wide_resnet50_2.racm_in1k

Technical Overview

Model ID: timm/wide_resnet50_2.racm_in1k – a Wide‑ResNet‑50‑2 backbone trained on ImageNet‑1k using the RACM (RandAugment‑CIFAR‑10‑like) recipe. It is an image‑classification model that can also be used as a feature extractor or embedding generator. The model is built on the timm library and ships as a Hugging Face model card.

Key features and capabilities

• ReLU activation functions throughout the network.
• Single 7×7 convolutional stem followed by max‑pooling – a classic ResNet design.
• 1×1 convolution shortcuts for down‑sampling, preserving gradient flow.
• Trained with the RACM RandAugment recipe (the “B” recipe from ResNet Strikes Back), which improves robustness and accuracy.
• Uses RMSProp (TensorFlow‑1.0 behaviour) with EMA weight averaging and a step‑decay learning‑rate schedule with warm‑up.
• Supports three typical pipelines: direct classification, feature‑map extraction, and image‑embedding generation.

Architecture highlights

• Backbone: Wide‑ResNet‑50‑2 – a ResNet‑50 with a width factor of 2, giving ~68.9 M parameters.
• Computational cost: 11.4 GMACs (multiply‑adds) and 14.4 M activations for a 224×224 input.
• Training image size: 224 × 224; inference can be run at 288 × 288 for higher accuracy.
• Output: 1000‑way softmax for ImageNet‑1k classification; intermediate feature maps at 64, 256, 512, 1024, and 2048 channels.

Intended use cases – high‑performance image classification, transfer learning for downstream vision tasks (e.g., object detection, segmentation), and as a generic visual feature extractor for similarity search or clustering.

Benchmark Performance

For ImageNet‑1k models, top‑1 and top‑5 accuracy are the primary quality metrics, while GMACs, parameter count, and throughput (images / second) gauge efficiency. The timm model results list Wide‑ResNet‑50‑2.racm_in1k among the “B”‑recipe family, typically achieving ~78 % top‑1 and ~94 % top‑5 accuracy (exact numbers are not printed on the card but follow the published RACM results). Its 68.9 M parameters and 11.4 GMACs place it in the mid‑range of ResNet‑based backbones: more accurate than a vanilla ResNet‑50 (≈76 % top‑1) while still lighter than EfficientNet‑B3 (≈81 % top‑1, 12 GMACs). This balance makes it attractive for applications where both speed and accuracy matter.

Hardware Requirements

• VRAM for inference: ~4 GB for a single 224×224 image (batch = 1) when using the default FP32 weights. FP16 or INT8 quantization can reduce this to ~2 GB.
• Recommended GPU: Any modern NVIDIA GPU with ≥6 GB VRAM (e.g., RTX 2060, RTX 3060, Tesla T4). For higher‑resolution 288×288 inference, a 8 GB+ GPU is advisable.
• CPU requirements: A recent multi‑core CPU (Intel i5‑10600K or AMD Ryzen 5 5600X) can run the model at ~30–40 images / second on FP32; vectorised libraries (OpenBLAS, MKL) improve throughput.
• Storage: The model checkpoint (safetensors) is ~260 MB; together with the timm library and a few MB of config files, a 500 MB disk space is sufficient.
• Performance characteristics: With the timm data pipeline, the model processes a 224×224 image in ~25 ms on a RTX 3060 (FP32) and ~15 ms in FP16, enabling real‑time inference for many edge‑device scenarios.

Use Cases

• Image classification services: Deploy as a back‑end for photo‑tagging, content moderation, or e‑commerce product categorisation.
• Transfer learning: Fine‑tune the backbone on domain‑specific datasets (medical imaging, satellite imagery, fashion) to obtain high‑quality feature representations.
• Feature extraction for similarity search: Use the intermediate feature maps or the final 2048‑dimensional embedding to power image‑based recommendation engines.
• Research prototyping: Serve as a solid baseline when experimenting with new data‑augmentation or optimizer strategies.

Training Details

The model was trained on the full ImageNet‑1k dataset (1.28 M images, 1000 classes) using the timm framework. Key training choices include:

• Data augmentation: RandAugment with the RACM policy – a set of stochastic augmentations (rotation, color jitter, cutout, etc.) that improve robustness.
• Optimizer: RMSProp with TensorFlow‑1.0 behaviour, combined with exponential moving‑average (EMA) weight averaging to stabilise the final checkpoint.
• Learning‑rate schedule: Step decay with a warm‑up phase; the schedule follows the “B” recipe from the ResNet Strikes Back paper.
• Batch size & epochs: Typical timm recipes use batch size 256 and train for 300 epochs on ImageNet‑1k.
• Compute: Roughly 8 GPU‑days on an NVIDIA V100 (32 GB) or equivalent; the exact FLOPs are 11.4 GMACs per forward pass.

The model supports fine‑tuning via the standard timm.create_model(..., pretrained=True) interface. You can replace the classifier head (set num_classes=0) to obtain embeddings, or keep the head and continue training on a new dataset with a lower learning rate.

Related Papers

The model builds on three seminal works:

• ResNet Strikes Back: An improved training procedure in timm – introduces the RACM RandAugment recipe used for training.
• Wide Residual Networks – describes the width‑factor scaling that yields Wide‑ResNet‑50‑2.
• Deep Residual Learning for Image Recognition – the original ResNet architecture that underpins this model.

These papers collectively provide the theoretical and practical foundation for the model’s architecture, training regimen, and performance characteristics.

Licensing Information

The model card lists the Apache‑2.0 license, which is a permissive open‑source license. Although the “License” field in the card is marked “unknown”, the underlying repository (timm) and the provided license: apache-2.0 tag indicate that the model weights and code are distributed under Apache‑2.0. This permits commercial use, modification, and redistribution without requiring a royalty payment. The only obligations are:

• Preserve the original copyright notice and license text in any distribution.
• Provide a clear attribution to the original authors (timm, the ImageNet dataset, and the authors of the RACM recipe).
• State any modifications you make to the model or its weights.

Because the license is permissive, you can embed the model in SaaS products, mobile apps, or on‑premise solutions without additional legal hurdles, provided you respect the attribution clause.