Skip to main content
Version: 0.9.x [Latest Beta]

Custom pipeline construction

This example shows how to build a pipeline manually, adding an object detection node that uses an .denkmodel model file.

📊 Architecture overview​

Input Image → Resize → Object Detection → Filter → Extract Patches → OCR → Results
              ↓                                      ↑
              └──────────────────────────────────────┘

🎨 Visual pipeline diagram​

from denkflow import Pipeline, ImageTensor, HubLicenseSource
import numpy as np
import cv2

# --- Configuration ---
pat = "xxx"
license_id = "xxx"
# Path to the individual model file downloaded from Network Details
od_model_path = "od.denkmodel"
ocr_model_path = "ocr.denkmodel"

# --- License Source Creation ---
# For custom pipelines with per-model `.denkmodel` files, a LicenseSource is typically needed
# when adding AI model nodes. The simplest way is using HubLicenseSource with a PAT:
license_source = HubLicenseSource.from_pat(pat, license_id=license_id)

# --- Pipeline Construction ---
pipeline = Pipeline()

# Pipeline boundary input and constant topics (empty node name → /port_name)
input_image_topic = "/image"
pipeline.set_constant_value("/target_size", np.array([1088, 1280], dtype=np.int64))
pipeline.set_constant_value("/patch_size", np.array([224, 224], dtype=np.int64))
pipeline.set_constant_value("/iou_threshold", np.array([0.7], dtype=np.float32))
pipeline.set_constant_value("/score_threshold", np.array([0.5], dtype=np.float32))

# Add nodes (example: resize -> detect)

### object detection
image_resize_node = pipeline.add_image_resize_node(
    node_name="image-resize",
    image_source=input_image_topic,
    target_size_source="/target_size",
    resize_mode="CenterPadBlack",  # also: "Stretch", "CenterPadWhite"
    execution_provider="cpu",  # Execution provider ("cpu", "cuda", "tensorrt", "directml", "openvino")
    device_id=0,  # Device id for the execution provider
)

object_detection_node = pipeline.add_object_detection_node(
    node_name="object-detection",
    image_source=image_resize_node.output,  # Input from resize node
    model_path=od_model_path,  # Path to the .denkmodel file
    license_source=license_source,  # Pass the created license source
    execution_provider="cpu",  # Execution provider
    device_id=0,  # Device id for the execution provider
)

bb_filter = pipeline.add_bounding_box_filter_node(
    node_name="bb-filter",
    bounding_boxes_source=object_detection_node.output,
    iou_threshold_source="/iou_threshold",
    score_threshold_source="/score_threshold",
    execution_provider="cpu",
    device_id=0,
)

###
### ocr
image_patches_node = pipeline.add_image_patches_node(
    node_name="image-patches",
    image_source=image_resize_node.output,  # Input from resize node
    bboxes_source=bb_filter.output,
    target_size_source="/patch_size",
    resize_mode="CenterPadBlack",       # also: "Stretch", "CenterPadWhite"
    resize_method="Bilinear",           # also: "Nearest", "Bicubic", "Area"
    bounding_box_extend_ratio=0.0,      # extend each bbox by this fraction before cropping
    execution_provider="cpu",  # Execution provider
    device_id=0,  # Device id for the execution provider
)

ocr_node = pipeline.add_ocr_node(
    node_name="ocr-node",
    image_source="image-patches/cutouts",  # Patch cutouts topic from image-patches node
    model_path=ocr_model_path,  # Path to the .denkmodel file
    license_source=license_source,  # Pass the created license source
    execution_provider="cpu",  # Execution provider
    device_id=0,  # Device id for the execution provider
)

###
# --- Inspection ---
print("Custom Pipeline Structure:")
print(pipeline)

# --- Initialization ---
pipeline.initialize()

for topic in pipeline.get_topics():
    print(topic)

# --- Subscribe to Final Output ---
object_receiver = pipeline.subscribe(bb_filter.output)
text_receiver = pipeline.subscribe(ocr_node.output)

# --- Publish Input Image to the defined topic ---
image_path = "test.png"  # Define image_path closer to its use
image_tensor = ImageTensor.from_file(image_path)

# For in-memory image data such as NumPy arrays, see the Creating ImageTensors guide:
# ../creating-image-tensors.md

pipeline.publish_image_tensor(input_image_topic, image_tensor)

# --- Run Pipeline ---
pipeline.run()

# --- Receive and Process Results ---
objects = object_receiver.receive_bounding_box_tensor().to_objects(0.5)
texts = text_receiver.receive_ocr_tensor().to_objects()

print(f"objects {objects}:")
print(f"texts {texts}:")

for obj, text in zip(objects, texts):
    print(f"\n📦 Object Detection Result:")
    print(f"   • Class:      {obj.class_label.name}")
    print(f"   • Confidence: {obj.confidence:.2%}")
    print(f"   • Location:   Top-Left ({obj.x1}, {obj.y1})")
    print(f"                 Bottom-Right ({obj.x2}, {obj.y2})")
    print(f"   • Text:       {text}")
    print("   " + "─" * 50)


# Read the original image
image = cv2.imread(image_path)

# Colors for visualization
colors = {
    "box": (0, 255, 0),  # Green for bounding box
    "text_bg": (0, 0, 0),  # Black for text background
    "text": (255, 255, 255),  # White for text
}

for obj, text in zip(objects, texts):
    # Draw bounding box
    start_point = (int(obj.x1 * image.shape[1]), int(obj.y1 * image.shape[0]))
    end_point = (int(obj.x2 * image.shape[1]), int(obj.y2 * image.shape[0]))
    cv2.rectangle(image, start_point, end_point, colors["box"], 2)

    # Prepare label text
    label = f"{obj.class_label.name} ({obj.confidence:.1%}): {text}"

    # Calculate text size and position
    font = cv2.FONT_HERSHEY_SIMPLEX
    font_scale = 0.6
    thickness = 1
    (text_width, text_height), baseline = cv2.getTextSize(
        label, font, font_scale, thickness
    )

    # Draw text background
    text_x = int(obj.x1 * image.shape[1])
    text_y = (
        int(obj.y1 * image.shape[0]) - 10
        if int(obj.y1 * image.shape[0]) - 10 > text_height
        else int(obj.y1 * image.shape[0]) + text_height
    )
    cv2.rectangle(
        image,
        (text_x, text_y - text_height - baseline),
        (text_x + text_width, text_y + baseline),
        colors["text_bg"],
        -1,
    )  # Filled rectangle

    # Draw text
    cv2.putText(
        image, label, (text_x, text_y), font, font_scale, colors["text"], thickness
    )

# Save the annotated image
cv2.imwrite("output.png", image)
print("\n✨ Annotated image saved as 'output.png'")

Node details​

Topic / NodeTypeInput(s)OutputPurpose
/imagePipeline input-Image tensorExternal image input
/target_sizeConstant-[1088, 1280]Define resize dimensions
image-resizeImage Resize/image + /target_sizeResized imageStandardize image size
object-detectionObject DetectionResized imageBounding boxesDetect objects in image
/iou_thresholdConstant-0.7IoU threshold for bb-filter
/score_thresholdConstant-0.5Score threshold for bb-filter
bb-filterBounding Box FilterRaw detections + thresholdsFiltered bboxesRemove low-confidence/overlapping detections
/patch_sizeConstant-[224, 224]Define patch dimensions
image-patchesImage PatchesResized image + filtered bboxes + patch sizeImage cutoutsExtract object regions
ocr-nodeOCRImage cutoutsText resultsExtract text from patches

Filtering parameters​

  • IoU Threshold: 0.7 (removes overlapping detections)
  • Score Threshold: 0.5 (removes low-confidence detections)