regnety_016.tv2_in1k

Technical Overview

What is this model? regnety_016.tv2_in1k is a RegNetY‑1.6 GF backbone trained for image‑classification on the ImageNet‑1k dataset. It is part of the timm (PyTorch‑Image‑Models) library and follows the “design‑space” philosophy of RegNet: a simple, scalable architecture whose depth, width, and bottleneck ratios are derived from a mathematically‑defined space.

Key features and capabilities

• Stochastic depth regularisation for improved generalisation.
• Gradient checkpointing to reduce memory during training.
• Layer‑wise learning‑rate decay, making fine‑tuning easier.
• Configurable output stride (dilation) and optional pre‑activation bottleneck blocks (RegNetV style).
• Supports feature‑map extraction, embedding generation, and classic classification heads.

Architecture highlights – The model contains ~11.2 M parameters, performs ~1.6 GMACs per 224 × 224 image, and produces ~8 M activation elements. Its stage‑wise widths follow the RegNetY‑1.6 GF pattern (32‑48‑120‑336‑888 channels) and each stage uses a bottleneck block with a 1×1‑3×3‑1×1 design. The final linear head maps the 888‑channel feature map (7 × 7) to 1000 ImageNet classes.

Intended use cases – General‑purpose computer‑vision tasks that require a strong, lightweight backbone: image classification, feature extraction for downstream vision pipelines, transfer learning on custom datasets, and as a plug‑in component for object detection or segmentation frameworks.

Benchmark Performance

Benchmarks for image‑classification backbones typically report Top‑1 and Top‑5 accuracy on ImageNet‑1k, FLOPs (GMACs), parameter count, and latency on common hardware. For RegNetY‑016 the published results are:

• Top‑1 accuracy ≈ 81.5 % (ImageNet‑1k validation).
• Top‑5 accuracy ≈ 95.5 %.
• Parameters: 11.2 M.
• GMACs (224 × 224): 1.6 GMAC.

These numbers place RegNetY‑016 comfortably between classic ResNet‑50 (≈76 % Top‑1) and larger RegNetY‑1.6 GF variants, offering a sweet spot of accuracy versus compute. Compared with other timm models of similar size (e.g. regnety_160), RegNetY‑016 delivers ~5 % higher Top‑1 at a modest increase in FLOPs, making it attractive for latency‑sensitive deployments.

Hardware Requirements

VRAM for inference

• FP32 inference on a single 224 × 224 image needs roughly 2 GB of GPU memory.
• FP16 (torch.cuda.amp) can reduce this to 1.2 GB.

Recommended GPU

• Any recent NVIDIA GPU with ≥ 4 GB VRAM (e.g., RTX 2060, GTX 1660 Super).
• For batch‑size > 32, consider a GPU with ≥ 8 GB VRAM (RTX 3060, RTX A4000).

CPU & storage

• CPU inference is feasible but slower; a modern 8‑core processor (e.g., AMD Ryzen 7 5800X) is recommended.
• The model file (safetensors) is ≈ 45 MB. Including the timm library and sample scripts, allocate ≈ 150 MB of disk space.

Performance characteristics – On an RTX 3060, a single forward pass (FP16) takes ~2 ms, enabling > 500 fps for real‑time video streams. CPU latency is ~30 ms per image (single‑core) and ~12 ms with 8‑core parallelism.

Use Cases

Primary applications – The model excels at:

• Standard image‑classification pipelines (e.g., product tagging, content moderation).
• Feature extraction for similarity search or clustering.
• Transfer learning on domain‑specific datasets (medical imaging, satellite imagery, retail).
• Backbone for object‑detection frameworks such as Faster‑RCNN or YOLO‑v5.

Real‑world examples – A fashion e‑commerce platform can use RegNetY‑016 to classify clothing categories in milliseconds, while a wildlife‑monitoring project can extract embeddings for species identification without the computational overhead of larger CNNs.

Industries – Retail, automotive (visual inspection), healthcare (radiology), manufacturing, and any sector that needs fast, accurate visual perception on modest hardware.

Training Details

Methodology – The model was trained by the torchvision contributors using the standard ImageNet‑1k pipeline (1 000 classes, 224 × 224 crops). Enhancements from the timm implementation include:

• Stochastic depth with a drop rate of ~0.2.
• Gradient checkpointing to keep memory usage low.
• Layer‑wise learning‑rate decay (typically 0.9 per stage).
• Configurable activation (SiLU/Swish) and normalization (BatchNorm) layers.

Datasets – Primary pre‑training on ImageNet‑1k. No additional data augmentation beyond standard random resized crops, horizontal flips, and color jitter.

Compute – Training was performed on 8‑GPU NVIDIA V100 (32 GB) clusters for ~90 epochs, taking roughly 12 hours. The model can be fine‑tuned on smaller datasets with as few as 2‑4 GPUs.

Fine‑tuning – Users can set num_classes=0 to obtain raw embeddings, or use features_only=True to retrieve intermediate feature maps. The timm API supports mixed‑precision training and layer‑wise LR schedules out‑of‑the‑box.

Related Papers

The core research behind this model is described in:

• Designing Network Design Spaces (Radosavovic et al., 2020). This paper introduces the RegNet family and demonstrates how a simple, mathematically‑defined design space can produce high‑performing CNNs.

The ImageNet‑1k training recipe follows the torchvision “V2” weight generation (see GitHub discussion), which builds on the same design principles.

Licensing Information

The model is released under the BSD‑3‑Clause license, as indicated in the README. This permissive license permits:

• Commercial and non‑commercial use, modification, and distribution.
• Integration into proprietary software without the need to open‑source your own code.
• Re‑licensing of derived works, provided the original copyright notice and license text are retained.

Restrictions – The only requirement is attribution: you must keep the original copyright notice and disclaimer in any redistribution. No patent grant is provided, so you should verify that downstream use does not infringe on third‑party patents.