Implementation:LaurentMazare Tch rs No Grad Apply T

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

Overview

Concrete pattern for executing gradient-free forward passes using tch's no_grad context combined with the apply_t helper.

Description

This pattern combines tch::no_grad (a closure that disables gradient tracking) with Tensor::apply_t (a convenience method for calling ModuleT::forward_t). Together they enable efficient feature extraction from frozen models. The no_grad_guard() variant returns an RAII guard that disables gradients until dropped.

Usage

Use to pre-compute features for transfer learning. Wrap the entire feature extraction in tch::no_grad and call apply_t with train: false.

Code Reference

Source Location

• Repository: tch-rs
• File: src/nn/module.rs
• Lines: 50-52 (apply_t)

Signature

// Closure-based API
pub fn no_grad<F: FnOnce() -> T, T>(f: F) -> T

// RAII guard API
pub fn no_grad_guard() -> NoGradGuard

// Tensor helper
impl Tensor {
    pub fn apply_t<M: ModuleT>(&self, m: &M, train: bool) -> Tensor
}

Import

use tch;
use tch::nn::ModuleT;

I/O Contract

Inputs

Name	Type	Required	Description
self (Tensor)	&Tensor	Yes	Input data tensor (e.g., all training images)
m	&M: ModuleT	Yes	Frozen model to apply
train	bool	Yes	Set to false for feature extraction

Outputs

Name	Type	Description
Tensor	Tensor	Feature tensor with no gradient tracking (e.g., [N, 512])

Usage Examples

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

// Pre-compute features for entire dataset
let train_features = tch::no_grad(|| {
    dataset.train_images.apply_t(&backbone, false)
});
let test_features = tch::no_grad(|| {
    dataset.test_images.apply_t(&backbone, false)
});

Related Pages

Implements Principle

• Principle:LaurentMazare_Tch_rs_Frozen_Feature_Computation