Implementation:NVIDIA DALI TF ResNet Model

Knowledge Sources	NVIDIA_DALI
Domains	Image_Classification, TensorFlow
Last Updated	2026-02-08 16:00 GMT

Overview

Implements the ResNet-50 architecture in TensorFlow/Keras with identity and convolutional residual blocks, L2 regularization, and support for both channels-first and channels-last data formats.

Description

This module provides a complete implementation of the ResNet-50 convolutional neural network for image classification, built using TensorFlow Keras functional API. The architecture follows the standard ResNet-50 design with four stages of residual blocks, each consisting of bottleneck (1x1, 3x3, 1x1 convolution) layers with batch normalization and ReLU activation.

Two types of residual blocks are implemented: `identity_block` for blocks where the input and output dimensions match (using a direct skip connection), and `conv_block` for blocks where spatial dimensions or channel counts change (using a 1x1 convolution on the shortcut path). Both block types use He normal initialization for convolutional kernels and support optional L2 weight regularization (default weight decay: 1e-4).

The `resnet50()` function assembles the complete network: an initial 7x7 convolution with stride 2 and max pooling, followed by stage 2 (3 blocks with 64/64/256 filters), stage 3 (4 blocks with 128/128/512 filters), stage 4 (6 blocks with 256/256/1024 filters), and stage 5 (3 blocks with 512/512/2048 filters). The network concludes with global average pooling and a dense classification layer with softmax activation. The model accepts 224x224x3 input images and supports configurable batch normalization parameters (decay=0.9, epsilon=1e-5).

The implementation supports both channels-first and channels-last data formats (auto-detected from Keras backend), optional input rescaling from [0,1] to the range expected by ImageNet-pretrained models, and configurable L2 regularization.

Usage

Use this module to create a ResNet-50 model for ImageNet classification training with the DALI-accelerated data pipeline. The model is typically instantiated by the `train_ctl` runner function.

Code Reference

Source Location

Repository: NVIDIA_DALI
File: docs/examples/use_cases/tensorflow/resnet-n/resnet_model.py
Lines: 1-415

Signature

def identity_block(input_tensor, kernel_size, filters, stage, block,
                   use_l2_regularizer=True) -> tf.Tensor:
    """The identity block with no conv layer at shortcut."""
    ...

def conv_block(input_tensor, kernel_size, filters, stage, block,
               strides=(2, 2), use_l2_regularizer=True) -> tf.Tensor:
    """A block that has a conv layer at shortcut."""
    ...

def resnet50(num_classes, batch_size=None, use_l2_regularizer=True,
             rescale_inputs=False) -> tf.keras.Model:
    """Instantiates the ResNet50 architecture."""
    ...

Import

from resnet_model import resnet50

model = resnet50(num_classes=1000, batch_size=256)

I/O Contract

Inputs

Name	Type	Required	Description
num_classes	int	Yes	Number of output classes (e.g., 1000 for ImageNet)
batch_size	int	No	Static batch size for input shape (default: None for dynamic)
use_l2_regularizer	bool	No	Whether to apply L2 regularization to Conv/Dense layers (default: True)
rescale_inputs	bool	No	Whether to rescale inputs from [0,1] to ImageNet range (default: False)
input images	tf.Tensor	Yes	Image tensor of shape [batch, 224, 224, 3]

Outputs

Name	Type	Description
model	tf.keras.Model	ResNet-50 Keras model with input shape (224, 224, 3)
predictions	tf.Tensor	Softmax classification probabilities of shape [batch, num_classes]

Usage Examples

Create and Train ResNet-50

from resnet_model import resnet50
from nvutils import image_processing

# Create model
model = resnet50(num_classes=1000, batch_size=256)
model.summary()

# Compile
model.compile(
    optimizer=tf.keras.optimizers.SGD(learning_rate=0.1, momentum=0.9),
    loss=tf.keras.losses.SparseCategoricalCrossentropy(),
    metrics=[
        tf.keras.metrics.SparseTopKCategoricalAccuracy(k=1, name='top1'),
        tf.keras.metrics.SparseTopKCategoricalAccuracy(k=5, name='top5'),
    ],
)

# Train with DALI data
dataset = image_processing.image_set(
    train_files, batch_size=256, height=224, width=224,
    training=True, use_dali="GPU", idx_filenames=idx_files,
)

Related Pages

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment