Some fixes for yolo-v3. (#529)

* Some fixes for yolo-v3.

* Use the running stats for inference in the batch-norm layer.

* Get some proper predictions for yolo.

* Avoid the quadratic insertion.

candle-core/src/tensor.rs

@@ -497,10 +497,7 @@ impl Tensor {
         let repeats = shape.into();
         let repeats = repeats.dims();
         let mut inp = if self.rank() < repeats.len() {
-            let mut shape = self.dims().to_vec();
+            let shape = [vec![1; repeats.len() - self.rank()], self.dims().to_vec()].concat();
-            while shape.len() < repeats.len() {
-                shape.push(1)
-            }
             self.reshape(shape)?
         } else {
             self.clone()

candle-examples/examples/yolo-v3/darknet.rs

@@ -1,4 +1,4 @@
-use candle::{Device, IndexOp, Result, Tensor};
+use candle::{DType, Device, IndexOp, Result, Tensor};
 use candle_nn::{batch_norm, conv2d, conv2d_no_bias, Func, Module, VarBuilder};
 use std::collections::BTreeMap;
 use std::fs::File;
@@ -145,11 +145,12 @@ fn conv(vb: VarBuilder, index: usize, p: usize, b: &Block) -> Result<(usize, Bl)
             Some(bn) => bn.forward(&xs)?,
             None => xs,
         };
-        if leaky {
+        let xs = if leaky {
-            xs.maximum(&(&xs * 0.1)?)
+            xs.maximum(&(&xs * 0.1)?)?
         } else {
+            xs
+        };
         Ok(xs)
-        }
     });
     Ok((filters, Bl::Layer(Box::new(func))))
 }
@@ -225,12 +226,13 @@ fn detect(
     let grid = Tensor::arange(0u32, grid_size as u32, &Device::Cpu)?;
     let a = grid.repeat((grid_size, 1))?;
     let b = a.t()?.contiguous()?;
-    let x_offset = a.unsqueeze(1)?;
+    let x_offset = a.flatten_all()?.unsqueeze(1)?;
-    let y_offset = b.unsqueeze(1)?;
+    let y_offset = b.flatten_all()?.unsqueeze(1)?;
     let xy_offset = Tensor::cat(&[&x_offset, &y_offset], 1)?
         .repeat((1, nanchors))?
         .reshape((grid_size * grid_size * nanchors, 2))?
-        .unsqueeze(0)?;
+        .unsqueeze(0)?
+        .to_dtype(DType::F32)?;
     let anchors: Vec<f32> = anchors
         .iter()
         .flat_map(|&(x, y)| vec![x as f32 / stride as f32, y as f32 / stride as f32].into_iter())
@@ -245,7 +247,8 @@ fn detect(
     let ys02 = (candle_nn::ops::sigmoid(&ys02)?.add(&xy_offset)? * stride as f64)?;
     let ys24 = (ys24.exp()?.mul(&anchors)? * stride as f64)?;
     let ys4 = candle_nn::ops::sigmoid(&ys4)?;
-    Tensor::cat(&[ys02, ys24, ys4], 2)
+    let ys = Tensor::cat(&[ys02, ys24, ys4], 2)?;
+    Ok(ys)
 }
 
 impl Darknet {

candle-examples/examples/yolo-v3/extract-weights.py

@@ -0,0 +1,7 @@
+def remove_prefix(text, prefix):
+  return text[text.startswith(prefix) and len(prefix):]
+nps = {}
+for k, v in model.state_dict().items():
+  k = remove_prefix(k, 'module_list.')
+  nps[k] = v.detach().numpy()
+np.savez('yolo-v3.ot', **nps)

candle-examples/examples/yolo-v3/main.rs

@@ -11,22 +11,22 @@ use anyhow::Result;
 use candle::{DType, Device, Tensor};
 use candle_nn::{Module, VarBuilder};
 use clap::Parser;
+use image::{DynamicImage, ImageBuffer};
 
-const CONFIG_NAME: &str = "candle-examples/examples/yolo/yolo-v3.cfg";
+const CONFIDENCE_THRESHOLD: f32 = 0.5;
-const CONFIDENCE_THRESHOLD: f64 = 0.5;
+const NMS_THRESHOLD: f32 = 0.4;
-const NMS_THRESHOLD: f64 = 0.4;
 
 #[derive(Debug, Clone, Copy)]
 struct Bbox {
-    xmin: f64,
+    xmin: f32,
-    ymin: f64,
+    ymin: f32,
-    xmax: f64,
+    xmax: f32,
-    ymax: f64,
+    ymax: f32,
-    confidence: f64,
+    confidence: f32,
 }
 
 // Intersection over union of two bounding boxes.
-fn iou(b1: &Bbox, b2: &Bbox) -> f64 {
+fn iou(b1: &Bbox, b2: &Bbox) -> f32 {
     let b1_area = (b1.xmax - b1.xmin + 1.) * (b1.ymax - b1.ymin + 1.);
     let b2_area = (b2.xmax - b2.xmin + 1.) * (b2.ymax - b2.ymin + 1.);
     let i_xmin = b1.xmin.max(b2.xmin);
@@ -38,18 +38,35 @@ fn iou(b1: &Bbox, b2: &Bbox) -> f64 {
 }
 
 // Assumes x1 <= x2 and y1 <= y2
-pub fn draw_rect(_: &mut Tensor, _x1: usize, _x2: usize, _y1: usize, _y2: usize) {
+pub fn draw_rect(
-    todo!()
+    img: &mut ImageBuffer<image::Rgb<u8>, Vec<u8>>,
+    x1: u32,
+    x2: u32,
+    y1: u32,
+    y2: u32,
+) {
+    for x in x1..=x2 {
+        let pixel = img.get_pixel_mut(x, y1);
+        *pixel = image::Rgb([255, 0, 0]);
+        let pixel = img.get_pixel_mut(x, y2);
+        *pixel = image::Rgb([255, 0, 0]);
+    }
+    for y in y1..=y2 {
+        let pixel = img.get_pixel_mut(x1, y);
+        *pixel = image::Rgb([255, 0, 0]);
+        let pixel = img.get_pixel_mut(x2, y);
+        *pixel = image::Rgb([255, 0, 0]);
+    }
 }
 
-pub fn report(pred: &Tensor, img: &Tensor, w: usize, h: usize) -> Result<Tensor> {
+pub fn report(pred: &Tensor, img: DynamicImage, w: usize, h: usize) -> Result<DynamicImage> {
     let (npreds, pred_size) = pred.dims2()?;
     let nclasses = pred_size - 5;
     // The bounding boxes grouped by (maximum) class index.
     let mut bboxes: Vec<Vec<Bbox>> = (0..nclasses).map(|_| vec![]).collect();
     // Extract the bounding boxes for which confidence is above the threshold.
     for index in 0..npreds {
-        let pred = Vec::<f64>::try_from(pred.get(index)?)?;
+        let pred = Vec::<f32>::try_from(pred.get(index)?)?;
         let confidence = pred[4];
         if confidence > CONFIDENCE_THRESHOLD {
             let mut class_index = 0;
@@ -91,24 +108,21 @@ pub fn report(pred: &Tensor, img: &Tensor, w: usize, h: usize) -> Result<Tensor>
         bboxes_for_class.truncate(current_index);
     }
     // Annotate the original image and print boxes information.
-    let (_, initial_h, initial_w) = img.dims3()?;
+    let (initial_h, initial_w) = (img.height(), img.width());
-    let mut img = (img.to_dtype(DType::F32)? * (1. / 255.))?;
+    let w_ratio = initial_w as f32 / w as f32;
-    let w_ratio = initial_w as f64 / w as f64;
+    let h_ratio = initial_h as f32 / h as f32;
-    let h_ratio = initial_h as f64 / h as f64;
+    let mut img = img.to_rgb8();
     for (class_index, bboxes_for_class) in bboxes.iter().enumerate() {
         for b in bboxes_for_class.iter() {
             println!("{}: {:?}", coco_classes::NAMES[class_index], b);
-            let xmin = ((b.xmin * w_ratio) as usize).clamp(0, initial_w - 1);
+            let xmin = ((b.xmin * w_ratio) as u32).clamp(0, initial_w - 1);
-            let ymin = ((b.ymin * h_ratio) as usize).clamp(0, initial_h - 1);
+            let ymin = ((b.ymin * h_ratio) as u32).clamp(0, initial_h - 1);
-            let xmax = ((b.xmax * w_ratio) as usize).clamp(0, initial_w - 1);
+            let xmax = ((b.xmax * w_ratio) as u32).clamp(0, initial_w - 1);
-            let ymax = ((b.ymax * h_ratio) as usize).clamp(0, initial_h - 1);
+            let ymax = ((b.ymax * h_ratio) as u32).clamp(0, initial_h - 1);
-            draw_rect(&mut img, xmin, xmax, ymin, ymax.min(ymin + 2));
+            draw_rect(&mut img, xmin, xmax, ymin, ymax);
-            draw_rect(&mut img, xmin, xmax, ymin.max(ymax - 2), ymax);
-            draw_rect(&mut img, xmin, xmax.min(xmin + 2), ymin, ymax);
-            draw_rect(&mut img, xmin.max(xmax - 2), xmax, ymin, ymax);
         }
     }
-    Ok((img * 255.)?.to_dtype(DType::U8)?)
+    Ok(DynamicImage::ImageRgb8(img))
 }
 
 #[derive(Parser, Debug)]
@@ -118,6 +132,9 @@ struct Args {
     #[arg(long)]
     model: String,
 
+    #[arg(long)]
+    config: String,
+
     images: Vec<String>,
 }
 
@@ -128,18 +145,36 @@ pub fn main() -> Result<()> {
     let weights = unsafe { candle::safetensors::MmapedFile::new(&args.model)? };
     let weights = weights.deserialize()?;
     let vb = VarBuilder::from_safetensors(vec![weights], DType::F32, &Device::Cpu);
-    let darknet = darknet::parse_config(CONFIG_NAME)?;
+    let darknet = darknet::parse_config(&args.config)?;
     let model = darknet.build_model(vb)?;
 
-    for image in args.images.iter() {
+    for image_name in args.images.iter() {
+        println!("processing {image_name}");
+        let mut image_name = std::path::PathBuf::from(image_name);
         // Load the image file and resize it.
         let net_width = darknet.width()?;
         let net_height = darknet.height()?;
-        let image = candle_examples::load_image_and_resize(image, net_width, net_height)?;
+
+        let original_image = image::io::Reader::open(&image_name)?
+            .decode()
+            .map_err(candle::Error::wrap)?;
+        let image = {
+            let data = original_image
+                .resize_exact(
+                    net_width as u32,
+                    net_height as u32,
+                    image::imageops::FilterType::Triangle,
+                )
+                .to_rgb8()
+                .into_raw();
+            Tensor::from_vec(data, (net_width, net_height, 3), &Device::Cpu)?.permute((2, 0, 1))?
+        };
         let image = (image.unsqueeze(0)?.to_dtype(DType::F32)? * (1. / 255.))?;
         let predictions = model.forward(&image)?.squeeze(0)?;
-        let _image = report(&predictions, &image, net_width, net_height)?;
+        let image = report(&predictions, original_image, net_width, net_height)?;
-        println!("converted {image}");
+        image_name.set_extension("pp.jpg");
+        println!("writing {image_name:?}");
+        image.save(image_name)?
     }
     Ok(())
 }

candle-nn/src/batch_norm.rs

@@ -40,6 +40,8 @@ impl From<f64> for BatchNormConfig {
 
 #[derive(Debug)]
 pub struct BatchNorm {
+    running_mean: Tensor,
+    running_var: Tensor,
     weight_and_bias: Option<(Tensor, Tensor)>,
     remove_mean: bool,
     eps: f64,
@@ -47,7 +49,14 @@ pub struct BatchNorm {
 }
 
 impl BatchNorm {
-    pub fn new(num_features: usize, weight: Tensor, bias: Tensor, eps: f64) -> Result<Self> {
+    pub fn new(
+        num_features: usize,
+        running_mean: Tensor,
+        running_var: Tensor,
+        weight: Tensor,
+        bias: Tensor,
+        eps: f64,
+    ) -> Result<Self> {
         if eps < 0. {
             candle::bail!("batch-norm eps cannot be negative {eps}")
         }
@@ -64,6 +73,8 @@ impl BatchNorm {
             )
         }
         Ok(Self {
+            running_mean,
+            running_var,
             weight_and_bias: Some((weight, bias)),
             remove_mean: true,
             eps,
@@ -71,11 +82,18 @@ impl BatchNorm {
         })
     }
 
-    pub fn new_no_bias(num_features: usize, eps: f64) -> Result<Self> {
+    pub fn new_no_bias(
+        num_features: usize,
+        running_mean: Tensor,
+        running_var: Tensor,
+        eps: f64,
+    ) -> Result<Self> {
         if eps < 0. {
             candle::bail!("batch-norm eps cannot be negative {eps}")
         }
         Ok(Self {
+            running_mean,
+            running_var,
             weight_and_bias: None,
             remove_mean: true,
             eps,
@@ -84,8 +102,8 @@ impl BatchNorm {
     }
 }
 
-impl crate::Module for BatchNorm {
+impl BatchNorm {
-    fn forward(&self, x: &Tensor) -> Result<Tensor> {
+    pub fn forward_learning(&self, x: &Tensor) -> Result<Tensor> {
         let x_dtype = x.dtype();
         let internal_dtype = match x_dtype {
             DType::F16 | DType::BF16 => DType::F32,
@@ -129,6 +147,29 @@ impl crate::Module for BatchNorm {
     }
 }
 
+impl crate::Module for BatchNorm {
+    fn forward(&self, x: &Tensor) -> Result<Tensor> {
+        let target_shape: Vec<usize> = x
+            .dims()
+            .iter()
+            .enumerate()
+            .map(|(idx, v)| if idx == 1 { *v } else { 1 })
+            .collect();
+        let target_shape = target_shape.as_slice();
+        let x = x
+            .broadcast_sub(&self.running_mean.reshape(target_shape)?)?
+            .broadcast_div(&(self.running_var.reshape(target_shape)? + self.eps)?.sqrt()?)?;
+        match &self.weight_and_bias {
+            None => Ok(x),
+            Some((weight, bias)) => {
+                let weight = weight.reshape(target_shape)?;
+                let bias = bias.reshape(target_shape)?;
+                x.broadcast_mul(&weight)?.broadcast_add(&bias)
+            }
+        }
+    }
+}
+
 pub fn batch_norm<C: Into<BatchNormConfig>>(
     num_features: usize,
     config: C,
@@ -138,6 +179,8 @@ pub fn batch_norm<C: Into<BatchNormConfig>>(
     if config.eps < 0. {
         candle::bail!("batch-norm eps cannot be negative {}", config.eps)
     }
+    let running_mean = vb.get_or_init(num_features, "running_mean", crate::Init::Const(0.))?;
+    let running_var = vb.get_or_init(num_features, "running_var", crate::Init::Const(1.))?;
     let weight_and_bias = if config.affine {
         let weight = vb.get_or_init(num_features, "weight", crate::Init::Const(1.))?;
         let bias = vb.get_or_init(num_features, "bias", crate::Init::Const(0.))?;
@@ -146,6 +189,8 @@ pub fn batch_norm<C: Into<BatchNormConfig>>(
         None
     };
     Ok(BatchNorm {
+        running_mean,
+        running_var,
         weight_and_bias,
         remove_mean: config.remove_mean,
         eps: config.eps,

candle-nn/tests/batch_norm.rs

@@ -7,8 +7,8 @@ extern crate accelerate_src;
 mod test_utils;
 
 use anyhow::Result;
-use candle::{Device, Tensor};
+use candle::{DType, Device, Tensor};
-use candle_nn::{BatchNorm, Module};
+use candle_nn::BatchNorm;
 
 /* The test below has been generated using the following PyTorch code:
 import torch
@@ -21,7 +21,9 @@ print(output.flatten())
 */
 #[test]
 fn batch_norm() -> Result<()> {
-    let bn = BatchNorm::new_no_bias(5, 1e-8)?;
+    let running_mean = Tensor::zeros(5, DType::F32, &Device::Cpu)?;
+    let running_var = Tensor::ones(5, DType::F32, &Device::Cpu)?;
+    let bn = BatchNorm::new_no_bias(5, running_mean.clone(), running_var.clone(), 1e-8)?;
     let input: [f32; 120] = [
         -0.7493, -1.0410, 1.6977, -0.6579, 1.7982, -0.0087, 0.2812, -0.1190, 0.2908, -0.5975,
         -0.0278, -0.2138, -1.3130, -1.6048, -2.2028, 0.9452, 0.4002, 0.0831, 1.0004, 0.1860,
@@ -37,7 +39,7 @@ fn batch_norm() -> Result<()> {
         1.4252, -0.9115, -0.1093, -0.3100, -0.6734, -1.4357, 0.9205,
     ];
     let input = Tensor::new(&input, &Device::Cpu)?.reshape((2, 5, 3, 4))?;
-    let output = bn.forward(&input)?;
+    let output = bn.forward_learning(&input)?;
     assert_eq!(output.dims(), &[2, 5, 3, 4]);
     let output = output.flatten_all()?;
     assert_eq!(
@@ -59,11 +61,13 @@ fn batch_norm() -> Result<()> {
     );
     let bn2 = BatchNorm::new(
         5,
+        running_mean.clone(),
+        running_var.clone(),
         Tensor::new(&[0.5f32], &Device::Cpu)?.broadcast_as(5)?,
         Tensor::new(&[-1.5f32], &Device::Cpu)?.broadcast_as(5)?,
         1e-8,
     )?;
-    let output2 = bn2.forward(&input)?;
+    let output2 = bn2.forward_learning(&input)?;
     assert_eq!(output2.dims(), &[2, 5, 3, 4]);
     let output2 = output2.flatten_all()?;
     let diff2 = ((output2 - (output * 0.5)?)? + 1.5)?.sqr()?;