Implementation:NVIDIA DALI Synset Mapping JSON

Knowledge Sources	NVIDIA_DALI
Domains	Vision, Training
Last Updated	2026-02-08 16:00 GMT

Overview

Provides a JSON array mapping ImageNet class indices to human-readable synset labels for the EfficientNet training example.

Description

This is a static JSON data file containing an array of 1000 strings, where each string at position i corresponds to the human-readable label for ImageNet class index i. Each entry contains the primary class name followed by synonyms and scientific names separated by commas (for example, "tench, Tinca tinca" at index 0 and "goldfish, Carassius auratus" at index 1).

Unlike the Python dictionary variant (synsets.py) used in other DALI examples, this file uses JSON format for broader compatibility and is structured as a flat JSON array rather than a key-value mapping. The array ordering implicitly encodes the class index, making position-based lookup straightforward from any language or tool that supports JSON parsing.

This file is used by the EfficientNet PyTorch training example within the DALI repository to translate numeric prediction outputs into class names during evaluation and visualization.

Usage

Use this file when working with the EfficientNet example to convert model prediction indices to human-readable labels. Load it with any JSON parser and index the resulting array by class index.

Code Reference

Source Location

• Repository: NVIDIA_DALI
• File: docs/examples/use_cases/pytorch/efficientnet/LOC_synset_mapping.json
• Lines: 1-1000

Signature

[
    "tench, Tinca tinca",
    "goldfish, Carassius auratus",
    "great white shark, white shark, man-eater, ...",
    ...
]

Import

import json

with open("LOC_synset_mapping.json", "r") as f:
    synset_mapping = json.load(f)

I/O Contract

Inputs

Name	Type	Required	Description
class_index	int	Yes	An integer in the range [0, 999] used as array index to look up the class label.

Outputs

Name	Type	Description
label	str	Human-readable class name string with comma-separated synonyms and scientific names.

Usage Examples

Loading and querying the synset mapping

import json

with open("LOC_synset_mapping.json", "r") as f:
    synset_mapping = json.load(f)

# Look up class 207 (golden retriever)
predicted_class = 207
label = synset_mapping[predicted_class]
print(f"Predicted: {label}")
# Output: "Predicted: golden retriever"

# Get total number of classes
print(f"Total classes: {len(synset_mapping)}")
# Output: "Total classes: 1000"

Annotating top-k predictions

import json
import torch

with open("LOC_synset_mapping.json", "r") as f:
    synset_mapping = json.load(f)

# Suppose 'output' is model output tensor of shape [1, 1000]
probs = torch.softmax(output, dim=1)
top5_probs, top5_indices = torch.topk(probs, 5, dim=1)

for prob, idx in zip(top5_probs[0], top5_indices[0]):
    print(f"  {synset_mapping[idx.item()]}: {prob.item():.4f}")

Related Pages

• Environment:NVIDIA_DALI_CUDA_GPU_Environment