Implementation:LaurentMazare Tch rs Resnet18

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

Overview

Concrete tool for instantiating a ResNet-18 vision model provided by the tch vision module.

Description

resnet::resnet18 creates a ResNet-18 model as a FuncT closure implementing ModuleT. The model consists of conv1 (7x7), batch norm, 4 residual layers (2 blocks each with 64/128/256/512 channels), adaptive average pooling, and a final fully-connected layer mapping to num_classes. All parameters are registered in the provided VarStore path.

Usage

Use this to create a ResNet-18 model for image classification. Load pretrained weights via VarStore::load after instantiation. The model expects input of shape [batch, 3, 224, 224].

Code Reference

Source Location

• Repository: tch-rs
• File: src/vision/resnet.rs
• Lines: 78-80

Signature

pub fn resnet18(p: &nn::Path, num_classes: i64) -> FuncT<'static>

Import

use tch::vision::resnet;

I/O Contract

Inputs

Name	Type	Required	Description
p	&nn::Path	Yes	VarStore path for parameter registration
num_classes	i64	Yes	Number of output classes (1000 for ImageNet)

Outputs

Name	Type	Description
FuncT<'static>	impl ModuleT	ResNet-18 model implementing ModuleT (forward_t with train flag)

Usage Examples

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

let mut vs = nn::VarStore::new(Device::Cpu);
let model = resnet::resnet18(&vs.root(), imagenet::CLASS_COUNT);

// Load pretrained weights
vs.load("resnet18.ot")?;

// Inference
let image = imagenet::load_image_and_resize224("photo.jpg")?;
let output = model.forward_t(&image.unsqueeze(0), false);
let probs = output.softmax(-1, Kind::Float);

Related Pages

Implements Principle

• Principle:LaurentMazare_Tch_rs_Pretrained_Model_Instantiation