v5 export: candidate_states + records.onnx assignment head

README.md

@@ -17,63 +17,43 @@ tags:
 # gliner2.5-multi-v1-onnx
 
 ONNX export of [fastino/gliner2.5-multi-v1](https://huggingface.co/fastino/gliner2.5-multi-v1) (GLiNER 2.5
-`BoundaryExtractor`) for onnxruntime-web / WebGPU. **Revision 4:** same fused
-encoder as revision 3, plus `[R]` role-state gather and a separate relation
-scorer graph.
+`BoundaryExtractor`) for onnxruntime-web / WebGPU. **Revision 5:** revision 4
+plus `candidate_states` on the main graph and a tiny `records.onnx` assignment
+head.
 
-One encoder pass. JointIE beam search stays in JavaScript.
+One encoder pass. JointIE beam, classification `implies`/`excludes`, and record
+assignment stay in JavaScript.
 
 Host packing and decode: [Pastel-Org/gliner2.5-onnx-webgpu](https://github.com/Pastel-Org/gliner2.5-onnx-webgpu)
 (live: [gliner25-onnx-webgpu.pages.dev](https://gliner25-onnx-webgpu.pages.dev)).
 
 ## Files
 
-| File | Role |
-|------|------|
-| `onnx/model.onnx` | Encoder + entity pair path + classifier + cached states |
-| `onnx/heads.onnx` | `SparseRelationScorer` only (no encoder). Directional 2H + biaffine. |
+| File | Role | Size |
+|------|------|------|
+| `onnx/model.onnx` | Encoder + entity pair path + classifier + `text_states` + `candidate_states` | 1121.5 MB |
+| `onnx/heads.onnx` | `SparseRelationScorer` only (no encoder) | 23.7 MB |
+| `onnx/records.onnx` | RecordHead assignment (inst/field/cand projections + null column) | 1.19 MB |
 
-## model.onnx inputs
+## model.onnx outputs (new in v5)
 
-| Name | Shape | Dtype |
-|------|-------|-------|
-| input_ids | [B, T] | int64 |
-| attention_mask | [B, T] | int64 |
-| text_word_indices | [B, L] | int64 |
-| text_word_mask | [B, L] | float32 |
-| query_marker_indices | [B, Q] | int64 |
-| query_marker_mask | [B, Q] | float32 |
-| cls_marker_indices | [B, K] | int64 |
-| cls_marker_mask | [B, K] | float32 |
-| rel_marker_indices | [B, R] | int64 |
-| rel_marker_mask | [B, R] | float32 |
+`candidate_states` `[B, Q, C, H]` — C=192. Needed by `records.onnx`. Unused heads:
+pass length-1 indices with mask 0.
 
-Unused heads: pass length-1 indices with mask 0.
+## records.onnx
 
-## model.onnx outputs
-
-| Name | Shape |
-|------|-------|
-| start_logits / end_logits | [B, Q, L+1] |
-| pair_indices / pair_logits / pair_valid | C=192 |
-| cls_logits | [B, K] |
-| text_states | [B, L, H] |
-| query_states | [B, Q, H] |
-| rel_role_states | [B, R, H] |
-
-Host concatenates each `[R] head` + `[R] tail` pair into a 2H relation state.
-
-## heads.onnx
-
-Inputs: `text_states`, `rel_states [B, Rel, 2H]`, `head_start/end`, `tail_start/end`,
-`rel_index`, `pair_mask` (all pair tensors `[B, P]`). Output: `rel_logits [B, P]`.
-`sigmoid(rel_logits)` then JS beam (width 16). Schema constraints stay on the host.
+Inputs: `inst_states [B,N,H]`, `inst_mask [B,N]`, `field_query_states [B,F,H]`,
+`field_cand_states [B,F,C,H]`, `field_cand_mask [B,F,C]`.
+Outputs: `assign_logits [B,N,F,1+C]` (col 0 is ABSENT), `object_logits`,
+`latent_logits`. Natural mode seeds instances from the first `::str` field's
+candidates above threshold.
 
 ## Still not in ONNX
 
-Neural RecordHead instance formation. Repeated JSON objects on the demo are
-host assignment (i-th mention of each field). Constrained `implies`/`excludes`
-for classification is JS when used.
+`score_explicit_spans` (pair-reranker broadcast does not export). The demo
+attaches attributes by looking up the same `(start,end)` on attribute `[E]`
+queries. Full Kleene classification AST is not ported; README `implies` /
+`excludes` is a JS beam.
 
 ## Credits
 

export_config.json

@@ -1,46 +1,5 @@
 {
-  "architecture": "boundary",
-  "export_version": 4,
+  "export_version": 5,
   "base_model": "fastino/gliner2.5-multi-v1",
-  "opset": 17,
-  "inputs": [
-    "input_ids",
-    "attention_mask",
-    "text_word_indices",
-    "text_word_mask",
-    "query_marker_indices",
-    "query_marker_mask",
-    "cls_marker_indices",
-    "cls_marker_mask",
-    "rel_marker_indices",
-    "rel_marker_mask"
-  ],
-  "outputs": [
-    "start_logits",
-    "end_logits",
-    "pair_indices",
-    "pair_logits",
-    "pair_valid",
-    "cls_logits",
-    "text_states",
-    "query_states",
-    "rel_role_states"
-  ],
-  "heads_inputs": [
-    "text_states",
-    "rel_states",
-    "head_start",
-    "head_end",
-    "tail_start",
-    "tail_end",
-    "rel_index",
-    "pair_mask"
-  ],
-  "heads_outputs": [
-    "rel_logits"
-  ],
-  "hidden_size": 768,
-  "directional_relation_states": true,
-  "relation_temperature": 1.0,
-  "notes": "v4: v3 plus rel_marker gather and query_states. heads.onnx is SparseRelationScorer (directional 2H, biaffine on). Host concats [R] head||tail role states. Beam stays in JS. Dummy R=1/P=1 with mask 0 when unused."
+  "notes": "v4 plus candidate_states for RecordHead"
 }

onnx/model.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:42d0f11cb4313d66542c6966ae7e88c6df90cdc77ddbec877a0e33b3c59b4114
-size 1120694273
+oid sha256:4d42b04535a40d5f1da5e1fde57fdf5fc96d897a083895acc58734c7612bc5fb
+size 1121494477

onnx/records.onnx

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bef6dbb75ffbe584fdfc9af19a763c9b6539ea32f5821268f1663dc61beac2a5
+size 1191956