update note

src/models/deepseek_ocr/model.rs

@@ -151,8 +151,8 @@ impl Attention {
     ) -> Result<(Tensor, Tensor)> {
         let (q_h, q_w) = q_size;
         let (k_h, k_w) = k_size;
-        let rh = self.get_rel_pos(q_h, k_h, rel_pos_h)?; // (h, k, dim)
+        let rh = self.get_rel_pos(q_h, k_h, rel_pos_h)?; // (q_h, k_h, dim)
-        let rw = self.get_rel_pos(q_w, k_w, rel_pos_w)?; // (w, k, dim)
+        let rw = self.get_rel_pos(q_w, k_w, rel_pos_w)?; // (q_w, k_w, dim)
         let (b, _, dim) = q.dims3()?;
         let r_q = q.reshape((b, q_h, q_w, dim))?.contiguous()?;
         let r_q_ = r_q.unsqueeze(D::Minus2)?; // (b, q_h, q_w, 1, dim)
@@ -610,7 +610,8 @@ impl CLIPVisionEmbeddings {
     }
 
     fn get_abs_pos(&self, tgt_size: usize) -> Result<Tensor> {
-        let abs_pos_new = self.pos_embeds.squeeze(0)?;
+        println!("self.pos_embeds: {:?}", self.pos_embeds);
+        let abs_pos_new = self.pos_embeds.clone();
         let (len, dim) = abs_pos_new.dims2()?;
         let src_size = ((len - 1) as f32).sqrt() as usize;
         let tgt_size = (tgt_size as f32).sqrt() as usize;
@@ -991,17 +992,6 @@ impl DeepseekV2MoE {
         Ok(final_xs)
     }
 
-    // pub fn farward(&self, xs: &Tensor) -> Result<Tensor> {
-    //     let identity = xs.clone();
-    //     let (bs, seq_len, embedding_dim) = xs.dims3()?;
-    //     let (topk_idx, topk_weight) = self.gate.forward(xs)?;
-    //     let xs = xs.reshape((bs * seq_len, embedding_dim))?;
-    //     let xs = self.moe_infer(&xs, &topk_idx, &topk_weight)?;
-    //     let xs = xs.reshape((bs, seq_len, embedding_dim))?;
-    //     let xs_shared_experts = self.shared_experts.forward(&identity)?;
-    //     let xs = xs.add(&xs_shared_experts)?;
-    //     Ok(xs)
-    // }
 }
 
 impl Module for DeepseekV2MoE {

src/utils/img_utils.rs

@@ -130,6 +130,7 @@ pub fn find_closest_aspect_ratio(
             best_ratio_diff = ratio_diff;
             best_ratio = ratio;
         } else if (ratio_diff - best_ratio_diff).abs() < 1e-10 {
+            // 当多个候选比例具有相同的宽高比差异时，根据图像的实际面积来选择最优比例。
             let target_area = 0.5 * (image_size as f64).powi(2) * (ratio.0 * ratio.1) as f64;
             if area as f64 > target_area {
                 best_ratio = ratio;
@@ -148,6 +149,7 @@ pub fn dynamic_preprocess(
     let orig_width = image.width();
     let orig_height = image.height();
     let aspect_ratio = orig_width as f64 / orig_height as f64;
+    // 控制分块数量在2-9之间
     let target_ratios = generate_target_ratios_sorted(2, 9);
     let target_aspect_ratio = find_closest_aspect_ratio(
         aspect_ratio,
@@ -196,7 +198,7 @@ pub fn resize_with_edge_padding(
 ) -> DynamicImage {
     // 按图像原比例resize,可能不是输入的宽高
     let mut img = img.resize(width, height, image::imageops::FilterType::CatmullRom);
-    // 使用全0像素填充为输入宽高
+    // 使用输入像素颜色填充为输入宽高
     if img.height() != height || img.width() != width {
         let (img_h, img_w) = (img.height(), img.width());
         let img_buffer = img.to_rgb8();

tests/weight_test.rs

@@ -71,7 +71,7 @@ fn deepseekocr_weight() -> Result<()> {
     for m in &model_list {
         let weights = safetensors::load(m, &device)?;
         for (key, tensor) in weights.iter() {
-            if key.contains("lm_head") {
+            if key.contains("rel_pos_h") {
                 println!("=== {} === {:?}", key, tensor.shape());
             }
             // println!("=== {} === {:?}", key, tensor.shape());