumps(predictions, indent=2))
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

scripts/run_tests.sh

@@ -0,0 +1,15 @@
+echo "${bamboo_nexus_password}" | docker login --username "${bamboo_nexus_user}" --password-stdin nexus.iqser.com:5001
+
+pip install dvc
+pip install 'dvc[ssh]'
+echo "Pulling dvc data"
+dvc pull
+
+docker build -f Dockerfile_tests -t image-prediction-tests .
+
+rnd=$(date +"%s")
+name=image-prediction-tests-${rnd}
+
+echo "running tests container"
+
+docker run --rm --name $name -v $PWD:$PWD -w $PWD -v /var/run/docker.sock:/var/run/docker.sock image-prediction-tests

setup.py

@@ -1,13 +0,0 @@
-#!/usr/bin/env python
-
-from distutils.core import setup
-
-setup(
-    name="image_prediction",
-    version="0.1.0",
-    description="",
-    author="",
-    author_email="",
-    url="",
-    packages=["image_prediction"],
-)

setup/docker.sh

@@ -1,15 +0,0 @@
-#!/bin/bash
-set -e
-
-python3 -m venv build_venv
-source build_venv/bin/activate
-python3 -m pip install --upgrade pip
-
-pip install dvc
-pip install 'dvc[ssh]'
-dvc pull
-
-git submodule update --init --recursive
-
-docker build -f Dockerfile_base -t image-prediction-base .
-docker build -f Dockerfile -t image-prediction .

sonar-project.properties

@@ -1,4 +0,0 @@
-sonar.exclusions=bamboo-specs/**, **/test_data/**
-sonar.c.file.suffixes=-
-sonar.cpp.file.suffixes=-
-sonar.objc.file.suffixes=-

src/image_prediction/__init__.py

@@ -0,0 +1,13 @@
+import logging
+import sys
+
+# log config
+LOG_FORMAT = "%(asctime)s [%(levelname)s] - [%(filename)s -> %(funcName)s() -> %(lineno)s] : %(message)s"
+DATE_FORMAT = "%Y-%m-%d %H:%M:%S"
+stream_handler = logging.StreamHandler(sys.stdout)
+stream_handler_format = logging.Formatter(LOG_FORMAT, datefmt=DATE_FORMAT)
+stream_handler.setFormatter(stream_handler_format)
+
+logger = logging.getLogger(__name__)
+logger.propagate = False
+logger.addHandler(stream_handler)

src/image_prediction/classifier/classifier.py

@@ -0,0 +1,35 @@
+from typing import List, Union, Tuple
+
+import numpy as np
+from PIL.Image import Image
+from funcy import rcompose
+
+from image_prediction.estimator.adapter.adapter import EstimatorAdapter
+from image_prediction.label_mapper.mapper import LabelMapper
+from image_prediction.utils import get_logger
+
+logger = get_logger()
+
+
+class Classifier:
+    def __init__(self, estimator_adapter: EstimatorAdapter, label_mapper: LabelMapper):
+        """Abstraction layer over different estimator backends (e.g. keras or scikit-learn). For each backend to be used
+        an EstimatorAdapter must be implemented.
+
+        Args:
+            estimator_adapter: adapter for a given estimator backend
+        """
+        self.__estimator_adapter = estimator_adapter
+        self.__label_mapper = label_mapper
+        self.__pipe = rcompose(self.__estimator_adapter, self.__label_mapper)
+
+    def predict(self, batch: Union[np.array, Tuple[Image]]) -> List[str]:
+
+        if isinstance(batch, np.ndarray) and batch.shape[0] == 0:
+            return []
+
+        return self.__pipe(batch)
+
+    def __call__(self, batch: np.array) -> List[str]:
+        logger.debug("Classifier.predict")
+        return self.predict(batch)

src/image_prediction/classifier/image_classifier.py

@@ -0,0 +1,32 @@
+from itertools import chain
+from typing import Iterable
+
+from PIL.Image import Image
+from funcy import rcompose, chunks
+
+from image_prediction.classifier.classifier import Classifier
+from image_prediction.estimator.preprocessor.preprocessor import Preprocessor
+from image_prediction.estimator.preprocessor.preprocessors.identity import IdentityPreprocessor
+from image_prediction.utils import get_logger
+
+logger = get_logger()
+
+
+class ImageClassifier:
+    """Combines a classifier with a preprocessing pipeline: Receives images, chunks into batches, converts to tensors,
+    applies transformations and finally sends to internal classifier.
+    """
+
+    def __init__(self, classifier: Classifier, preprocessor: Preprocessor = None):
+        self.estimator = classifier
+        self.preprocessor = preprocessor if preprocessor else IdentityPreprocessor()
+        self.pipe = rcompose(self.preprocessor, self.estimator)
+
+    def predict(self, images: Iterable[Image], batch_size=16):
+        batches = chunks(batch_size, images)
+        predictions = chain.from_iterable(map(self.pipe, batches))
+        return predictions
+
+    def __call__(self, images: Iterable[Image], batch_size=16):
+        logger.debug("ImageClassifier.predict")
+        yield from self.predict(images, batch_size=batch_size)

src/image_prediction/compositor/compositor.py

@@ -0,0 +1,16 @@
+from funcy import rcompose
+
+from image_prediction.transformer.transformer import Transformer
+from image_prediction.utils import get_logger
+
+logger = get_logger()
+
+
+class TransformerCompositor(Transformer):
+    def __init__(self, formatter: Transformer, *formatters: Transformer):
+        formatters = (formatter, *formatters)
+        self.pipe = rcompose(*formatters)
+
+    def transform(self, obj):
+        logger.debug("TransformerCompositor.transform")
+        return self.pipe(obj)

src/image_prediction/config.py

@@ -0,0 +1,7 @@
+from pathlib import Path
+
+from pyinfra.config.loader import load_settings
+
+from image_prediction.locations import PROJECT_ROOT_DIR
+
+CONFIG = load_settings(root_path=PROJECT_ROOT_DIR, settings_path="config")

src/image_prediction/default_objects.py

@@ -0,0 +1,43 @@
+from funcy import juxt
+
+from image_prediction.classifier.classifier import Classifier
+from image_prediction.classifier.image_classifier import ImageClassifier
+from image_prediction.compositor.compositor import TransformerCompositor
+from image_prediction.encoder.encoders.hash_encoder import HashEncoder
+from image_prediction.estimator.adapter.adapter import EstimatorAdapter
+from image_prediction.formatter.formatters.camel_case import Snake2CamelCaseKeyFormatter
+from image_prediction.formatter.formatters.enum import EnumFormatter
+from image_prediction.image_extractor.extractors.parsable import ParsablePDFImageExtractor
+from image_prediction.label_mapper.mappers.probability import ProbabilityMapper
+from image_prediction.model_loader.loader import ModelLoader
+from image_prediction.model_loader.loaders.mlflow import MlflowConnector
+from image_prediction.redai_adapter.mlflow import MlflowModelReader
+from image_prediction.transformer.transformers.coordinate.pdfnet import PDFNetCoordinateTransformer
+from image_prediction.transformer.transformers.response import ResponseTransformer
+
+
+def get_mlflow_model_loader(mlruns_dir):
+    model_loader = ModelLoader(MlflowConnector(MlflowModelReader(mlruns_dir)))
+    return model_loader
+
+
+def get_image_classifier(model_loader, model_identifier):
+    model, classes = juxt(model_loader.load_model, model_loader.load_classes)(model_identifier)
+    return ImageClassifier(Classifier(EstimatorAdapter(model), ProbabilityMapper(classes)))
+
+
+def get_extractor(**kwargs):
+    image_extractor = ParsablePDFImageExtractor(**kwargs)
+
+    return image_extractor
+
+
+def get_formatter():
+    formatter = TransformerCompositor(
+        PDFNetCoordinateTransformer(), EnumFormatter(), ResponseTransformer(), Snake2CamelCaseKeyFormatter()
+    )
+    return formatter
+
+
+def get_encoder():
+    return HashEncoder()

src/image_prediction/encoder/encoder.py

@@ -0,0 +1,13 @@
+import abc
+from typing import Iterable
+
+from PIL.Image import Image
+
+
+class Encoder(abc.ABC):
+    @abc.abstractmethod
+    def encode(self, images: Iterable[Image]):
+        raise NotImplementedError
+
+    def __call__(self, images: Iterable[Image], batch_size=16):
+        yield from self.encode(images)

src/image_prediction/encoder/encoders/hash_encoder.py

@@ -0,0 +1,26 @@
+from typing import Iterable
+
+from PIL import Image
+
+from image_prediction.encoder.encoder import Encoder
+
+
+class HashEncoder(Encoder):
+    def encode(self, images: Iterable[Image.Image]):
+        yield from map(hash_image, images)
+
+    def __call__(self, images: Iterable[Image.Image], batch_size=16):
+        yield from self.encode(images)
+
+
+def hash_image(image: Image.Image) -> str:
+    """See: https://stackoverflow.com/a/49692185/3578468"""
+    image = image.resize((10, 10), Image.ANTIALIAS)
+    image = image.convert("L")
+    pixel_data = list(image.getdata())
+    avg_pixel = sum(pixel_data) / len(pixel_data)
+    bits = "".join(["1" if (px >= avg_pixel) else "0" for px in pixel_data])
+    hex_representation = str(hex(int(bits, 2)))[2:][::-1].upper()
+    # Note: For each 4 leading zeros, the hex representation will be shorter by one character.
+    # To ensure that all hashes have the same length, we pad the hex representation with zeros (also see RED-3813).
+    return hex_representation.zfill(25)

src/image_prediction/estimator/adapter/adapter.py

@@ -0,0 +1,15 @@
+from image_prediction.utils import get_logger
+
+logger = get_logger()
+
+
+class EstimatorAdapter:
+    def __init__(self, estimator):
+        self.estimator = estimator
+
+    def predict(self, batch):
+        return self.estimator(batch)
+
+    def __call__(self, batch):
+        logger.debug("EstimatorAdapter.predict")
+        return self.predict(batch)

src/image_prediction/estimator/preprocessor/preprocessor.py

@@ -0,0 +1,10 @@
+import abc
+
+
+class Preprocessor(abc.ABC):
+    @abc.abstractmethod
+    def preprocess(self, batch):
+        raise NotImplementedError
+
+    def __call__(self, batch):
+        return self.preprocess(batch)

src/image_prediction/estimator/preprocessor/preprocessors/basic.py

@@ -0,0 +1,10 @@
+from image_prediction.estimator.preprocessor.preprocessor import Preprocessor
+from image_prediction.estimator.preprocessor.utils import images_to_batch_tensor
+
+
+class BasicPreprocessor(Preprocessor):
+    """Converts images to tensors"""
+
+    @staticmethod
+    def preprocess(images):
+        return images_to_batch_tensor(images)

src/image_prediction/estimator/preprocessor/preprocessors/identity.py

@@ -0,0 +1,10 @@
+from image_prediction.estimator.preprocessor.preprocessor import Preprocessor
+
+
+class IdentityPreprocessor(Preprocessor):
+    @staticmethod
+    def preprocess(images):
+        return images
+
+    def __call__(self, images):
+        return self.preprocess(images)

src/image_prediction/estimator/preprocessor/utils.py

@@ -0,0 +1,10 @@
+import numpy as np
+from PIL.Image import Image
+
+
+def image_to_normalized_tensor(image: Image) -> np.ndarray:
+    return np.array(image) / 255
+
+
+def images_to_batch_tensor(images) -> np.ndarray:
+    return np.array(list(map(image_to_normalized_tensor, images)))

src/image_prediction/exceptions.py

@@ -0,0 +1,42 @@
+class UnknownEstimatorAdapter(ValueError):
+    pass
+
+
+class UnknownImageExtractor(ValueError):
+    pass
+
+
+class UnknownModelLoader(ValueError):
+    pass
+
+
+class UnknownDatabaseType(ValueError):
+    pass
+
+
+class UnknownLabelFormat(ValueError):
+    pass
+
+
+class UnexpectedLabelFormat(ValueError):
+    pass
+
+
+class IncorrectInstantiation(RuntimeError):
+    pass
+
+
+class IntentionalTestException(RuntimeError):
+    pass
+
+
+class InvalidBox(Exception):
+    pass
+
+
+class ParsingError(Exception):
+    pass
+
+
+class BadXref(ValueError):
+    pass

src/image_prediction/extraction.py

@@ -0,0 +1,13 @@
+from image_prediction.image_extractor.extractors.parsable import ParsablePDFImageExtractor
+
+
+def extract_images_from_pdf(pdf, extractor=None):
+
+    if not extractor:
+        extractor = ParsablePDFImageExtractor()
+
+    try:
+        images_extracted, metadata_extracted = zip(*extractor(pdf))
+        return images_extracted, metadata_extracted
+    except ValueError:
+        return [], []

src/image_prediction/flask.py

@@ -0,0 +1,60 @@
+from typing import Callable
+
+from flask import Flask, request, jsonify
+from prometheus_client import generate_latest, CollectorRegistry, Summary
+
+from image_prediction.utils import get_logger
+from image_prediction.utils.process_wrapping import wrap_in_process
+
+logger = get_logger()
+
+
+def make_prediction_server(predict_fn: Callable):
+    app = Flask(__name__)
+    registry = CollectorRegistry(auto_describe=True)
+    metric = Summary(
+        f"redactmanager_imageClassification_seconds", f"Time spent on image-service classification.", registry=registry
+    )
+
+    @app.route("/ready", methods=["GET"])
+    def ready():
+        resp = jsonify("OK")
+        resp.status_code = 200
+        return resp
+
+    @app.route("/health", methods=["GET"])
+    def healthy():
+        resp = jsonify("OK")
+        resp.status_code = 200
+        return resp
+
+    def __failure():
+        response = jsonify("Analysis failed")
+        response.status_code = 500
+        return response
+
+    @app.route("/predict", methods=["POST"])
+    @app.route("/", methods=["POST"])
+    @metric.time()
+    def predict():
+
+        # Tensorflow does not free RAM. Workaround: Run prediction function (which instantiates a model) in sub-process.
+        # See: https://stackoverflow.com/questions/39758094/clearing-tensorflow-gpu-memory-after-model-execution
+        predict_fn_wrapped = wrap_in_process(predict_fn)
+
+        logger.info("Analysing...")
+        predictions = predict_fn_wrapped(request.data)
+
+        if predictions is not None:
+            response = jsonify(predictions)
+            logger.info("Analysis completed.")
+            return response
+        else:
+            logger.error("Analysis failed.")
+            return __failure()
+
+    @app.route("/prometheus", methods=["GET"])
+    def prometheus():
+        return generate_latest(registry=registry)
+
+    return app

src/image_prediction/formatter/formatter.py

@@ -0,0 +1,15 @@
+import abc
+
+from image_prediction.transformer.transformer import Transformer
+
+
+class Formatter(Transformer):
+    @abc.abstractmethod
+    def format(self, obj):
+        raise NotImplementedError
+
+    def transform(self, obj):
+        raise NotImplementedError()
+
+    def __call__(self, obj):
+        return self.format(obj)

src/image_prediction/formatter/formatters/camel_case.py

@@ -0,0 +1,11 @@
+from image_prediction.formatter.formatters.key_formatter import KeyFormatter
+
+
+class Snake2CamelCaseKeyFormatter(KeyFormatter):
+    def format_key(self, key):
+
+        if isinstance(key, str):
+            head, *tail = key.split("_")
+            return head + "".join(map(str.title, tail))
+        else:
+            return key

src/image_prediction/formatter/formatters/enum.py

@@ -0,0 +1,23 @@
+from enum import Enum
+
+from image_prediction.formatter.formatters.key_formatter import KeyFormatter
+
+
+class EnumFormatter(KeyFormatter):
+    def format_key(self, key):
+        return key.value if isinstance(key, Enum) else key
+
+    def transform(self, obj):
+        raise NotImplementedError
+
+
+class ReverseEnumFormatter(KeyFormatter):
+    def __init__(self, enum):
+        self.enum = enum
+        self.reverse_enum = {e.value: e for e in enum}
+
+    def format_key(self, key):
+        return self.reverse_enum.get(key, key)
+
+    def transform(self, obj):
+        raise NotImplementedError

src/image_prediction/formatter/formatters/identity.py

@@ -0,0 +1,6 @@
+from image_prediction.formatter.formatter import Formatter
+
+
+class IdentityFormatter(Formatter):
+    def format(self, obj):
+        return obj

src/image_prediction/formatter/formatters/key_formatter.py

@@ -0,0 +1,28 @@
+import abc
+from typing import Iterable
+
+from image_prediction.formatter.formatter import Formatter
+
+
+class KeyFormatter(Formatter):
+    @abc.abstractmethod
+    def format_key(self, key):
+        raise NotImplementedError
+
+    def __format(self, data):
+
+        # If we wanted to do this properly, we would need handlers for all expected types and dispatch based
+        # on a type comparison. This is too much engineering for the limited use-case of this class though.
+        if isinstance(data, Iterable) and not isinstance(data, dict) and not isinstance(data, str):
+            f = map(self.__format, data)
+            return type(data)(f) if not isinstance(data, map) else f
+
+        if not isinstance(data, dict):
+            return data
+
+        keys_formatted = list(map(self.format_key, data))
+
+        return dict(zip(keys_formatted, map(self.__format, data.values())))
+
+    def format(self, data):
+        return self.__format(data)

src/image_prediction/image_extractor/extractor.py

@@ -0,0 +1,19 @@
+import abc
+from collections import namedtuple
+from typing import Iterable
+
+from image_prediction.utils import get_logger
+
+ImageMetadataPair = namedtuple("ImageMetadataPair", ["image", "metadata"])
+
+logger = get_logger()
+
+
+class ImageExtractor(abc.ABC):
+    @abc.abstractmethod
+    def extract(self, obj) -> Iterable[ImageMetadataPair]:
+        raise NotImplementedError
+
+    def __call__(self, obj, **kwargs):
+        logger.debug("ImageExtractor.extract")
+        return self.extract(obj, **kwargs)

src/image_prediction/image_extractor/extractors/mock.py

@@ -0,0 +1,7 @@
+from image_prediction.image_extractor.extractor import ImageExtractor, ImageMetadataPair
+
+
+class ImageExtractorMock(ImageExtractor):
+    def extract(self, image_container):
+        for i, image in enumerate(image_container):
+            yield ImageMetadataPair(image, {"image_id": i})

src/image_prediction/image_extractor/extractors/parsable.py

@@ -0,0 +1,300 @@
+import atexit
+import json
+import traceback
+from _operator import itemgetter
+from functools import partial, lru_cache
+from itertools import chain, starmap, filterfalse, tee
+from operator import itemgetter, truth
+from typing import Iterable, Iterator, List, Union
+
+import fitz
+import numpy as np
+from PIL import Image
+from funcy import merge, pluck, compose, rcompose, remove, keep
+from scipy.stats import gmean
+
+from image_prediction.config import CONFIG
+from image_prediction.exceptions import InvalidBox
+from image_prediction.formatter.formatters.enum import EnumFormatter
+from image_prediction.image_extractor.extractor import ImageExtractor, ImageMetadataPair
+from image_prediction.info import Info
+from image_prediction.stitching.stitching import stitch_pairs
+from image_prediction.stitching.utils import validate_box
+from image_prediction.transformer.transformers.response import compute_geometric_quotient
+from image_prediction.utils import get_logger
+
+logger = get_logger()
+
+
+class ParsablePDFImageExtractor(ImageExtractor):
+    def __init__(self, verbose=False, tolerance=0):
+        """
+
+        Args:
+            verbose: Whether to show progressbar
+            tolerance: The tolerance in pixels for the distance between images, beyond which they will not be stitched
+                together
+        """
+        self.doc: fitz.Document = None
+        self.verbose = verbose
+        self.tolerance = tolerance
+
+    def extract(self, pdf: bytes, page_range: range = None):
+        self.doc = fitz.Document(stream=pdf)
+
+        pages = extract_pages(self.doc, page_range) if page_range else self.doc
+
+        image_metadata_pairs = chain.from_iterable(map(self.__process_images_on_page, pages))
+
+        yield from image_metadata_pairs
+
+    def __process_images_on_page(self, page: fitz.Page):
+        metadata = extract_valid_metadata(self.doc, page)
+        images = get_images_on_page(self.doc, metadata)
+
+        clear_caches()
+
+        image_metadata_pairs = starmap(ImageMetadataPair, filter(all, zip(images, metadata)))
+        #  TODO: In the future, consider to introduce an image validator as a pipeline component rather than doing the
+        #   validation here. Invalid images can then be split into a different stream and joined with the intact images
+        #   again for the formatting step.
+        image_metadata_pairs = self.__filter_valid_images(image_metadata_pairs)
+        image_metadata_pairs = stitch_pairs(list(image_metadata_pairs), tolerance=self.tolerance)
+
+        yield from image_metadata_pairs
+
+    @staticmethod
+    def __filter_valid_images(image_metadata_pairs: Iterable[ImageMetadataPair]) -> Iterator[ImageMetadataPair]:
+        def validate_image_is_not_corrupt(image: Image.Image, metadata: dict):
+            """See RED-5148: Some images are corrupt and cannot be processed by the image classifier. This function
+            filters out such images by trying to resize and convert them to RGB. If this fails, the image is considered
+            corrupt and is dropped.
+            TODO: find cleaner solution
+            """
+            try:
+                image.resize((100, 100)).convert("RGB")
+                return ImageMetadataPair(image, metadata)
+            except (OSError, Exception) as err:
+                metadata = json.dumps(EnumFormatter()(metadata), indent=2)
+                logger.warning(f"Invalid image encountered. Image metadata:\n{metadata}\n\n{traceback.format_exc()}")
+                return None
+
+        def filter_small_images_on_scanned_pages(image_metadata_pairs) -> Iterable[ImageMetadataPair]:
+            """See RED-9746: Small images on scanned pages should be dropped, so they are not classified. This is a
+            heuristic to filter out images that are too small in relation to the page size if they are on a scanned page.
+
+            The ratio is computed as the geometric mean of the width and height of the image divided by the geometric mean
+            of the width and height of the page. If the ratio is below the threshold, the image is dropped.
+            """
+
+            def image_is_a_scanned_page(image_metadata_pair: ImageMetadataPair) -> bool:
+                tolerance = CONFIG.filters.is_scanned_page.tolerance
+                width_ratio = image_metadata_pair.metadata[Info.WIDTH] / image_metadata_pair.metadata[Info.PAGE_WIDTH]
+                height_ratio = (
+                    image_metadata_pair.metadata[Info.HEIGHT] / image_metadata_pair.metadata[Info.PAGE_HEIGHT]
+                )
+                return width_ratio >= 1 - tolerance and height_ratio >= 1 - tolerance
+
+            def image_fits_geometric_mean_ratio(image_metadata_pair: ImageMetadataPair) -> bool:
+                min_ratio = CONFIG.filters.image_to_page_quotient.min
+                metadatum = image_metadata_pair.metadata
+                image_gmean = gmean([metadatum[Info.WIDTH], metadatum[Info.HEIGHT]])
+                page_gmean = gmean([metadatum[Info.PAGE_WIDTH], metadatum[Info.PAGE_HEIGHT]])
+                ratio = image_gmean / page_gmean
+                return ratio >= min_ratio
+
+            pairs, pairs_copy = tee(image_metadata_pairs)
+
+            if any(map(image_is_a_scanned_page, pairs_copy)):
+                logger.debug("Scanned page detected, filtering out small images ...")
+                return filter(image_fits_geometric_mean_ratio, pairs)
+            else:
+                return pairs
+
+        image_metadata_pairs = filter_small_images_on_scanned_pages(image_metadata_pairs)
+
+        return filter(truth, starmap(validate_image_is_not_corrupt, image_metadata_pairs))
+
+
+def extract_pages(doc, page_range):
+    page_range = range(page_range.start + 1, page_range.stop + 1)
+    pages = map(doc.load_page, page_range)
+
+    yield from pages
+
+
+def get_images_on_page(doc, metadata):
+    xrefs = pluck(Info.XREF, metadata)
+    images = map(partial(xref_to_image, doc), xrefs)
+
+    yield from images
+
+
+def extract_valid_metadata(doc: fitz.Document, page: fitz.Page):
+    metadata = get_metadata_for_images_on_page(page)
+    metadata = filter_valid_metadata(metadata)
+    metadata = add_alpha_channel_info(doc, metadata)
+
+    return list(metadata)
+
+
+def get_metadata_for_images_on_page(page: fitz.Page):
+    metadata = map(get_image_metadata, get_image_infos(page))
+    metadata = add_page_metadata(page, metadata)
+
+    yield from metadata
+
+
+def filter_valid_metadata(metadata):
+    yield from compose(
+        # TODO: Disabled for now, since atm since the backend needs atm the metadata and the hash of every image, even
+        #  scanned pages. In the future, this should be resolved differently, e.g. by filtering all page-sized images
+        #  and giving the user the ability to reclassify false positives with a separate call.
+        # filter_out_page_sized_images,
+        filter_out_tiny_images,
+        filter_out_invalid_metadata,
+    )(metadata)
+
+
+def filter_out_invalid_metadata(metadata):
+    def __validate_box(box):
+        try:
+            return validate_box(box)
+        except InvalidBox as err:
+            logger.debug(f"Dropping invalid metadatum, reason: {err}")
+
+    yield from keep(__validate_box, metadata)
+
+
+def filter_out_page_sized_images(metadata):
+    yield from remove(breaches_image_to_page_quotient, metadata)
+
+
+def filter_out_tiny_images(metadata):
+    yield from filterfalse(tiny, metadata)
+
+
+@lru_cache(maxsize=None)
+def get_image_infos(page: fitz.Page) -> List[dict]:
+    return page.get_image_info(xrefs=True)
+
+
+@lru_cache(maxsize=None)
+def xref_to_image(doc, xref) -> Union[Image.Image, None]:
+    # NOTE: image extraction is done via pixmap to array, as this method is twice as fast as extraction via bytestream
+    try:
+        pixmap = fitz.Pixmap(doc, xref)
+        array = convert_pixmap_to_array(pixmap)
+        return Image.fromarray(array)
+    except ValueError:
+        logger.debug(f"Xref {xref} is invalid, skipping extraction ...")
+        return
+
+
+def convert_pixmap_to_array(pixmap: fitz.Pixmap):
+    array = np.frombuffer(pixmap.samples, dtype=np.uint8).reshape(pixmap.h, pixmap.w, pixmap.n)
+    array = _normalize_channels(array)
+    return array
+
+
+def _normalize_channels(array: np.ndarray):
+    if array.shape[-1] == 1:
+        array = array[:, :, 0]
+    elif array.shape[-1] == 4:
+        array = array[..., :3]
+    elif array.shape[-1] != 3:
+        logger.warning(f"Unexpected image format: {array.shape}.")
+        raise ValueError(f"Unexpected image format: {array.shape}.")
+
+    return array
+
+
+def get_image_metadata(image_info):
+    xref, coords = itemgetter("xref", "bbox")(image_info)
+    x1, y1, x2, y2 = map(rounder, coords)
+
+    width = abs(x2 - x1)
+    height = abs(y2 - y1)
+
+    return {
+        Info.WIDTH: width,
+        Info.HEIGHT: height,
+        Info.X1: x1,
+        Info.X2: x2,
+        Info.Y1: y1,
+        Info.Y2: y2,
+        Info.XREF: xref,
+    }
+
+
+def add_page_metadata(page, metadata):
+    yield from map(partial(merge, get_page_metadata(page)), metadata)
+
+
+def add_alpha_channel_info(doc, metadata):
+    def add_alpha_value_to_metadatum(metadatum):
+        alpha = metadatum_to_alpha_value(metadatum)
+        return {**metadatum, Info.ALPHA: alpha}
+
+    xref_to_alpha = partial(has_alpha_channel, doc)
+    metadatum_to_alpha_value = compose(xref_to_alpha, itemgetter(Info.XREF))
+
+    yield from map(add_alpha_value_to_metadatum, metadata)
+
+
+@lru_cache(maxsize=None)
+def load_image_handle_from_xref(doc, xref):
+    try:
+        return doc.extract_image(xref)
+    except ValueError:
+        logger.debug(f"Xref {xref} is invalid, skipping extraction ...")
+        return
+
+
+rounder = rcompose(round, int)
+
+
+def get_page_metadata(page):
+    page_width, page_height = map(rounder, page.mediabox_size)
+
+    return {
+        Info.PAGE_WIDTH: page_width,
+        Info.PAGE_HEIGHT: page_height,
+        Info.PAGE_IDX: page.number,
+    }
+
+
+def has_alpha_channel(doc, xref):
+    maybe_image = load_image_handle_from_xref(doc, xref)
+    maybe_smask = maybe_image["smask"] if maybe_image else None
+
+    if maybe_smask:
+        return any([doc.extract_image(maybe_smask) is not None, bool(fitz.Pixmap(doc, maybe_smask).alpha)])
+    else:
+        try:
+            return bool(fitz.Pixmap(doc, xref).alpha)
+        except ValueError:
+            logger.debug(f"Encountered invalid xref `{xref}` in {doc.metadata.get('title', '<no title>')}.")
+            return False
+
+
+def tiny(metadata):
+    return metadata[Info.WIDTH] * metadata[Info.HEIGHT] <= 4
+
+
+def clear_caches():
+    get_image_infos.cache_clear()
+    load_image_handle_from_xref.cache_clear()
+    xref_to_image.cache_clear()
+
+
+atexit.register(clear_caches)
+
+
+def breaches_image_to_page_quotient(metadatum):
+    page_width, page_height, x1, x2, y1, y2, width, height = itemgetter(
+        Info.PAGE_WIDTH, Info.PAGE_HEIGHT, Info.X1, Info.X2, Info.Y1, Info.Y2, Info.WIDTH, Info.HEIGHT
+    )(metadatum)
+    geometric_quotient = compute_geometric_quotient(page_width, page_height, x2, x1, y2, y1)
+    quotient_breached = bool(geometric_quotient > CONFIG.filters.image_to_page_quotient.max)
+    return quotient_breached

src/image_prediction/info.py

@@ -0,0 +1,15 @@
+from enum import Enum
+
+
+class Info(Enum):
+    PAGE_WIDTH = "page_width"
+    PAGE_HEIGHT = "page_height"
+    PAGE_IDX = "page_idx"
+    WIDTH = "width"
+    HEIGHT = "height"
+    X1 = "x1"
+    X2 = "x2"
+    Y1 = "y1"
+    Y2 = "y2"
+    ALPHA = "alpha"
+    XREF = "xref"

src/image_prediction/label_mapper/mapper.py

@@ -0,0 +1,10 @@
+import abc
+
+
+class LabelMapper(abc.ABC):
+    @abc.abstractmethod
+    def map_labels(self, items):
+        raise NotImplementedError
+
+    def __call__(self, items):
+        return self.map_labels(items)

src/image_prediction/label_mapper/mappers/numeric.py

@@ -0,0 +1,20 @@
+from typing import Mapping, Iterable
+
+from image_prediction.exceptions import UnexpectedLabelFormat
+from image_prediction.label_mapper.mapper import LabelMapper
+
+
+class IndexMapper(LabelMapper):
+    def __init__(self, labels: Mapping[int, str]):
+        self.__labels = labels
+
+    def __validate_index_label_format(self, index_label: int) -> None:
+        if not 0 <= index_label < len(self.__labels):
+            raise UnexpectedLabelFormat(f"Received index label  '{index_label}' that has no associated string label.")
+
+    def __map_label(self, index_label: int) -> str:
+        self.__validate_index_label_format(index_label)
+        return self.__labels[index_label]
+
+    def map_labels(self, index_labels: Iterable[int]) -> Iterable[str]:
+        return map(self.__map_label, index_labels)

src/image_prediction/label_mapper/mappers/probability.py

@@ -0,0 +1,39 @@
+from enum import Enum
+from operator import itemgetter
+from typing import Mapping, Iterable
+
+import numpy as np
+from funcy import rcompose, rpartial
+
+from image_prediction.exceptions import UnexpectedLabelFormat
+from image_prediction.label_mapper.mapper import LabelMapper
+
+
+class ProbabilityMapperKeys(Enum):
+    LABEL = "label"
+    PROBABILITIES = "probabilities"
+
+
+class ProbabilityMapper(LabelMapper):
+    def __init__(self, labels: Mapping[int, str]):
+        self.__labels = labels
+        # String conversion in the middle due to floating point precision issues.
+        # See: https://stackoverflow.com/questions/56820/round-doesnt-seem-to-be-rounding-properly
+        self.__rounder = rcompose(rpartial(round, 4), str, float)
+
+    def __validate_array_label_format(self, probabilities: np.ndarray) -> None:
+        if not len(probabilities) == len(self.__labels):
+            raise UnexpectedLabelFormat(
+                f"Received fewer probabilities ({len(probabilities)}) than labels were passed ({len(self.__labels)})."
+            )
+
+    def __map_array(self, probabilities: np.ndarray) -> dict:
+        self.__validate_array_label_format(probabilities)
+        cls2prob = dict(
+            sorted(zip(self.__labels, list(map(self.__rounder, probabilities))), key=itemgetter(1), reverse=True)
+        )
+        most_likely = [*cls2prob][0]
+        return {ProbabilityMapperKeys.LABEL: most_likely, ProbabilityMapperKeys.PROBABILITIES: cls2prob}
+
+    def map_labels(self, probabilities: Iterable[np.ndarray]) -> Iterable[dict]:
+        return map(self.__map_array, probabilities)

src/image_prediction/locations.py

@@ -0,0 +1,18 @@
+"""Defines constant paths relative to the module root path."""
+
+from pathlib import Path
+
+# FIXME: move these paths to config, only depending on 'ROOT_PATH' environment variable.
+MODULE_DIR = Path(__file__).resolve().parents[0]
+PACKAGE_ROOT_DIR = MODULE_DIR.parents[0]
+PROJECT_ROOT_DIR = PACKAGE_ROOT_DIR.parents[0]
+
+CONFIG_FILE = PROJECT_ROOT_DIR / "config" / "settings.toml"
+BANNER_FILE = PROJECT_ROOT_DIR / "banner.txt"
+
+DATA_DIR = PROJECT_ROOT_DIR / "data"
+MLRUNS_DIR = str(DATA_DIR / "mlruns")
+
+TEST_DIR = PROJECT_ROOT_DIR / "test"
+TEST_DATA_DIR = TEST_DIR / "data"
+TEST_DATA_DIR_DVC = TEST_DIR / "data.dvc"

src/image_prediction/model_loader/database/connector.py

@@ -0,0 +1,7 @@
+import abc
+
+
+class DatabaseConnector(abc.ABC):
+    @abc.abstractmethod
+    def get_object(self, identifier):
+        raise NotImplementedError

src/image_prediction/model_loader/database/connectors/mock.py

@@ -0,0 +1,9 @@
+from image_prediction.model_loader.database.connector import DatabaseConnector
+
+
+class DatabaseConnectorMock(DatabaseConnector):
+    def __init__(self, store: dict):
+        self.store = store
+
+    def get_object(self, identifier):
+        return self.store[identifier]

src/image_prediction/model_loader/loader.py

@@ -0,0 +1,18 @@
+from functools import lru_cache
+
+from image_prediction.model_loader.database.connector import DatabaseConnector
+
+
+class ModelLoader:
+    def __init__(self, database_connector: DatabaseConnector):
+        self.database_connector = database_connector
+
+    @lru_cache(maxsize=None)
+    def __get_object(self, identifier):
+        return self.database_connector.get_object(identifier)
+
+    def load_model(self, identifier):
+        return self.__get_object(identifier)["model"]
+
+    def load_classes(self, identifier):
+        return self.__get_object(identifier)["classes"]

src/image_prediction/model_loader/loaders/mlflow.py

@@ -0,0 +1,10 @@
+from image_prediction.model_loader.database.connector import DatabaseConnector
+from image_prediction.redai_adapter.mlflow import MlflowModelReader
+
+
+class MlflowConnector(DatabaseConnector):
+    def __init__(self, mlflow_reader: MlflowModelReader):
+        self.mlflow_reader = mlflow_reader
+
+    def get_object(self, run_id):
+        return self.mlflow_reader[run_id]

src/image_prediction/pipeline.py

@@ -0,0 +1,105 @@
+import os
+from functools import lru_cache, partial
+from itertools import chain, tee
+from typing import Iterable, Any
+
+from funcy import rcompose, first, compose, second, chunks, identity, rpartial
+from kn_utils.logging import logger
+from tqdm import tqdm
+
+from image_prediction.config import CONFIG
+from image_prediction.default_objects import (
+    get_formatter,
+    get_mlflow_model_loader,
+    get_image_classifier,
+    get_extractor,
+    get_encoder,
+)
+from image_prediction.locations import MLRUNS_DIR
+from image_prediction.utils.generic import lift, starlift
+
+os.environ["TF_CPP_MIN_LOG_LEVEL"] = "3"
+
+
+@lru_cache(maxsize=None)
+def load_pipeline(**kwargs):
+    logger.info(f"Loading pipeline with kwargs: {kwargs}")
+    model_loader = get_mlflow_model_loader(MLRUNS_DIR)
+    model_identifier = CONFIG.service.mlflow_run_id
+
+    pipeline = Pipeline(model_loader, model_identifier, **kwargs)
+
+    return pipeline
+
+
+def parallel(*fs):
+    return lambda *args: (f(a) for f, a in zip(fs, args))
+
+
+def star(f):
+    return lambda x: f(*x)
+
+
+class Pipeline:
+    def __init__(self, model_loader, model_identifier, batch_size=16, verbose=False, **kwargs):
+        self.verbose = verbose
+
+        extract = get_extractor(**kwargs)
+        classifier = get_image_classifier(model_loader, model_identifier)
+        reformat = get_formatter()
+        represent = get_encoder()
+
+        split = compose(star(parallel(*map(lift, (first, first, second)))), rpartial(tee, 3))
+        classify = compose(chain.from_iterable, lift(classifier), partial(chunks, batch_size))
+        pairwise_apply = compose(star, parallel)
+        join = compose(starlift(lambda prd, rpr, mdt: {"classification": prd, **mdt, "representation": rpr}), star(zip))
+
+        #                       />--classify--\
+        # --extract-->--split--+->--encode---->+--join-->reformat-->filter_duplicates
+        #                       \>--identity--/
+
+        self.pipe = rcompose(
+            extract,  # ... image-metadata-pairs as a stream
+            split,  # ... into an image stream and a metadata stream
+            pairwise_apply(classify, represent, identity),  # ... apply functions to the streams pairwise
+            join,  # ... the streams by zipping
+            reformat,  # ... the items
+            filter_duplicates,  # ... filter out duplicate images
+        )
+
+    def __call__(self, pdf: bytes, page_range: range = None):
+        yield from tqdm(
+            self.pipe(pdf, page_range=page_range),
+            desc="Processing images from document",
+            unit=" images",
+            disable=not self.verbose,
+        )
+
+
+def filter_duplicates(metadata: Iterable[dict[str, Any]]) -> Iterable[dict[str, Any]]:
+    """Filter out duplicate images from the `position` (image coordinates) and `page`, preferring the one with
+    `allPassed` set to True.
+    See RED-10765 (RM-241): Removed redactions reappear for why this is necessary.
+    """
+    keep = dict()
+    for image_meta in metadata:
+        key: tuple[int, int, int, int, int] = (
+            image_meta["position"]["x1"],
+            image_meta["position"]["x2"],
+            image_meta["position"]["y1"],
+            image_meta["position"]["y2"],
+            image_meta["position"]["pageNumber"],
+        )
+        if key in keep:
+            logger.warning(
+                f"Duplicate image found: x1={key[0]}, x2={key[1]}, y1={key[2]}, y2={key[3]}, pageNumber={key[4]}"
+            )
+            if image_meta["filters"]["allPassed"]:
+                logger.warning("Setting the image with allPassed flag set to True")
+                keep[key] = image_meta
+            else:
+                logger.warning("Keeping the previous image since the current image has allPassed flag set to False")
+        else:
+            keep[key] = image_meta
+
+    yield from keep.values()