Implementation:LaurentMazare Tch rs ModuleT Forward T

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

Overview

Concrete tool for executing a training-aware forward pass through a neural network provided by the tch nn module.

Description

The ModuleT::forward_t trait method and its helper Tensor::apply_t execute a forward pass with explicit training/evaluation mode control. The Module trait has a blanket implementation of ModuleT that ignores the train flag. The batch_accuracy_for_logits method on ModuleT provides a convenience utility for computing accuracy over batched evaluation.

Usage

Use forward_t(&input, false) or input.apply_t(&model, false) for inference. Use forward_t(&input, true) during training when the model contains dropout or batch normalization layers.

Code Reference

Source Location

• Repository: tch-rs
• File: src/nn/module.rs
• Lines: 14 (trait definition), 50-52 (apply_t helper)

Signature

pub trait ModuleT: std::fmt::Debug + Send {
    fn forward_t(&self, xs: &Tensor, train: bool) -> Tensor;
}

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

Import

use tch::nn::ModuleT;

I/O Contract

Inputs

Name	Type	Required	Description
xs	&Tensor	Yes	Input tensor (e.g., [batch, 3, 224, 224] for images)
train	bool	Yes	true for training mode, false for evaluation/inference

Outputs

Name	Type	Description
Tensor	Tensor	Model output (logits), typically [batch, num_classes]

Usage Examples

use tch::nn::ModuleT;

// Direct method call
let logits = model.forward_t(&input, false);

// Via Tensor helper
let logits = input.apply_t(&model, false);

// Batch evaluation
let accuracy = model.batch_accuracy_for_logits(
    &test_images, &test_labels, device, 256
);

Related Pages

Implements Principle

• Principle:LaurentMazare_Tch_rs_Forward_Pass_Inference