Implementation:NVIDIA DALI EfficientDet Net
| Knowledge Sources | |
|---|---|
| Domains | Object_Detection, TensorFlow |
| Last Updated | 2026-02-08 16:00 GMT |
Overview
Implements the complete EfficientDet object detection network as a Keras Model with integrated training step, validation step, and mAP computation.
Description
This module provides the `EfficientDetNet` class, a `tf.keras.Model` subclass that assembles the full EfficientDet architecture for object detection. It combines an EfficientNet backbone for feature extraction with a Bi-directional Feature Pyramid Network (BiFPN) for multi-scale feature fusion and separate classification (ClassNet) and box regression (BoxNet) prediction heads.
The model initialization configures the backbone network (using `efficientnet_builder`), constructs resampling layers for coarser feature levels, builds FPN cells for feature fusion, and sets up ClassNet and BoxNet heads with configurable parameters including number of anchors (derived from aspect ratios and scales), number of FPN filters, activation types, and separable convolution options. The class also provides built-in training metrics (mean loss, learning rate) and mAP tracking.
The training workflow is managed through `train_step()` which unpacks inputs into features and multi-level classification/box targets, computes forward pass through backbone and detection heads, calculates focal loss for classification and huber/IoU loss for box regression with L2 regularization, and applies gradient updates. A `test_step()` method performs inference and computes mAP using IoU-based matching. The model supports variable freezing via regex patterns and gradient checkpointing for memory efficiency.
Usage
Use this class as the primary model for EfficientDet-based object detection training. Instantiate with a model name (e.g., 'efficientdet-d1') or a custom params dictionary, then use standard Keras fit/evaluate workflows or custom training loops.
Code Reference
Source Location
- Repository: NVIDIA_DALI
- File: docs/examples/use_cases/tensorflow/efficientdet/model/efficientdet_net.py
- Lines: 1-329
Signature
class EfficientDetNet(tf.keras.Model):
def __init__(self, model_name=None, params=None, name=""): ...
def train_step(self, data): ...
def test_step(self, data): ...
def call(self, inputs, training): ...
# Internal methods
def _freeze_vars(self) -> list: ...
def _reg_l2_loss(self, weight_decay, regex=r".*(kernel|weight):0$") -> tf.Tensor: ...
def _unpack_inputs(self, inputs) -> Tuple[tf.Tensor, dict]: ...
def _unpack_outputs(self, cls_out_list, box_out_list) -> Tuple[dict, dict]: ...
def _calc_mAP(self, pred_boxes, pred_scores, pred_classes, gt_boxes, gt_classes): ...
Import
from model.efficientdet_net import EfficientDetNet
model = EfficientDetNet(model_name="efficientdet-d1")
I/O Contract
Inputs
| Name | Type | Required | Description |
|---|---|---|---|
| model_name | str | No | EfficientDet model variant (e.g., 'efficientdet-d0' through 'efficientdet-d7x') |
| params | dict | No | Custom parameter dictionary (alternative to model_name) |
| inputs | tuple | Yes | Training input tuple: (features, num_positives, boxes, classes, *cls_targets, *box_targets) |
Outputs
| Name | Type | Description |
|---|---|---|
| cls_outputs | dict | Per-level classification predictions keyed by level number |
| box_outputs | dict | Per-level box regression predictions keyed by level number |
| train metrics | dict | Dictionary with 'mean_loss', 'loss', 'lr' during training |
| eval metrics | dict | Dictionary with 'mAP' during evaluation |
Usage Examples
Train EfficientDet-D1
import hparams_config
from model.efficientdet_net import EfficientDetNet
# Create model
config = hparams_config.get_efficientdet_config("efficientdet-d1")
config.override({"batch_size": 8, "num_classes": 91})
model = EfficientDetNet(params=config.as_dict())
# Compile and train
model.compile(optimizer=tf.keras.optimizers.SGD(learning_rate=0.08, momentum=0.9))
model.fit(train_dataset, epochs=300, validation_data=val_dataset)