modify yolo loss function

richardyang92 · Mar 9, 2024 · 755112f · 755112f
1 parent 6bb5cef
commit 755112f
Show file tree

Hide file tree

Showing 4 changed files with 30 additions and 27 deletions.
diff --git a/fashion_mnist/src/data.rs b/fashion_mnist/src/data.rs
@@ -105,7 +105,7 @@ impl FashionMNISTDataset {
 
     fn new(split: &str) -> Self {
         // Download dataset
-        let root = Path::new("/Users/yangyang/Projects/burn-examples/fashion_mnist/data");
+        let root = Path::new("./fashion_mnist/data");
 
         // MNIST is tiny so we can load it in-memory
         // Train images (u8): 28 * 28 * 60000 = 47.04Mb

diff --git a/yolo_v1/src/main.rs b/yolo_v1/src/main.rs
@@ -1,5 +1,5 @@
 
-use burn::{backend::{Autodiff, LibTorch}, optim::AdamWConfig};
+use burn::{backend::{Autodiff, LibTorch}, optim::AdamWConfig, tensor::Tensor};
 // use data::{ItemLoader, VocItem, VocItemLoader};
 use training::TrainingConfig;
 

diff --git a/yolo_v1/src/model.rs b/yolo_v1/src/model.rs
@@ -17,7 +17,7 @@ pub struct BBox {
     ymin: f32,
     xmax: f32,
     ymax: f32,
-    prob: f32,
+    confident: f32,
     box_origin: [f32; 4],
 }
 
@@ -51,7 +51,7 @@ impl From<(&[f32], usize, usize)> for BBox {
             ymin,
             xmax: xmin + box_w,
             ymax: ymin + box_h,
-            prob: box_val[4],
+            confident: box_val[4],
             box_origin
         }
     }
@@ -117,30 +117,33 @@ impl YoloV1Loss {
             for (i, j) in iproduct!(0..segment_w, 0..segment_h) {
                 let (target_has_obj, (target_bbox, _), target_probs) = self.create_bboxes(&target, i, j);
                 let (_, (predict_bbox_1, predict_bbox_2), predict_probs) = self.create_bboxes(&predict, i, j);
+
+                let iou_1 = self.compute_iou::<B>(predict_bbox_1, target_bbox);
+                let iou_2 = self.compute_iou::<B>(predict_bbox_2, target_bbox);
+
+                let (resp_bbox, no_resp_bbox) = if iou_1 > iou_2 {
+                    (predict_bbox_1, predict_bbox_2)
+                } else {
+                    (predict_bbox_2, predict_bbox_1)
+                };
+
                 if target_has_obj {
-                    let iou_1 = self.compute_iou::<B>(predict_bbox_1, target_bbox);
-                    let iou_2 = self.compute_iou::<B>(predict_bbox_2, target_bbox);
-                    let choosen_bbox = if iou_1 > iou_2 {
-                        loss += (predict_bbox_1.prob - target_bbox.prob).powi(2);
-                        loss += self.l_noobj * (predict_bbox_2.prob - target_bbox.prob).powi(2);
-                        predict_bbox_1.box_origin
-                    } else {
-                        loss += self.l_noobj * (predict_bbox_1.prob - target_bbox.prob).powi(2);
-                        loss += (predict_bbox_2.prob - target_bbox.prob).powi(2);
-                        predict_bbox_2.box_origin
-                    };
-                    loss += self.l_coord * ((choosen_bbox[0] - target_bbox.box_origin[0])).powi(2);
-                    loss += self.l_coord * ((choosen_bbox[1] - target_bbox.box_origin[1])).powi(2);
-                    loss += self.l_coord * (choosen_bbox[2].sqrt() - target_bbox.box_origin[2].sqrt()).powi(2);
-                    loss += self.l_coord * (choosen_bbox[3].sqrt() - target_bbox.box_origin[3].sqrt()).powi(2);
-                    for i in 0..20 {
-                        if target_probs[i] > 0f32 {
-                            loss += (predict_probs[i] - target_probs[i]).powi(2);
-                        }
+                    if resp_bbox.confident > 0f32 {
+                        loss += self.l_coord * ((resp_bbox.box_origin[0] - target_bbox.box_origin[0])).powi(2);
+                        loss += self.l_coord * ((resp_bbox.box_origin[1] - target_bbox.box_origin[1])).powi(2);
+                        loss += self.l_coord * (resp_bbox.box_origin[2].sqrt() - target_bbox.box_origin[2].sqrt()).powi(2);
+                        loss += self.l_coord * (resp_bbox.box_origin[3].sqrt() - target_bbox.box_origin[3].sqrt()).powi(2);
+                        loss += (resp_bbox.confident - target_bbox.confident).powi(2);
                     }
                 } else {
-                    loss += self.l_noobj * (predict_bbox_1.prob - target_bbox.prob).powi(2);
-                    loss += self.l_noobj * (predict_bbox_2.prob - target_bbox.prob).powi(2);
+                    loss += self.l_noobj * (resp_bbox.confident - target_bbox.confident).powi(2);
+                    loss += self.l_noobj * (no_resp_bbox.confident - target_bbox.confident).powi(2);
+                }
+
+                for i in 0..20 {
+                    if target_probs[i] > 0f32 {
+                        loss += (predict_probs[i] - target_probs[i]).powi(2);
+                    }
                 }
             }
             loss_vec.push(loss);

diff --git a/yolo_v1/src/training.rs b/yolo_v1/src/training.rs
@@ -12,13 +12,13 @@ pub struct TrainingConfig {
     pub optimizer: AdamWConfig,
     #[config(default = 10)]
     pub num_epochs: usize,
-    #[config(default = 64)]
+    #[config(default = 32)]
     pub batch_size: usize,
     #[config(default = 8)]
     pub num_workers: usize,
     #[config(default = 35)]
     pub seed: u64,
-    #[config(default = 0.0001)]
+    #[config(default = 0.001)]
     pub learing_rate: f64,
 }