Move the test-utils bits to a shared place. (#619)

candle-core/src/lib.rs

@@ -63,6 +63,7 @@ pub mod shape;
 mod storage;
 mod strided_index;
 mod tensor;
+pub mod test_utils;
 pub mod utils;
 mod variable;
 

candle-core/src/test_utils.rs

@@ -1,9 +1,4 @@
-#![allow(dead_code)]
-
-#[cfg(feature = "accelerate")]
-extern crate accelerate_src;
-
-use candle_core::{Result, Tensor};
+use crate::{Result, Tensor};
 
 #[macro_export]
 macro_rules! test_device {
@@ -23,6 +18,12 @@ macro_rules! test_device {
     };
 }
 
+pub fn to_vec0_round(t: &Tensor, digits: i32) -> Result<f32> {
+    let b = 10f32.powi(digits);
+    let t = t.to_vec0::<f32>()?;
+    Ok(f32::round(t * b) / b)
+}
+
 pub fn to_vec1_round(t: &Tensor, digits: i32) -> Result<Vec<f32>> {
     let b = 10f32.powi(digits);
     let t = t.to_vec1::<f32>()?;
@@ -40,7 +41,7 @@ pub fn to_vec2_round(t: &Tensor, digits: i32) -> Result<Vec<Vec<f32>>> {
     Ok(t)
 }
 
-pub fn to_vec3_round(t: Tensor, digits: i32) -> Result<Vec<Vec<Vec<f32>>>> {
+pub fn to_vec3_round(t: &Tensor, digits: i32) -> Result<Vec<Vec<Vec<f32>>>> {
     let b = 10f32.powi(digits);
     let t = t.to_vec3::<f32>()?;
     let t = t

candle-core/tests/conv_tests.rs

@@ -1,6 +1,5 @@
-mod test_utils;
 use anyhow::Result;
-use candle_core::{Device, Tensor};
+use candle_core::{test_device, test_utils, Device, Tensor};
 
 /* This test is based on the following script.
 import torch

candle-core/tests/custom_op_tests.rs

@@ -1,10 +1,8 @@
 use candle_core::backend::BackendStorage;
 use candle_core::cpu_backend;
+use candle_core::test_utils::to_vec1_round;
 use candle_core::{CpuStorage, CustomOp1, DType, Device, Error, Layout, Result, Shape, Tensor};
 
-mod test_utils;
-use test_utils::to_vec1_round;
-
 fn fwd<T: num_traits::Float>(v: T, alpha: f64) -> T {
     if v.is_sign_positive() {
         v

candle-core/tests/grad_tests.rs

@@ -1,6 +1,5 @@
 use anyhow::{Context, Result};
-use candle_core::{Device, Shape, Tensor, Var};
-mod test_utils;
+use candle_core::{test_device, test_utils, Device, Shape, Tensor, Var};
 
 fn simple_grad(device: &Device) -> Result<()> {
     let x = Var::new(&[3f32, 1., 4.], device)?;

candle-core/tests/indexing_tests.rs

@@ -1,8 +1,6 @@
 use anyhow::Result;
 use candle_core::{Device, IndexOp, Tensor};
 
-mod test_utils;
-
 #[test]
 fn integer_index() -> Result<()> {
     let dev = Device::Cpu;

candle-core/tests/layout_tests.rs

@@ -1,5 +1,4 @@
-mod test_utils;
-use candle::{Device, IndexOp, Result, Tensor};
+use candle::{test_device, Device, IndexOp, Result, Tensor};
 use candle_core as candle;
 
 fn contiguous(device: &Device) -> Result<()> {

candle-core/tests/pool_tests.rs

@@ -1,5 +1,4 @@
-mod test_utils;
-use candle_core::{Device, IndexOp, Result, Tensor};
+use candle_core::{test_device, test_utils, Device, IndexOp, Result, Tensor};
 
 // https://github.com/huggingface/candle/issues/364
 fn avg_pool2d(dev: &Device) -> Result<()> {
@@ -56,14 +55,17 @@ fn avg_pool2d_pytorch(dev: &Device) -> Result<()> {
     .reshape((1, 2, 4, 4))?;
     let pool = t.avg_pool2d((2, 2), (2, 2))?.squeeze(0)?;
     assert_eq!(
-        test_utils::to_vec3_round(pool, 4)?,
+        test_utils::to_vec3_round(&pool, 4)?,
         [
             [[-1.1926, -0.0395], [0.2688, 0.1871]],
             [[0.1835, -0.1606], [0.6249, 0.3217]]
         ]
     );
     let pool = t.avg_pool2d((3, 3), (3, 3))?.squeeze(0)?;
-    assert_eq!(test_utils::to_vec3_round(pool, 4)?, [[[0.085]], [[0.0078]]]);
+    assert_eq!(
+        test_utils::to_vec3_round(&pool, 4)?,
+        [[[0.085]], [[0.0078]]]
+    );
 
     let t = t.reshape((1, 1, 4, 8))?;
     let pool = t.avg_pool2d((2, 2), (2, 2))?.squeeze(0)?.squeeze(0)?;

candle-core/tests/quantized_tests.rs

@@ -1,11 +1,10 @@
 use candle_core::{
     quantized::{self, GgmlDType},
+    test_utils::to_vec2_round,
     Device, Result, Tensor,
 };
 use quantized::{k_quants, GgmlType};
-mod test_utils;
 use rand::prelude::*;
-use test_utils::to_vec2_round;
 
 const GGML_TEST_SIZE: usize = 32 * 128;
 

candle-core/tests/tensor_tests.rs

@@ -1,5 +1,4 @@
-mod test_utils;
-use candle_core::{DType, Device, IndexOp, Result, Tensor};
+use candle_core::{test_device, DType, Device, IndexOp, Result, Tensor};
 
 fn zeros(device: &Device) -> Result<()> {
     let tensor = Tensor::zeros((5, 2), DType::F32, device)?;