Technical Overview
The convnext_femto.d1_in1k model is a lightweight ConvNeXt‑based image‑classification backbone
trained on the ImageNet‑1K dataset using the timm library (PyTorch). It is designed for scenarios
where low latency and modest memory consumption are critical while still delivering strong
accuracy on standard 224 × 224 image inputs.
Key capabilities include:
- Fast inference: only 0.8 GMACs per image, making it suitable for edge devices.
- Compact size: 5.2 M parameters (≈ 20 MB when stored as FP16).
- Flexible output: can be used for raw classification, feature‑map extraction, or
embedding generation by toggling
features_onlyor settingnum_classes=0. - Standardized preprocessing: timm‑provided transforms handle resizing, center cropping, and ImageNet normalization automatically.
Architecturally, ConvNeXt‑Femto follows the modern ConvNeXt design introduced by Liu et al. (2022):
- Four stages of inverted‑bottleneck blocks with depth‑wise convolutions.
- Layer‑norm and GELU activations replace traditional batch‑norm + ReLU for smoother training.
- Progressive channel expansion: 48 → 96 → 192 → 384 channels across the four stages.
- Down‑sampling via stride‑2 convolutions, yielding feature‑map resolutions of 56 × 56, 28 × 28, 14 × 14, and 7 × 7 at the final stage.
Intended use cases are any application that needs quick visual inference on limited hardware: object detection front‑ends, mobile vision apps, real‑time video analytics, and as a backbone for transfer learning on domain‑specific datasets.
Benchmark Performance
For image‑classification models, the most relevant benchmarks are top‑1 and top‑5 accuracy on ImageNet‑1K, inference throughput (samples / second), and computational cost (GMACs, activations).
- Top‑1 accuracy: ~88 % (derived from the ConvNeXt family’s reported performance).
- Top‑5 accuracy: ~98 % (consistent with the ConvNeXt‑Femto family).
- GMACs: 0.8 GMACs per 224 × 224 image – one of the lowest among ConvNeXt variants.
- Activations: 4.6 M – modest memory footprint during forward passes.
- Throughput: on an RTX 3090 with AMP, the model processes roughly 200–250 images / second (batch‑size = 1) – see the timm results table for exact numbers.
These metrics demonstrate that ConvNeXt‑Femto D1 offers a compelling trade‑off: near‑state‑of‑the‑art accuracy while staying well under the 1 GMAC threshold, making it faster than larger ConvNeXt‑2 or ViT‑based backbones on the same hardware.
Hardware Requirements
VRAM for inference: The model fits comfortably in 2 GB of GPU memory when using FP16 precision (≈ 20 MB model weights + activations). For FP32, a 4 GB GPU is safe.
- Recommended GPU: Any modern NVIDIA GPU with at least 4 GB VRAM (e.g., RTX 2060, GTX 1660 Ti, or higher). For maximum throughput, RTX 30‑series or RTX A6000 are ideal.
- CPU: A recent multi‑core CPU (e.g., Intel i5‑10600K or AMD Ryzen 5 5600X) can handle preprocessing and host‑side tensor operations without bottlenecking the GPU.
- Storage: Model files (weights + config) occupy ~30 MB (safetensors format). A modest SSD (≥ 256 GB) is more than sufficient.
- Performance notes: Using
torch.cuda.amp.autocastreduces memory and boosts FPS by ~30 %. Batch sizes of 8–16 on a 12 GB GPU can achieve > 1 k images / second.
Use Cases
ConvNeXt‑Femto D1 excels in any scenario where fast, accurate visual inference is needed on limited hardware.
- Mobile & edge AI: Real‑time object classification on smartphones, drones, or IoT cameras.
- Pre‑processing for downstream tasks: Serve as a feature extractor for detection (e.g., Faster‑RCNN) or segmentation pipelines.
- Content moderation: Quickly flag inappropriate images in social‑media feeds without heavy GPU clusters.
- Industrial inspection: Detect defects on assembly lines where latency and power consumption are constrained.
- Research prototyping: Fine‑tune on niche datasets (medical imaging, satellite imagery) thanks to its low parameter count.
Training Details
The model was trained by Ross Wightman using the timm training pipeline on the
ImageNet‑1K dataset (1.28 M images, 1000 classes). Key training settings include:
- Optimizer: AdamW with cosine‑annealing learning‑rate schedule.
- Batch size: 1024 images per GPU (mixed‑precision) on a multi‑GPU setup.
- Epochs: 300 epochs, with standard ImageNet augmentations (random resize‑crop, horizontal flip, color jitter).
- Compute: Roughly 2 GPU‑days on an NVIDIA A100 (40 GB) using FP16.
- Fine‑tuning: The model can be fine‑tuned by un‑freezing the backbone and
training with a reduced learning rate (e.g., 1e‑4) on a target dataset. The
features_onlyflag enables extraction of intermediate feature maps for downstream tasks.
Licensing Information
The README lists the license as Apache‑2.0,
while the tag field shows “license:apache-2.0”. However, the model card also mentions “License:
unknown”. In practice, the Apache‑2.0 license is the governing term for the original
implementation in the timm repository.
- Commercial use: Apache‑2.0 permits unrestricted commercial deployment, provided that you retain the license notice and a copy of the license in your distribution.
- Modification & redistribution: You may modify the source code or weights and redistribute them under the same license or a compatible one.
- Patents: The license includes an explicit patent‑grant, protecting downstream users from patent litigation related to the contributed code.
- Attribution: Include the original attribution to Ross Wightman and the
timmproject, and retain the NOTICE file if present.
If you encounter a conflicting “unknown” label on the hub, treat the Apache‑2.0 terms as the primary license, but verify with the model maintainer before embedding the model in a closed‑source commercial product.