Add invoke api method to template

Author: koopmant

app/grandchallenge/forge/templates/forge/partials/example_algorithm/Dockerfile

@@ -18,6 +18,9 @@ RUN python -m pip install \
     --no-color \
     --requirement /opt/app/requirements.txt
 
+COPY --chown=user:user app.py /opt/app/
 COPY --chown=user:user inference.py /opt/app/
 
-ENTRYPOINT ["python", "inference.py"]
+LABEL org.grand-challenge.api-method="invoke"
+
+ENTRYPOINT ["python", "app.py"]

app/grandchallenge/forge/templates/forge/partials/example_algorithm/app.py.template

@@ -9,9 +9,17 @@ Any implementation will do as long as it:
 3. On the invoke endpoint invokes the algorithm for inference and returns HTTP 201 CREATED
 
 """
+from contextlib import asynccontextmanager
 from pathlib import Path
 
+from fastapi import FastAPI, Response, status
 import torch
+import uvicorn
+
+import inference
+
+
+from uvicorn.config import LOGGING_CONFIG
 
 
 def _show_torch_cuda_info():
@@ -45,3 +53,56 @@ def init_model():
         print(f.read())
 
     return model
+
+
+MODELS = {}
+
+
+# During the lifespan of your inference server, your model should be ready
+# for invocations.It is important to load your model here, and not just
+# before running inference, to allow the inference time to be as short as
+# possible. Each invocation will have a timeout, so if your model still
+# needs to be loaded when the /invoke endpoint is called, there may not be
+# enough time for processing.
+@asynccontextmanager
+async def lifespan(app: FastAPI):
+    # Load the ML model
+    MODELS["answer_to_everything"] = init_model()
+    yield
+    # Clean up the models and release the resources
+    MODELS.clear()
+
+
+app = FastAPI(lifespan=lifespan)
+
+
+# After starting your inference server, the health endpoint will
+# be called repeatedly until it returns a 200 response.
+# Redirect responses will not be followed and will raise an exception.
+# Any other response will be ignored.
+@app.get("/health")
+async def health():
+    try:
+        # check if the model is initialized
+        _ = MODELS["answer_to_everything"]
+        return Response(status_code=status.HTTP_200_OK)
+    except KeyError:
+        return Response(status_code=status.HTTP_404_NOT_FOUND)
+
+
+# After the health endpoint returns a 200 response,
+# the invoke endpoint will be called (one or more times)
+# to invoke inference on the inputs in the input folder.
+# When inference is done, this endpoint should return a 201 response.
+# Any other response will raise an exception and fail.
+@app.post("/invoke")
+async def invoke():
+    model = MODELS["answer_to_everything"]
+    inference.run(model)
+    return Response(status_code=status.HTTP_201_CREATED)
+
+
+if __name__ == "__main__":
+    log_config = LOGGING_CONFIG.copy()
+    log_config["handlers"]["default"]["stream"] = "ext://sys.stdout"
+    uvicorn.run(app, host="0.0.0.0", port=4743, log_config=log_config)

app/grandchallenge/forge/templates/forge/partials/example_algorithm/do_test_run.sh.template

@@ -5,6 +5,8 @@ set -e
 
 SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
 DOCKER_IMAGE_TAG="example_algorithm_{{ object.slug }}"
+CONTAINER_NAME="example_algorithm_{{ object.slug }}_container"
+PORT=37847
 
 DOCKER_NOOP_VOLUME="${DOCKER_IMAGE_TAG}-volume"
 
@@ -26,6 +28,9 @@ cleanup() {
       $DOCKER_IMAGE_TAG \
       -c "chmod -R -f o+rwX /output/* || true"
 
+    docker rm -f "$CONTAINER_NAME" >/dev/null 2>&1 || true
+    echo "=+= Container stopped"
+
     # Ensure volume is removed
     docker volume rm "$DOCKER_NOOP_VOLUME" > /dev/null
 }
@@ -56,12 +61,14 @@ docker volume create "$DOCKER_NOOP_VOLUME" > /dev/null
 
 trap cleanup EXIT
 
-run_docker_forward_pass() {
+start_docker_container() {
     local interface_dir="$1"
 
-    echo "=+= Doing a forward pass on ${interface_dir}"
+    echo "=+= Starting container"
 
     ## Note the extra arguments that are passed here:
+    # '-p ${PORT}:4743'
+    #    maps local port to container port 4743
     # '--network none'
     #    entails there is no internet connection
     # '--gpus all'
@@ -70,20 +77,83 @@ run_docker_forward_pass() {
     #   is added because on Grand Challenge this directory cannot be used to store permanent files
     # '--volume ../model:/opt/ml/model/":ro'
     #   is added to provide access to the (optional) tarball-upload locally
-    docker run --rm {% if not no_gpus %}--gpus all {% endif %}\
+    docker run -d {% if not no_gpus %}--gpus all {% endif %}\
+        --name "$CONTAINER_NAME" \
         --platform=linux/amd64 \
-        --network none \
+        -p ${PORT}:4743 \
         --volume "${INPUT_DIR}/${interface_dir}":/input:ro \
         --volume "${OUTPUT_DIR}/${interface_dir}":/output \
         --volume "$DOCKER_NOOP_VOLUME":/tmp \
         --volume "${SCRIPT_DIR}/model":/opt/ml/model:ro \
-        "$DOCKER_IMAGE_TAG"
+        "$DOCKER_IMAGE_TAG" \
+        >/dev/null
+
+    echo "=+= Container started"
+}
+
+check_health() {
+    echo "=+= Waiting for health endpoint..."
+
+    local max_attempts=30
+    local delay=10
+
+    for ((i=1;i<=max_attempts;i++)); do
+        STATUS=$(curl -s -o /dev/null -w "%{http_code}" \
+            --max-time 10 \
+            http://localhost:${PORT}/health || echo "000")
+
+        echo "Health check attempt $i/$max_attempts returned $STATUS"
+
+        if [[ "$STATUS" == "200" ]]; then
+            echo "=+= API healthy"
+            return 0
+        fi
+
+        if [[ "$STATUS" == "302" ]]; then
+            echo "Health endpoint returned 302 — failing"
+            return 1
+        fi
+
+        echo "Retrying in ${delay}s"
+        sleep "$delay"
+    done
+
+    echo "Health endpoint never returned 200"
+    return 1
+}
+
+run_docker_forward_pass() {
+    local interface_dir="$1"
+
+    echo "=+= Doing a forward pass on ${interface_dir}"
+
+    echo "=+= Calling invoke endpoint"
+
+    STATUS=$(curl -s -o /dev/null -w "%{http_code}" \
+        --max-time 300 \
+        -X POST http://localhost:${PORT}/invoke || echo "000")
+
+    if [ "$STATUS" != "201" ]; then
+        echo "Invoke failed with status $STATUS"
+        exit 1
+    fi
+
+    echo "=+= Invoke completed"
+
+    echo "=+= Wrote results to ${OUTPUT_DIR}/${interface_dir}"
+}
 
-  echo "=+= Wrote results to ${OUTPUT_DIR}/${interface_dir}"
+stop_docker_container() {
+    echo "=+= Stopping container"
+    docker rm -f "$CONTAINER_NAME" >/dev/null 2>&1 || true
+    echo "=+= Container stopped"
 }
 
 {% for interface_name in object.algorithm_interface_names %}
+start_docker_container "{{ interface_name }}"
+check_health
 run_docker_forward_pass "{{ interface_name }}"
+stop_docker_container
 {% endfor %}
 
 

app/grandchallenge/forge/templates/forge/partials/example_algorithm/requirements.txt.template

@@ -12,3 +12,5 @@ pylibjpeg
 pylibjpeg-libjpeg
 pylibjpeg-openjpeg
 {% endif %}
+fastapi
+uvicorn