Implementation:LaurentMazare Tch rs Resnet18 No Final Layer

Knowledge Sources	tch-rs
Domains	Computer_Vision, Transfer_Learning
Last Updated	2026-02-08 14:00 GMT

Overview

Concrete tool for instantiating a ResNet-18 model without the final classification layer, for use as a frozen feature extractor, provided by the tch vision module.

Description

resnet::resnet18_no_final_layer creates a ResNet-18 model that outputs a 512-dimensional feature vector instead of class logits. It passes None for num_classes to the internal resnet constructor, which omits the final fc layer and instead applies adaptive_avg_pool2d followed by flat_view. Pretrained weights can still be loaded with VarStore::load — the missing fc weights are silently skipped.

Usage

Use for transfer learning when you need a frozen feature extractor. After loading pretrained weights, wrap feature computation in tch::no_grad for efficiency.

Code Reference

Source Location

• Repository: tch-rs
• File: src/vision/resnet.rs
• Lines: 82-84

Signature

pub fn resnet18_no_final_layer(p: &nn::Path) -> FuncT<'static>

Import

use tch::vision::resnet;

I/O Contract

Inputs

Name	Type	Required	Description
p	&nn::Path	Yes	VarStore path for parameter registration

Outputs

Name	Type	Description
FuncT<'static>	impl ModuleT	ResNet-18 backbone outputting [batch, 512] feature vectors

Usage Examples

use tch::{nn, nn::ModuleT, vision::resnet, Device};

let mut vs = nn::VarStore::new(Device::Cpu);
let backbone = resnet::resnet18_no_final_layer(&vs.root());
vs.load("resnet18.ot")?;  // fc weights ignored

// Pre-compute features
let features = tch::no_grad(|| {
    dataset.train_images.apply_t(&backbone, false)
});
println!("{:?}", features.size());  // [N, 512]

Related Pages

Implements Principle

• Principle:LaurentMazare_Tch_rs_Feature_Extraction