docling_pdf/
layout.rs

1//! Layout detection via the RT-DETR (`docling-layout-heron`) model exported to
2//! ONNX, run with `ort`. A port of docling-ibm-models' `LayoutPredictor`:
3//! resize the page image to 640×640 and rescale to `[0,1]` (the heron processor
4//! has `do_normalize=false`), run the model, then RT-DETR
5//! `post_process_object_detection` (sigmoid → top-k over query×class →
6//! center-to-corners boxes scaled to the page).
7
8use image::imageops::FilterType;
9use image::RgbImage;
10use ort::session::Session;
11use ort::value::Tensor;
12
13/// The 17 canonical layout classes, indexed by the model's class id
14/// (`config.json` `id2label`).
15pub const LABELS: [&str; 17] = [
16    "caption",
17    "footnote",
18    "formula",
19    "list_item",
20    "page_footer",
21    "page_header",
22    "picture",
23    "section_header",
24    "table",
25    "text",
26    "title",
27    "document_index",
28    "code",
29    "checkbox_selected",
30    "checkbox_unselected",
31    "form",
32    "key_value_region",
33];
34
35/// One detected region, in page points (top-left origin).
36#[derive(Debug, Clone)]
37pub struct Region {
38    pub label: &'static str,
39    pub score: f32,
40    pub l: f32,
41    pub t: f32,
42    pub r: f32,
43    pub b: f32,
44}
45
46/// Base confidence threshold (docling-ibm-models `base_threshold`): the raw
47/// RT-DETR floor before docling's `LayoutPostprocessor` applies its stricter
48/// per-label thresholds ([`label_threshold`]).
49const THRESHOLD: f32 = 0.3;
50const SIDE: u32 = 640;
51
52/// Per-label confidence threshold, ported from docling's
53/// `LayoutPostprocessor.CONFIDENCE_THRESHOLDS`. The raw predictor keeps every
54/// detection above the 0.3 base; the postprocessor then drops a cluster whose
55/// score is below its label's threshold. Applying it here (equivalent, since
56/// every per-label threshold is ≥ the 0.3 base) keeps low-confidence pictures /
57/// tables / list-items out of the assembly, matching docling.
58pub fn label_threshold(label: &str) -> f32 {
59    match label {
60        "section_header"
61        | "title"
62        | "code"
63        | "checkbox_selected"
64        | "checkbox_unselected"
65        | "form"
66        | "key_value_region"
67        | "document_index" => 0.45,
68        // caption, footnote, formula, list_item, page_footer, page_header,
69        // picture, table, text — all 0.5 in docling.
70        _ => 0.5,
71    }
72}
73
74pub struct LayoutModel {
75    session: Session,
76    /// Set when a multi-page inference fails — e.g. a locally built pre-#73
77    /// static graph (fixed batch=1) via `DOCLING_LAYOUT_ONNX` or a stale
78    /// `layout_heron_int8.onnx`. Batched calls then fall back to per-page runs
79    /// instead of failing the conversion.
80    batch_unsupported: bool,
81}
82
83impl LayoutModel {
84    /// Load the ONNX model from `DOCLING_LAYOUT_ONNX`. Without the override,
85    /// prefers `models/layout_heron_int8.onnx` when present (the quantized
86    /// default; `DOCLING_RS_FP32=1` opts out), else `models/layout_heron.onnx`.
87    pub fn load() -> Result<Self, String> {
88        Self::load_with(crate::intra_threads())
89    }
90
91    /// Like [`load`](Self::load) but with an explicit intra-op thread count. A
92    /// parallel page-worker pool loads its helper models on a single thread each
93    /// and gets its speed-up from running pages concurrently instead.
94    pub fn load_with(intra: usize) -> Result<Self, String> {
95        let path = crate::model_path(
96            "DOCLING_LAYOUT_ONNX",
97            "models/layout_heron.onnx",
98            "models/layout_heron_int8.onnx",
99        );
100        let session = Session::builder()
101            .map_err(|e| format!("layout: builder: {e}"))?
102            // Let inference use the available cores (ort otherwise defaults low);
103            // a large PDF runs this model once per page.
104            .with_intra_threads(intra)
105            .map_err(|e| format!("layout: intra_threads: {e}"))?
106            .commit_from_file(&path)
107            .map_err(|e| format!("layout: load {path}: {e}"))?;
108        Ok(Self {
109            session,
110            batch_unsupported: false,
111        })
112    }
113
114    /// Detect layout regions on a page image. `page_w`/`page_h` are the page size
115    /// in points; returned boxes are in those coordinates.
116    pub fn predict(
117        &mut self,
118        img: &RgbImage,
119        page_w: f32,
120        page_h: f32,
121    ) -> Result<Vec<Region>, String> {
122        Ok(self
123            .predict_batch(&[(img, page_w, page_h)])?
124            .pop()
125            .expect("one result per input page"))
126    }
127
128    /// Detect layout regions on several page images with **one** inference call
129    /// (issue #73). The ONNX export has a dynamic batch dimension, so a worker
130    /// can amortize the per-run framework overhead and keep its cores busier on
131    /// multi-page documents. Results are per-image, index-aligned with `pages`,
132    /// and identical to calling [`predict`](Self::predict) per page.
133    pub fn predict_batch(
134        &mut self,
135        pages: &[(&RgbImage, f32, f32)],
136    ) -> Result<Vec<Vec<Region>>, String> {
137        if pages.len() > 1 && self.batch_unsupported {
138            return self.predict_singly(pages);
139        }
140        match self.run_batch(pages) {
141            Err(e) if pages.len() > 1 => {
142                // A graph without the dynamic batch dim (pre-#73 export) fails
143                // only for batch > 1 — remember and recover per page.
144                eprintln!(
145                    "docling-pdf: layout model rejected a {}-page batch ({e}); \
146                     falling back to per-page inference — re-export with \
147                     scripts/install/export_layout.py for batched layout",
148                    pages.len()
149                );
150                self.batch_unsupported = true;
151                self.predict_singly(pages)
152            }
153            other => other,
154        }
155    }
156
157    fn predict_singly(
158        &mut self,
159        pages: &[(&RgbImage, f32, f32)],
160    ) -> Result<Vec<Vec<Region>>, String> {
161        pages
162            .iter()
163            .map(|p| Ok(self.run_batch(&[*p])?.pop().expect("one result")))
164            .collect()
165    }
166
167    fn run_batch(&mut self, pages: &[(&RgbImage, f32, f32)]) -> Result<Vec<Vec<Region>>, String> {
168        if pages.is_empty() {
169            return Ok(Vec::new());
170        }
171        // Resize each page to 640×640 (RT-DETR ignores aspect ratio), rescale to
172        // [0,1], lay out as NCHW.
173        let n = (SIDE * SIDE) as usize;
174        let batch = pages.len();
175        let mut data = vec![0f32; batch * 3 * n];
176        for (p, (img, _, _)) in pages.iter().enumerate() {
177            let resized = image::imageops::resize(*img, SIDE, SIDE, FilterType::Triangle);
178            let page_off = p * 3 * n;
179            for (i, px) in resized.pixels().enumerate() {
180                data[page_off + i] = px[0] as f32 / 255.0;
181                data[page_off + n + i] = px[1] as f32 / 255.0;
182                data[page_off + 2 * n + i] = px[2] as f32 / 255.0;
183            }
184        }
185        let input = Tensor::from_array(([batch, 3, SIDE as usize, SIDE as usize], data))
186            .map_err(|e| format!("layout: input tensor: {e}"))?;
187        let outputs = self
188            .session
189            .run(ort::inputs!["pixel_values" => input])
190            .map_err(|e| format!("layout: inference: {e}"))?;
191        let (lshape, logits) = outputs["logits"]
192            .try_extract_tensor::<f32>()
193            .map_err(|e| format!("layout: extract logits: {e}"))?;
194        let (_, boxes) = outputs["pred_boxes"]
195            .try_extract_tensor::<f32>()
196            .map_err(|e| format!("layout: extract boxes: {e}"))?;
197
198        let num_queries = lshape[1] as usize;
199        let num_classes = lshape[2] as usize;
200
201        let mut all = Vec::with_capacity(batch);
202        for (p, (_, page_w, page_h)) in pages.iter().enumerate() {
203            let logits =
204                &logits[p * num_queries * num_classes..(p + 1) * num_queries * num_classes];
205            let boxes = &boxes[p * num_queries * 4..(p + 1) * num_queries * 4];
206
207            // sigmoid over every (query, class); take the top `num_queries` scores.
208            let mut scored: Vec<(f32, usize)> = (0..num_queries * num_classes)
209                .map(|idx| (sigmoid(logits[idx]), idx))
210                .collect();
211            scored.sort_unstable_by(|a, b| b.0.total_cmp(&a.0));
212            scored.truncate(num_queries);
213
214            let mut regions = Vec::new();
215            for (score, idx) in scored {
216                if score <= THRESHOLD {
217                    continue;
218                }
219                let label_id = idx % num_classes;
220                let q = idx / num_classes;
221                let cx = boxes[q * 4];
222                let cy = boxes[q * 4 + 1];
223                let w = boxes[q * 4 + 2];
224                let h = boxes[q * 4 + 3];
225                // center_to_corners, then scale normalized coords to page points.
226                let l = (cx - w / 2.0) * page_w;
227                let t = (cy - h / 2.0) * page_h;
228                let r = (cx + w / 2.0) * page_w;
229                let b = (cy + h / 2.0) * page_h;
230                regions.push(Region {
231                    label: LABELS.get(label_id).copied().unwrap_or("text"),
232                    score,
233                    l,
234                    t,
235                    r,
236                    b,
237                });
238            }
239            all.push(regions);
240        }
241        Ok(all)
242    }
243}
244
245fn sigmoid(x: f32) -> f32 {
246    1.0 / (1.0 + (-x).exp())
247}
docling_pdf/layout.rs

docling_pdf/
layout.rs
