huggingface/candle
1
0
Fork 0
mirror of https://github.com/huggingface/candle.git synced 2025-06-16 18:48:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

Move some shared functions to the nn module. (#221)

Browse Source
This commit is contained in:
Laurent Mazare
2023-07-22 14:25:11 +02:00
committed by GitHub
parent 43c7223292
commit 1f26042693
4 changed files with 24 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
candle-examples/examples/simple-training/main.rs
View File

@ -3,27 +3,12 @@
extern crate intel_mkl_src;
use anyhow::Result;
use candle::{DType, Tensor, Var, D};
use candle::{DType, Var, D};
use candle_nn::{loss, ops};
const IMAGE_DIM: usize = 784;
const LABELS: usize = 10;
fn log_softmax<D: candle::shape::Dim>(xs: &Tensor, d: D) -> candle::Result<Tensor> {
let d = d.to_index(xs.shape(), "log-softmax")?;
let max = xs.max_keepdim(d)?;
let diff = xs.broadcast_sub(&max)?;
let sum_exp = diff.exp()?.sum_keepdim(d)?;
let log_sm = diff.broadcast_sub(&sum_exp.log()?)?;
Ok(log_sm)
}
fn nll_loss(inp: &Tensor, target: &Tensor) -> candle::Result<Tensor> {
let b_sz = target.dim(0)?;
inp.gather(target, 1)?
.sum_all()?
.affine(-1f64 / b_sz as f64, 0.)
}
pub fn main() -> Result<()> {
let dev = candle::Device::cuda_if_available(0)?;
let m = candle_nn::vision::mnist::load_dir("data")?;
@ -41,8 +26,8 @@ pub fn main() -> Result<()> {
let test_labels = m.test_labels.to_dtype(DType::U32)?;
for epoch in 1..200 {
let logits = train_images.matmul(&ws)?.broadcast_add(&bs)?;
let log_sm = log_softmax(&logits, D::Minus1)?;
let loss = nll_loss(&log_sm, &train_labels)?;
let log_sm = ops::log_softmax(&logits, D::Minus1)?;
let loss = loss::nll(&log_sm, &train_labels)?;
sgd.backward_step(&loss)?;
let test_logits = test_images.matmul(&ws)?.broadcast_add(&bs)?;

2
candle-nn/src/lib.rs
View File

@ -6,6 +6,8 @@ pub mod embedding;
pub mod init;
pub mod layer_norm;
pub mod linear;
pub mod loss;
pub mod ops;
pub mod optim;
pub mod var_builder;
pub mod vision;

8
candle-nn/src/loss.rs Normal file
View File

@ -0,0 +1,8 @@
use candle::{Result, Tensor};
pub fn nll(inp: &Tensor, target: &Tensor) -> Result<Tensor> {
let b_sz = target.dim(0)?;
inp.gather(target, 1)?
.sum_all()?
.affine(-1f64 / b_sz as f64, 0.)
}

10
candle-nn/src/ops.rs Normal file
View File

@ -0,0 +1,10 @@
use candle::{Result, Tensor};
pub fn log_softmax<D: candle::shape::Dim>(xs: &Tensor, d: D) -> Result<Tensor> {
let d = d.to_index(xs.shape(), "log-softmax")?;
let max = xs.max_keepdim(d)?;
let diff = xs.broadcast_sub(&max)?;
let sum_exp = diff.exp()?.sum_keepdim(d)?;
let log_sm = diff.broadcast_sub(&sum_exp.log()?)?;
Ok(log_sm)
}