pimm export (config.json)

README.md

@@ -1,27 +1,16 @@
 ---
-license: mit
-datasets:
-- DeepLearnPhysics/PILArNet-M
 library_name: pimm
 tags:
-- particle
-- physics
-- 3D
-- lartpc
-- panda
-- point-transformer-v3
+  - particle-physics
+  - lartpc
+  - point-cloud
 ---
 
-
 # Panda — semantic segmentation
 
-A 5-class semantic segmentation model for LArTPC images: a
-PointTransformerV3 encoder–decoder backbone with a linear per-point classification
-head.
+5-class LArTPC semantic segmentation: PTv3 encoder–decoder + linear per-point head.
 
-- **pimm type:** `DefaultSegmentorV2`
-- **classes:** 5 · **backbone_out_channels:** 64
-- Backbone is the full PTv3 encoder–decoder (`enc_mode=False`).
+- **pimm type:** `DefaultSegmentorV2` · 5 classes
 
 ## Loading
 
@@ -30,24 +19,9 @@ import pimm
 model = pimm.from_pretrained("hf://deeplearnphysics/panda-semantic")
 ```
 
-The architecture and all hyper-parameters are carried in `training_config.json`,
-so no config file is needed. Weights are bitwise-identical to the original
-checkpoint.
-
-## Input
-
-A pimm `Point` dict for LArTPC (PILArNet) point clouds: `coord` (xyz) + `energy`,
-collected as `feat_keys=("coord", "energy")` (4 input channels). Coordinates are
-normalized in the model forward.
-
-## Notes
-
-- Config sets `enable_flash=True`. On CPU or non-FlashAttention setups, override
-  it (e.g. pass `enable_flash=False` to `from_pretrained`, or for detectors set it
-  on the `backbone`).
+Architecture + hyper-parameters travel in `config.json`, so no config file is
+needed. Weights are bitwise-identical to the original checkpoint.
 
 ## Provenance
 
-Repackaged from the original [`DeepLearnPhysics/panda`](https://github.com/DeepLearnPhysics/panda) checkpoints into the
-[`pimm`](https://github.com/youngsm/particle-imaging-models) export format. Inherits
-the license of the source repository.
+Repackaged from the original [`panda`](https://github.com/DeepLearnPhysics/panda) checkpoints into the [`pimm`](https://github.com/youngsm/particle-imaging-models) export format. Inherits the source repo's license.

config.json

@@ -0,0 +1,100 @@
+{
+  "model": {
+    "type": "DefaultSegmentorV2",
+    "num_classes": 5,
+    "backbone_out_channels": 64,
+    "freeze_backbone": false,
+    "backbone": {
+      "type": "PT-v3m2",
+      "in_channels": 4,
+      "order": [
+        "hilbert",
+        "hilbert-trans",
+        "z",
+        "z-trans"
+      ],
+      "stride": [
+        2,
+        2,
+        2,
+        2
+      ],
+      "enc_depths": [
+        3,
+        3,
+        3,
+        9,
+        3
+      ],
+      "enc_channels": [
+        48,
+        96,
+        192,
+        384,
+        512
+      ],
+      "enc_num_head": [
+        3,
+        6,
+        12,
+        24,
+        32
+      ],
+      "enc_patch_size": [
+        256,
+        256,
+        256,
+        256,
+        256
+      ],
+      "dec_depths": [
+        2,
+        2,
+        2,
+        2
+      ],
+      "dec_channels": [
+        64,
+        96,
+        192,
+        384
+      ],
+      "dec_num_head": [
+        4,
+        6,
+        12,
+        24
+      ],
+      "dec_patch_size": [
+        256,
+        256,
+        256,
+        256
+      ],
+      "mlp_ratio": 4,
+      "qkv_bias": true,
+      "qk_scale": null,
+      "layer_scale": 0.0,
+      "attn_drop": 0.0,
+      "proj_drop": 0.0,
+      "drop_path": 0.3,
+      "shuffle_orders": true,
+      "pre_norm": true,
+      "enable_rpe": false,
+      "enable_flash": true,
+      "upcast_attention": false,
+      "upcast_softmax": false,
+      "traceable": false,
+      "mask_token": false,
+      "enc_mode": false,
+      "freeze_encoder": false
+    },
+    "criteria": [
+      {
+        "type": "CrossEntropyLoss",
+        "loss_weight": 1.0,
+        "ignore_index": -1
+      }
+    ]
+  }
+}